From Data to Behavior: A Comprehensive Guide to GPS and Accelerometer Analysis in Animal Movement Studies

Abstract

This article provides a comprehensive framework for researchers and scientists utilizing integrated GPS and accelerometer data in animal movement analysis. It covers foundational principles of sensor technology and data collection, explores advanced methodological approaches including machine learning classification and movement ecology metrics, addresses critical troubleshooting for data accuracy and sensor deployment, and compares validation techniques for behavioral inference. By synthesizing current methodologies and analytical best practices, this guide aims to enhance the reliability and biological relevance of movement data across research applications, from basic behavioral ecology to conservation and biomedical studies.

The Building Blocks: Understanding Sensor Technology and Core Movement Metrics

Fundamentals of Tri-axial Accelerometers and GPS Sensor Technology

The integration of Global Positioning System (GPS) receivers and tri-axial accelerometers forms the technological cornerstone of modern animal movement analysis research. These micro-sensors, often combined into a single wearable device or "wearable technology," enable researchers to capture detailed data on an animal's location, movement, and behavior in near real-time [1] [2].

Global Positioning System (GPS) Receivers

GPS technology operates as a satellite-based navigation system. The receiver in an animal-borne tag determines its location by communicating with a network of satellites orbiting the Earth. To calculate a position, the receiver must lock onto three or more satellites and perform a calculation known as trilateration to determine the distance to each, thereby fixing its own latitude and longitude [2].

A critical parameter for data quality is the sampling frequency, measured in Hertz (Hz), which dictates how often the unit recalculates and reports its position per second [1]. Higher sampling frequencies generally yield a path that is closer to the animal's true movement, especially during rapid, non-linear locomotion [1].

Table 1: Common GPS Sampling Frequencies and Their Characteristics in Animal Research

Sampling Frequency	Data Points per Second	Typical Use Cases and Considerations
1 Hz [1]	1	Suitable for tracking long-distance, slow-to-moderate speed movements, such as large mammal migration [1].
5 Hz [1]	5	A common frequency for tracking various terrestrial animals; offers a balance between detail and data storage [1].
10 Hz [1]	10	Provides higher resolution for capturing short-distance, high-speed movements and rapid directional changes [1].
15 Hz [1]	15	May provide the highest path resolution; some commercial 15Hz units use interpolation from 10Hz GPS and accelerometer data [2].

Data quality can be compromised by environmental factors that reduce satellite visibility, including dense forest canopy, steep terrain, and man-made structures like stadiums or urban canyons. Other influencing factors are atmospheric conditions, electronic interference, and satellite geometry, collectively known as 'positional dilution of precision' [2].

Tri-axial Accelerometers

A tri-axial accelerometer is a piezo-electrical sensor that measures proper acceleration—the physical acceleration experienced by an object—in three perpendicular planes: X (medial-lateral), Y (anterior-posterior), and Z (vertical) [1] [2]. By measuring the frequency and magnitude of movements in these planes, the accelerometer calculates the total G-forces (with 1g = 9.81 m/s², Earth's gravity) an animal experiences, expressed as a composite vector magnitude [1].

Unlike GPS, which records zero acceleration at a constant velocity, accelerometers capture all movements and impacts, making them ideal for quantifying specific behaviors [1]. They typically operate at much higher frequencies than GPS, such as 100 Hz, to capture the full detail of fine-scale movements and forces [2].

Figure 1: Logical workflow of integrated GPS and accelerometer data collection and analysis for animal movement studies.

Application in Animal Movement Analysis

In movement ecology, understanding why, how, where, and when animals move is fundamental [3]. The synergy of GPS and accelerometer data allows researchers to move beyond simple path trajectories to a mechanistic understanding of behavior.

The primary applications include:

• Behavioral Classification: Machine learning models are applied to high-frequency accelerometer data to classify specific behaviors (e.g., foraging, running, resting) without direct observation [4].
• Energetics and Migration Modeling: Accelerometer data can act as a proxy for energy expenditure. This is crucial for modeling long-distance migrations, as demonstrated in studies of globe-skimmer dragonflies, where energy constraints and wind patterns were used to predict transoceanic migration routes [3].
• Quantifying Human-Wildlife Interactions: Overlaying animal movement data with maps of anthropogenic threats (e.g., shipping lanes, urban development) helps identify collision risks and cumulative stress exposure, informing conservation policy [3].
• Collective Behavior Analysis: Simultaneous tracking of multiple individuals in a group using GPS and accelerometers reveals how collective movements, such as predator evasion in bird flocks, are coordinated [3].

Table 2: Quantitative Metrics Derived from GPS and Accelerometer Data in Animal Research

Sensor	Primary Metrics	Derived / Calculated Variables
GPS Receiver	- Position (Latitude, Longitude) - Timestamp - Number of connected satellites	- Velocity (m/s) and speed - Distance travelled (m) - Home range and habitat use - Changes of direction and turning angles
Tri-axial Accelerometer	- Acceleration in X, Y, Z axes (m/s² or g) - Composite Vector Magnitude (VM)	- Overall Dynamic Body Acceleration (ODBA) - Behavior-specific signatures (e.g., foraging, running) - "Player/Body Load" / Impact quantification - Step count or stroke frequency

Experimental Protocols

Protocol: Field Deployment for Terrestrial Mammal Tracking

This protocol outlines the procedure for deploying a combined GPS-accelerometer tag on a medium-to-large terrestrial mammal (e.g., elk, wolf, caribou) to study movement ecology and behavior.

I. Pre-Deployment Preparation

• Device Selection and Configuration:

  • Select a device with a GPS sampling frequency of 10 Hz or higher to capture fine-scale movement, especially in complex terrain [1].
  • Ensure the integrated accelerometer has a sampling frequency of ≥ 100 Hz for detailed behavioral classification [2].
  • Program the device with a predefined sampling schedule that balances research questions with battery life (e.g., 1 second every 5 minutes during expected active periods).
  • Configure the device to record the number of connected satellites for each fix to allow for post-hoc data filtering based on accuracy.

• Tag Assembly and Testing:

  • Securely house the device in a weatherproof casing.
  • Attach the device to an appropriate collar or harness, ensuring the weight is <5% of the animal's body mass.
  • Conduct bench-testing to verify all sensors are operational and the attachment system is robust.

II. Field Deployment and Data Collection

• Animal Capture and Handling:

  • Following institutional animal care and use committee guidelines, capture target animals using safe and approved methods (e.g., chemical immobilization by a veterinarian, box trapping).
  • Minimize handling time and stress. Fit the collar snugly but with enough space to prevent chafing. Ensure the device sits squarely on the animal, with the accelerometer axes aligned as consistently as possible (e.g., Y-axis anterior-posterior).

• Data Retrieval:

  • Data can be retrieved via UHF/VHF download when in proximity, via cellular networks, or by satellite transmission, depending on the device.
  • For long-term studies, plan for automatic drop-off mechanisms or recapture for device retrieval.

Protocol: Data Processing and Analysis Workflow

This protocol describes the computational steps to transform raw sensor data into ecologically meaningful information.

I. Data Pre-processing and Cleaning

• GPS Data Filtering:

  • Import raw GPS data (latitude, longitude, time, satellite count) into a analysis environment (e.g., R, Python, MoveApps [5]).
  • Remove 3D fixes with a low number of connected satellites (e.g., <4) to reduce positional error [2].
  • Filter out physiologically impossible locations based on speed and turning angle between consecutive points.

• Accelerometer Data Calibration and Integration:

  • Import raw acceleration values for all three axes.
  • For behavioral analysis, often only the dynamic component (movement) is needed. Subtract the static component (gravity) from each axis to isolate animal-induced acceleration.
  • Calculate the Overall Dynamic Body Acceleration (ODBA) or Vectorial Dynamic Body Acceleration (VeDBA) from the dynamic components of the three axes as a proxy for energy expenditure.

II. Integrated Analysis and Modeling

• Behavioral Classification:

  • Extract features from the cleaned accelerometer data (e.g., standard deviation, variance, signal magnitude area) over rolling windows (e.g., 3-5 seconds).
  • Use a machine learning classifier (e.g., Random Forest, Support Vector Machine) trained on ground-truthed data to predict behavioral states (e.g., resting, foraging, travelling) for each window [4].

• Movement Path and Habitat Analysis:

  • Link the classified behaviors to the simultaneous GPS locations.
  • Use the annotated GPS tracks in Resource Selection Functions (RSFs) or Step Selection Functions (SSFs) to model habitat selection relative to environmental variables (e.g., vegetation, distance to road) [6].

Figure 2: Data analysis workflow from raw sensor data to ecological insight.

The Scientist's Toolkit: Research Reagent Solutions

Table 3: Essential Materials and Tools for GPS-Accelerometer Animal Tracking Research

Item / "Reagent"	Function / Application	Examples / Specifications
GPS-Accelerometer Biologger	Primary data collection unit. Captures spatio-temporal location and tri-axial acceleration.	Units with GPS ≥10 Hz and 100 Hz tri-axial accelerometer (e.g., measuring up to 16g on each axis) [1] [2].
Animal Attachment System	Securely and humanely attaches the biologger to the study animal.	Custom-fitted collars (terrestrial mammals), harnesses (birds, some mammals), or glue-on mounts (marine animals).
Data Visualization Software	Explores and communicates animal movements in an environmental context.	ECODATA suite: Creates custom animated maps combining movement tracks with remote sensing and GIS data layers [7] [5].
Analysis Platform (Cloud-Based)	Provides accessible, code-based tools for processing and analyzing movement data.	MoveApps: An interactive, open-source platform for creating and sharing analysis workflows without extensive coding skills [5].
Deep Learning Tracking Toolbox	Provides highly accurate posture and movement tracking from video for model validation.	DeepLabCut, DeepBhvTracking: Uses deep learning (e.g., YOLO algorithm) to track animals in video under complex lab conditions, validating accelerometer-based behavior classification [4].
Statistical Modeling Environment	For developing and applying advanced statistical models to understand movement mechanisms.	R programming language with packages (move, amt, momentuHMM); Python with similar libraries. Used for SSFs, RSFs, and hidden Markov models (HMMs) [6].
5-Chlorotubercidin	5-Chlorotubercidin, CAS:24386-95-6, MF:C11H13ClN4O4, MW:300.70 g/mol	Chemical Reagent
1-Chloroazulene	1-Chloroazulene	1-Chloroazulene (C10H7Cl) for research. A functionalized azulene building block for organic synthesis and materials science. For Research Use Only. Not for human use.

The quantitative analysis of animal movement is fundamental to ecology, conservation, and related biological sciences. High-resolution data from GPS and accelerometer sensors have revolutionized this field, enabling researchers to decipher patterns across scales—from fine-scale foraging decisions to broad-scale migratory strategies [3] [8]. Among the myriad of available metrics, step length, turning angles, and net squared displacement (NSD) form a core set for characterizing movement paths and inferring underlying behavioral states [8]. This document provides detailed application notes and experimental protocols for the use of these key metrics within the context of GPS and accelerometer-based animal movement analysis, framed for an audience of researchers, scientists, and drug development professionals.

Metric Definitions and Quantitative Summaries

Core Metric Definitions

The following table defines the three core movement metrics and their primary ecological interpretations.

Table 1: Definitions and Ecological Interpretations of Key Movement Metrics

Metric	Definition	Ecological Interpretation & Behavioral Context
Step Length	The straight-line displacement between two consecutive GPS coordinate fixes in a trajectory [8].	A primary indicator of movement speed and scale. Longer steps suggest directed travel, exploration, or fleeing, while shorter steps are associated with area-restricted search behaviors like foraging or resting [8].
Turning Angle	The change in the direction of heading (absolute angle) from one movement step to the next [8].	A measure of path tortuosity. Small turning angles (near 0°) indicate directed, persistent movement. Large turning angles (near ±180°) suggest looping or highly tortuous paths, common during intensive searching or habitat sampling [8].
Net Squared Displacement (NSD)	The square of the Euclidean distance between the starting location of a movement path and each subsequent location [9] [8].	Used to identify broad-scale movement strategies at coarse (e.g., annual) temporal scales. Characteristic patterns differentiate migration (double-sigmoid curve), dispersal (sigmoid curve), nomadism (linear), and sedentarism (asymptotic) [9].

The calculation of these metrics relies on high-quality spatio-temporal data. The table below summarizes the data requirements and computational formulae.

Table 2: Data Requirements and Computational Formulae for Movement Metrics

Metric	Required Input Data	Sampling Considerations	Computational Formula
Step Length	A time-ordered series of animal locations (e.g., from GPS).	Highly sensitive to spatio-temporal resolution. Too low a rate may miss fine-scale behavior [8].	( L = \sqrt{(x{t+1} - xt)^2 + (y{t+1} - yt)^2} )Where ( (xt, yt) ) and ( (x{t+1}, y{t+1}) ) are consecutive coordinates.
Turning Angle	A time-ordered series of animal locations from which step vectors can be derived [8].	Sensitive to data resolution and GPS error, which can introduce noise in turning angle calculation.	( \thetat = \arg(\vec{v}{t}) - \arg(\vec{v}_{t-1}) )Where ( \arg(\vec{v}) ) is the direction of the step vector.
Net Squared Displacement (NSD)	A long-term trajectory with a defined origin point [9].	Effective for classifying annual strategies; less suited for fine-scale, gap-ridden data without specialized models [9].	( NSDt = (xt - x0)^2 + (yt - y0)^2 )Where ( (x0, y_0) ) is the path origin.

Experimental Protocols for Data Collection and Analysis

Sensor Deployment and Data Logging Protocol

This protocol outlines a standardized method for collecting movement data, drawing from field experiments in cattle monitoring [10].

• Device Selection: Use commercial or custom devices integrating a tri-axial accelerometer and a GPS sensor. The device should be packaged in a weatherproof case and attached via a collar (for mammals) or harness/other means appropriate to the taxon [10] [11].
• Sensor Configuration:
  • Accelerometer: Sample at a frequency of 10 Hz or higher. Use a dynamic range suitable for the animal's expected movements (e.g., ±2g for cattle) [10]. Data should be stored directly on a micro-SD card within the device.
  • GPS: Set a sampling interval to balance battery life and ecological question (e.g., every 5 minutes for pasture use). Configure the GPS to use a maximum Dilution of Precision (DOP) threshold (e.g., 1) and seek signals from a minimum number of satellites (e.g., 7) to ensure high accuracy [10].
• Field Deployment: Deploy devices on a representative sample of animals. For behavioral classification, simultaneously record video footage of the animals' activities (e.g., grazing, ruminating, lying down) to create a labeled dataset for training machine learning models [10].
• Data Pre-processing: After retrieval, download raw data from the SD card. For accelerometer data, extract features in the time and frequency domains from each axis. Match accelerometer data streams with the corresponding GPS fixes and video annotations using synchronized timestamps [10].

Data Processing and Calibration Protocol

The accuracy of subsequent analysis is critically dependent on proper sensor calibration and data handling [11].

• Accelerometer Calibration:
  • Laboratory Calibration (6-O Method): Prior to deployment, place the static device in six orthogonal orientations, with each accelerometer axis perpendicular to the Earth's surface [11].
  • Record the raw output for each axis in each orientation. The vectorial sum ( \|a\| = \sqrt{x^2 + y^2 + z^2} ) for a static device should be 1.0 g. Deviations indicate sensor error [11].
  • Calculate correction factors (offset and gain) for each axis to ensure the vector sum equals 1.0 g in all orientations. Apply this calibration to all subsequent field data [11].
• GPS Data Processing: Filter location data based on DOP values and number of satellites to remove low-quality fixes. Calculate primary movement metrics (see Table 2) from the cleaned trajectory data [8].

Behavioral Classification and Movement Strategy Analysis

This protocol describes a hybrid approach using accelerometer and GPS data to classify behavior across scales.

• Fine-Scale Behavioral State Classification:
  • Feature Extraction: From the calibrated accelerometer data, calculate 108 features in the time and frequency domains for each axis [10].
  • Model Training: Use a supervised machine learning classifier (e.g., Random Forest). Train the model using the extracted features and the corresponding behaviors labeled from video observation [10].
  • Behavioral Inference: Apply the trained model to the full accelerometer dataset to classify behavior at every time point (e.g., grazing, ruminating, lying, walking) [10].
• Broad-Scale Movement Strategy Analysis:
  • Calculate NSD: Compute the Net Squared Displacement time series for each individual's long-term trajectory [9].
  • Apply Latent State Model: To overcome the limitations of rigid parametric models, analyze the NSD time series using a discrete latent state model (e.g., a hidden Markov model). This model can identify underlying "modes" of movement (e.g., encamped, exploratory) based on the distribution of NSD values [9].
  • Classify Strategy: Use patterns in the time spent within and transitions between these latent modes to classify the overall movement strategy (e.g., migration, nomadism, sedentarism) [9].

Figure 1: Integrated workflow for animal movement data analysis, combining fine-scale accelerometer and broad-scale GPS data.

The Scientist's Toolkit: Research Reagent Solutions

Table 3: Essential Materials and Analytical Tools for Movement Research

Item	Function & Application	Specification Notes
GPS/Accelerometer Tag	Primary data logger for collecting spatio-temporal and dynamic motion data.	Should integrate a 3-D accelerometer (e.g., MEMS) and a GPS receiver. Must be low-power, weatherproof, and suitable for attachment to the study species [10].
Tri-axial Accelerometer	Senses acceleration forces in three orthogonal directions (X, Y, Z), providing detailed information on body motion and orientation [10].	Sample at ≥10 Hz. Dynamic range should be selected based on animal size and movement dynamics (e.g., ±2g for cattle, ±16g for birds) [10] [11].
Calibration Platform	Used for pre-deployment accelerometer calibration to ensure data accuracy and comparability [11].	A simple, level platform is sufficient for the 6-O calibration method to correct for sensor offset and gain errors [11].
Random Forest Classifier	A supervised machine learning algorithm used to classify fine-scale behaviors from accelerometer feature data [10].	Achieves high accuracy (e.g., >0.93 for grazing in cattle) when trained with video-validated data. Implementable in R (package randomForest) or Python (Scikit-learn) [10].
Latent State Model (HMM)	A statistical model for identifying discrete, underlying behavioral modes from time-series data like NSD or step-length/turning-angle distributions [9].	Provides a flexible alternative to rigid parametric models for classifying movement strategies. Can be implemented in R with packages such as moveHMM [9].
Net Squared Displacement (NSD)	A synthetic statistic for visualizing and quantifying large-scale movement patterns and classifying migratory strategies [9] [8].	Calculated as the squared distance from a trajectory origin. Its time-series pattern is diagnostic of migration, dispersal, nomadism, and sedentarism [9].
Dibenz(b,h)acridine	Dibenz(b,h)acridine|	Dibenz(b,h)acridine is a heterocyclic aromatic compound for research. This product is For Research Use Only. Not for diagnostic or therapeutic use.
Dhodh-IN-18	Dhodh-IN-18, MF:C21H16ClF5N6O4, MW:546.8 g/mol	Chemical Reagent

Integrated Analysis and Ecological Inference

The power of modern movement ecology lies in the integration of metrics across sensors and scales. Fine-scale accelerometer classifications can be contextualized within the broader movement strategies revealed by GPS and NSD analysis [10] [9]. For instance, a switch from sedentary behavior to directed, long-distance movement detected via NSD could be further resolved using accelerometry to reveal increased travel time and reduced foraging bouts during migration.

This integrated approach enables robust ecological inference and anomaly detection. Applications include monitoring pasture consumption in livestock, detecting early signs of disease (manifest as abnormal resting or movement), assessing predation threats by identifying herd alert behaviors, and understanding the impacts of environmental change on animal movement and distribution [10] [3]. By adhering to careful protocols for data collection, calibration, and analysis, researchers can ensure that inferences drawn from step length, turning angles, and net squared displacement are both biologically meaningful and statistically sound.

Figure 2: Conceptual diagram showing the integration of key metrics derived from sensor data to address ecological questions.

Dynamic Body Acceleration (DBA) as a Proxy for Energy Expenditure

Dynamic Body Acceleration (DBA) has emerged as a powerful proxy for estimating energy expenditure in free-ranging animals, revolutionizing the field of movement ecology. DBA is a metric derived from tri-axial accelerometers that measures acceleration associated with movement after removing the static component associated with posture [12]. The theoretical foundation rests on the principle that work is equal to the integral of force over distance, and therefore mass-specific energy expenditure at a constant speed is proportional to DBA, provided all work is in the direction of travel [12]. This relationship has opened new avenues for understanding the conservation energetics of species in rapidly changing ecosystems, particularly for animals that are difficult to observe directly [13].

The significance of DBA lies in its ability to circumvent long-standing limitations in ecological research. Traditional methods for estimating field metabolic rate, including mass loss, heart rate monitoring, and respirometry, all pose certain limitations or biases for field applications [12]. Accelerometers, in contrast, can quantify fine-scale movements and body postures unlimited by visibility, observer bias, or the scale of space use [13]. This enables researchers to address fundamental questions about how animals allocate energy among activities such as resting, commuting, and foraging—decisions that ultimately influence life history outcomes, breeding strategies, and survival [12].

Validation and Correlation with Established Methods

Comparison with Doubly Labelled Water

The doubly labelled water (DLW) method represents the gold standard for measuring energy expenditure in free-living conditions and has served as the primary benchmark for validating DBA [14]. The DLW technique involves enriching the body water of a subject with heavy hydrogen (²H) and heavy oxygen (¹⁸O), then determining the difference in washout kinetics between both isotopes, which is a function of carbon dioxide production [14]. This method provides accurate estimates of field metabolic rate over 24-48 hour periods with an accuracy of 1-2% and precision of approximately 5-7% [15].

Studies validating DBA against DLW have demonstrated strong correlations, though the strength varies across species and conditions. Research on Peruvian boobies (Sula variegata) revealed that DBA alone provided the best-fitting model to estimate mass-specific DEE compared with models partitioned per activity and time budget models, with a correlation of r=0.6 [12]. This correlation, while high, is lower than in other avian studies, suggesting that temperature is not the main cause of DBA-DEE decoupling in birds [12]. The validation process typically involves simultaneously deploying both accelerometers and administering DLW to subjects, then comparing the resulting energy expenditure estimates [12].

Table 1: Key Validation Studies of DBA Against Reference Methods

Study Subject	Reference Method	Correlation Coefficient	Key Findings	Source
Peruvian Boobies	Doubly Labelled Water	r = 0.6	DBA alone provided best-fitting model for mass-specific DEE	[12]
Obese Humans	Doubly Labelled Water	N/A	ActiReg underestimated TEE by 3.9%	[16]
Laboratory Validation	Indirect Calorimetry	r = 0.6-0.99	Correlation range across multiple avian studies	[12]

Applications Across Taxa

The application of DBA has expanded to include more than 120 species of animals to date, with studies of wild aquatic species currently outnumbering wild terrestrial species [13]. In domestic animals, DBA has been successfully implemented for behavior classification in cattle, achieving an accuracy of 0.93 for grazing behavior when combined with machine learning algorithms [10]. The methodology has proven particularly valuable for studying species that are notoriously difficult to observe in the wild, including deep-diving marine mammals, nocturnal species, and animals inhabiting complex three-dimensional environments [13].

Experimental Protocols and Methodologies

Accelerometer Deployment and Data Collection

The standard protocol for DBA estimation involves deploying tri-axial accelerometers on the study subjects. These sensors are typically aligned orthogonally to measure acceleration in three dimensions: surge (forward/backward), sway (left/right), and heave (up/down) [13]. For most applications, sensors should be programmed to sample acceleration at frequencies ≥10 Hz to capture the necessary detail of animal movement [10]. The sensors can be set to record continuously or in repeated bursts to conserve battery life and data storage capacity [13].

Proper attachment is crucial for obtaining accurate measurements. Accelerometers should be firmly secured to the animal's body using species-appropriate attachments such as collars, harnesses, or adhesives, with consideration for minimizing impacts on natural behavior [10]. The specific placement location depends on the species and research questions, with neck-mounted sensors proving effective for classifying behaviors like grazing, ruminating, laying, and steady standing in cattle [10].

DBA Calculation and Analysis

The calculation of DBA involves several processing steps to extract meaningful metrics from raw accelerometer data. First, the static acceleration component representing posture must be separated from the dynamic acceleration component representing movement [12]. This is typically achieved through high-pass filtering or by subtracting a running mean from the signal [12]. The vectorial norm of the dynamic acceleration components is then calculated to obtain the overall DBA [12].

The resulting DBA values can be correlated with energy expenditure through calibration studies, either using DLW as a reference or through laboratory-based calorimetry [12]. For behavioral classification, machine learning approaches such as random forest classifiers have demonstrated high accuracy, achieving 0.93 accuracy for classifying grazing behavior in cattle [10]. These methods typically extract features in both time and frequency domains from the accelerometer signals, with studies reporting the extraction of up to 108 features for comprehensive behavioral classification [10].

Table 2: DBA Calculation Methods and Applications

Calculation Method	Key Features	Best Applications	Limitations
Overall Dynamic Body Acceleration	Simple calculation, good for overall energy expenditure	Comparative studies across individuals or species	May miss activity-specific variations
Activity-Specific DBA	Higher precision for specific behaviors	Studies linking specific behaviors to energy costs	Requires additional behavior validation
Machine Learning Classification	Can identify multiple behavior patterns	Comprehensive behavioral ecology studies	Requires extensive training data

Research Toolkit: Essential Materials and Reagents

Table 3: Essential Research Reagents and Equipment for DBA Studies

Item	Specifications	Function/Purpose	Example Sources/Models
Tri-axial Accelerometers	Sampling rate ≥10 Hz, 3-axis measurement, weatherproof housing	Measures acceleration in surge, sway, and heave dimensions	Technosmart Axy, Digitanimal collars [10]
Doubly Labelled Water	²H₂O and H₂¹⁸O mixture, isotope enrichment 99.9% for ²H and 10.0% for ¹⁸O	Gold standard validation of energy expenditure	Medical Isotope, Isotec Inc. [15]
GPS Sensors	5-min sampling interval, ≤5.2m error for 90% of measurements	Tracks animal location and spatial movements	Digitanimal GPS collars [10]
Data Loggers	SD memory cards, sufficient capacity for study duration	Stores accelerometer and sensor data	Various commercial suppliers [10]
Isotope Ratio Mass Spectrometer	High precision for ²H and ¹⁸O measurement	Analyzes isotope enrichment in DLW method	Thermoquest Finnigan MAT Delta Plus [16]
Mesoxalaldehyde	Mesoxalaldehyde|2-Oxopropanedial|CAS 497-16-5	Mesoxalaldehyde (2-Oxopropanedial), CAS 497-16-5. A highly reactive dialdehyde for research. This product is For Research Use Only. Not for human or therapeutic use.	Bench Chemicals
VPhos Pd G3	VPhos Pd G3, MF:C45H57NO4PPdS+, MW:845.4 g/mol	Chemical Reagent	Bench Chemicals

Integration with Movement Ecology Frameworks

The interpretation of DBA data is enhanced through integration with broader movement ecology frameworks. Animal movement tracks can be conceptualized as a hierarchical organization of segments relevant at different spatiotemporal scales [17]. At the most fundamental level are Statistical Movement Elements (StaMEs), which represent the smallest achievable building blocks for hierarchical construction of animal movement tracks [17]. Sequences of StaMEs form Canonical Activity Modes (CAMs), which represent short fixed-length sequences of interpretable activity such as dithering, ambling, or directed walking [17]. These in turn combine to form Behavioral Activity Modes (BAMs), such as gathering resources or beelining, which ultimately compose Diel Activity Routines (DARs) [17].

This hierarchical framework enables researchers to dissect real movement tracks and generate realistic synthetic ones, providing a general tool for testing hypotheses in movement ecology [17]. The approach is particularly valuable for simulating animal movement in diverse contexts such as evaluating an individual's response to landscape changes, release into novel environments, or identifying when individuals are sick or unusually stressed [17].

Applications in Conservation and Management

The application of DBA extends beyond basic research to directly inform conservation strategies and wildlife management. In the Peruvian Humboldt Current system, once supporting 10 million tons of seabird guano prior to the collapse of the anchovy fishery, DBA measurements are being used to understand energy limitations hampering seabird recovery [12]. By quantifying the costs of flying and plunge-diving in species like Peruvian boobies, researchers can better understand the role of anchovy density, distance to anchovy schools, and depth of anchovies in limiting net energy gain and thus reproductive success [12].

In livestock management, accelerometers combined with GPS tracking can detect anomalous behaviors indicative of health issues, predator presence, or parturition events [10]. This enables early intervention and improves animal welfare. The technology also supports sustainable pasture management by identifying unbalanced use of pasture land, helping farmers develop strategies for more rational consumption of natural resources [10]. The ability to continuously monitor animals without human presence eliminates observer effects that can subtly influence animal behavior, providing more accurate data on natural behavior patterns [13].

Designing Effective Data Logging Protocols for Field Studies

The study of animal movement has been revolutionized by biologging technologies, which use animal-borne sensors to monitor location, behavior, and physiology over time and space [18]. Effective data logging protocols form the backbone of rigorous animal movement analysis research, enabling researchers to address fundamental questions in ecology, evolution, and conservation science. This application note provides a comprehensive framework for designing and implementing effective data logging protocols for field studies utilizing GPS and accelerometer technologies. We synthesize current methodologies and provide standardized approaches for data collection, processing, and validation to ensure the collection of high-quality, comparable data across studies and species. As the field moves toward larger data synthesis and smart conservation systems, standardized protocols become increasingly critical for enabling cross-study comparisons and meta-analyses [19] [20] [18].

Hardware Selection and Configuration

Choosing appropriate hardware is fundamental to successful data collection in animal movement studies. The selection process must balance research objectives with practical constraints including device weight, battery life, sensor specifications, and environmental durability.

Table 1: Key Research Reagent Solutions for Animal Movement Studies

Component	Specifications & Examples	Primary Function
GPS Sensor	Sample rate: 5 min to 30 min intervals; Accuracy: ~1.7m average error [21]; DOP threshold: 1; Min satellites: 7 [21]	Records precise location coordinates to track animal movement paths and spatial distribution.
Triaxial Accelerometer	Sampling: 10-25 Hz; Range: ±2-4 g; Axes: 3 orthogonal directions [21] [22]	Captures high-resolution movement dynamics for behavior classification and energy expenditure estimation.
Communication Module	LTE, LoRaWAN, or satellite transmission [23]	Enables remote data offloading, reducing the need for device recovery.
Power System	Rechargeable battery, often with solar panel supplementation [23]	Provides sustained power for extended deployment durations in field conditions.
Casing & Attachment	Weatherproof plastic case; Neck collar, harness, or ear tag [21] [22]	Protects electronics from environment and ensures secure, humane attachment to the study animal.
Data Storage	SD memory card or onboard storage with periodic transmission [21]	Safely retains recorded sensor data until it can be retrieved or transmitted.

The configuration of these components requires careful consideration of trade-offs. For GPS sensors, higher sampling frequencies provide more detailed movement trajectories but significantly reduce battery life. In cattle studies, a 5-minute GPS sampling interval effectively balances battery consumption with spatial resolution for detecting pasture usage patterns [21]. For accelerometers, sampling rates of 10-25 Hz are typically sufficient for classifying major behavioral classes such as grazing, ruminating, and lying [21] [22]. Device positioning also critically influences data quality; neck-mounted accelerometers in cattle effectively distinguish feeding behaviors, whereas leg-mounted sensors might better characterize locomotion patterns [21].

Field Deployment Protocol

Animal Selection and Device Fitting

Proper animal selection and device fitting are crucial for both data quality and animal welfare. Researchers should select subjects representative of the population while considering age, sex, and health status. A sample size of 30 animals has been effectively used in cattle behavior studies to capture representative behavioral patterns [21]. Device weight must not exceed 2-5% of the animal's body mass to avoid impacting natural behavior or causing injury [22]. For neck-collar deployments on cattle, ensure sufficient space for normal swallowing and neck movement while preventing the device from slipping over the head. For thoracic harnesses on birds, proper fit is critical to prevent feather wear while maintaining sensor orientation [22]. Document all deployment details including animal biometrics, device orientation, and deployment timestamp for subsequent data interpretation.

Data Collection and Management

Implement a systematic approach to data collection to ensure consistency throughout the study period. For GPS data, configure devices to record timestamps, coordinates, dilution of precision (DOP), and number of satellites used for each fix. For accelerometer data, record raw acceleration values for all three axes simultaneously. Continuous monitoring over extended periods (days to months) is typically necessary to capture meaningful behavioral patterns and temporal cycles [21] [22]. Establish regular data retrieval schedules via SD card replacement or remote transmission, implementing robust backup procedures to prevent data loss. Metadata documentation should include deployment logs, animal health observations, and environmental conditions that might influence behavior or sensor performance.

Data Pre-processing Workflow

Raw sensor data requires substantial pre-processing before analysis. The following workflow outlines the critical steps for transforming raw data into analysis-ready datasets, incorporating both GPS and accelerometer data streams.

Figure 1: Data pre-processing workflow showing parallel processing paths for GPS and accelerometer data, culminating in a structured dataset ready for analysis.

GPS Data Processing

GPS data requires cleaning to remove erroneous locations before analysis. Implement automated filtering to exclude fixes with high dilution of precision (DOP > 1) and those based on few satellites (<7), as these typically have lower accuracy [21]. Additional filters should remove physiologically implausible locations based on maximum realistic movement speeds between consecutive fixes. For advanced applications, consider using grid search algorithms for received signal strength (RSS) localization, which can provide more than 2 times greater spatial accuracy compared to traditional multilateration methods in wildlife tracking applications [24]. The cleaned location data can then be used to calculate movement metrics such as step lengths, turning angles, residence time, and home range size using methods such as kernel density estimation.

Accelerometer Data Processing

Accelerometer data processing involves multiple transformation steps to enable behavior classification. The raw signal is first segmented into fixed time windows, typically ranging from 3-15 seconds, with or without overlap [25]. For each axis and window, extract numerous features in both time and frequency domains - one effective cattle behavior identification study extracted 108 features including statistical measures (mean, variance, skewness), entropy measures, and frequency components [21]. Consider applying axis-agnostic feature selection methods to ensure robustness to device orientation changes [25]. The resulting feature set creates a structured table where each row represents a time window and columns contain the extracted features, ready for model training.

Table 2: Quantitative Performance of Behavior Classification Models in Various Species

Species	Behaviors Classified	Best-Performing Model	Accuracy/Performance	Key Pre-processing Factors
Beef Cattle [21]	Grazing, Ruminating, Laying, Standing	Random Forest	0.93 (grazing)	108 time/frequency features; 10Hz sampling
Dairy Goats [25]	Rumination, Head in Feeder, Lying, Standing	Custom ML Pipeline (ACT4Behav)	AUC: 0.800-0.829	Behavior-specific pre-processing; Filtering techniques
Sandgrouse [22]	Incubation behavior	Threshold-based Classification	>90% success rate	ODBA calculation; Sex-specific time windows

Machine Learning Classification and Validation

Behavior Classification Methodology

Supervised machine learning represents the state-of-the-art approach for classifying animal behavior from accelerometer data. The process begins with creating a labeled training dataset by matching accelerometer records to directly observed behaviors, typically using video recordings [21] [25]. Random Forest algorithms have demonstrated strong performance for cattle behavior classification, achieving 93% accuracy for distinguishing grazing behavior [21]. For each behavior of interest, train a separate classification model and optimize its pre-processing pipeline independently, as different behaviors may benefit from different window sizes, filtering techniques, and feature selections [25]. This behavior-specific optimization approach has yielded area under curve (AUC) scores of 0.800-0.829 for classifying rumination, feeding, and posture behaviors in dairy goats [25].

Model Validation Protocols

Robust validation is essential for ensuring machine learning models generalize beyond the training data. A recent review revealed that 79% of animal accelerometer studies did not adequately validate their models for overfitting [26]. To address this, implement rigorous validation using completely independent test sets comprising data from animals not included in the training set [26]. This approach reveals the true generalizability of models to new individuals. When testing on unseen goats, one study observed a decrease in AUC scores from 0.800-0.829 to 0.644-0.749, highlighting the importance of independent validation [25]. Always report performance metrics on the independent test set rather than training or validation sets, and consider using nested cross-validation approaches for reliable performance estimation when sample sizes are limited [26].

Data Integration and Synthesis

Modern movement ecology increasingly requires integrating multiple data sources and collaborating across studies. Effective data logging protocols should facilitate future data integration by implementing standardized formatting and comprehensive metadata collection. A recent compilation pipeline for sage-grouse successfully integrated 53 tracking datasets comprising nearly 5 million locations by standardizing data attributes and implementing robust error checking [19]. This integration enabled powerful large-scale analyses not possible with individual datasets. Similarly, emerging frameworks for combining animal tracking data with trait databases (e.g., morphological, physiological, and life history characteristics) create exciting opportunities to address novel research questions about how animal attributes influence movement patterns [18]. When designing logging protocols, anticipate future integration needs by adopting common data standards, vocabularies, and thorough metadata documentation following existing models such as the Movebank data repository [18].

The Movement Ecology Paradigm provides a unified theoretical and conceptual framework for studying the movement of organisms, encompassing the internal state, motion capacity, navigation capacity, and external factors that influence movement trajectories [27]. This paradigm has emerged as a response to the traditionally fragmented study of animal movement, integrating disciplines from biophysics to population ecology. The framework posits that movement results from the continuous interaction between an individual's internal state (why move?), its movement capabilities (how to move?), and its navigation capacity (when and where to move?), all modulated by external environmental factors [27]. The ongoing miniaturization and sophistication of tracking devices has significantly broadened the range of species that can be studied with unprecedented spatial and temporal resolution, fueling the development and application of this paradigm across ecological research.

The paradigm is particularly relevant in contemporary research given its utility for addressing pressing ecological challenges including wildlife conservation, disease ecology, and predicting species responses to environmental change. By offering a holistic lens through which to analyze movement phenomena ranging from foraging movements to long-distance migrations, the Movement Ecology Paradigm enables researchers to identify general principles governing organismal movement across taxa and ecosystems.

Core Principles and Conceptual Framework

The Movement Ecology Paradigm is built upon four foundational components that collectively determine movement paths:

• Internal State (Why move?): This component encompasses the physiological, neurological, and psychological drivers that motivate movement, such as hunger, reproductive state, fear, or curiosity. It represents the "why" behind movement decisions, often framed in terms of fulfilling fundamental biological needs.

• Motion Capacity (How to move?): This refers to the biomechanical and physiological mechanisms that enable movement, including anatomical adaptations for flying, swimming, walking, or running. It sets the physical constraints on how an organism can traverse its environment.

• Navigation Capacity (When and where to move?): This involves the sensory, cognitive, and memory capabilities that allow organisms to determine their position relative to targets and navigate through space. It includes abilities like compass orientation, map-based navigation, and cue-based movement.

• External Factors (How does environment influence movement?): These are the environmental variables that affect all other components, including abiotic factors (e.g., topography, wind, temperature) and biotic factors (e.g., resource distribution, predators, competitors) [27] [28].

These components interact continuously to produce the observed movement path of an organism. The paradigm emphasizes that a complete understanding of movement requires investigating all four components and their interactions, rather than focusing on any single element in isolation.

Figure 1: The Movement Ecology Framework depicting the four core components and their interactions leading to a movement path.

Application Notes: Integrating GPS and Accelerometer Data

Empirical Case Studies

The integration of GPS and accelerometer technologies has enabled rigorous testing of Movement Ecology Paradigm predictions across diverse species. The following case studies demonstrate practical applications:

Case Study 1: Lesser Kestrel (Falco naumanni) Foraging Strategies Researchers investigated the foraging behavior of central-place foraging lesser kestrels during breeding season using combined GPS and tri-axial accelerometers [27]. The study revealed that:

• Males devoted more time and energy to flight behaviors than females, corresponding with their role as primary food provisioners for the nest [27].
• During commuting flights, kestrels strategically switched between time-saving flapping flights and energy-saving soaring-gliding flights depending on solar radiation and thermal updraft strength [27].
• During prey searching, kestrels replaced energy-saving perch-hunting with time-saving hovering as wind speed increased, which provided stronger lift [27].
• Notably, kestrels maintained constant energy expenditure per foraging trip despite dramatic changes in flight and hunting strategies, suggesting a fixed energy budget per trip to which they adjusted their behavioral strategies in response to weather conditions [27].

Case Study 2: Gray Wolf (Canis lupus) Movement Energetics A study on gray wolves in interior Alaska utilized ACC-GPS collars to quantify energy expenditure, ranging patterns, and movement ecology [28]. Key findings included:

• Wolves exhibited substantial variability in home range size (500-8300 km²) that was not correlated with daily energy expenditure [28].
• Mean daily energy expenditure and travel distance were 22 MJ and 18 km day⁻¹, respectively [28].
• Wolves spent 20% and 17% more energy during summer pup rearing and autumn recruitment seasons than spring breeding season, regardless of pack reproductive status [28].
• Wolves were predominantly crepuscular but during night spent 2.4× more time engaged in high-energy activities (such as running) during pup rearing season than breeding season [28].
• Heavy precipitation, deep snow, and high ambient temperatures each reduced wolf mobility, demonstrating how abiotic conditions impact movement decisions [28].

Table 1: Quantitative Findings from Movement Ecology Studies Integrating GPS and Accelerometer Technologies

Study Species	Key Behavioral Metrics	Energy Expenditure Findings	Environmental Correlates
Lesser Kestrel (Falco naumanni)	Behavioral compensation between flight strategies; Sex-specific time and energy allocation	Maintained constant energy expenditure per trip despite strategy shifts	Solar radiation, thermal updrafts, wind speed, and air temperature influenced flight and hunting decisions [27]
Gray Wolf (Canis lupus)	Mean daily travel distance: 18 km; Home range: 500-8300 km²	Mean DEE: 22 MJ/day; 20% higher in pup-rearing vs. breeding season	Heavy precipitation, deep snow, and high ambient temperatures reduced mobility [28]

Data Analysis Protocols

Protocol 1: Assessing Bias and Robustness in Social Network Metrics from GPS Telemetry For studies investigating social dynamics within the Movement Ecology framework, a structured protocol has been developed to assess the reliability of social network metrics derived from partial population sampling [29]. This five-step workflow includes:

• Pre-network data permutation: Determine if the network structure captures non-random aspects of association by comparing observed networks to null models generated through data stream permutation [29].
• Bias assessment via sub-sampling: Evaluate how bias in network summary statistics varies with decreasing proportions of sampled individuals [29].
• Bootstrapping for uncertainty estimation: Apply bootstrapping techniques to subsamples of the observed network to quantify how different the network properties would have been if a completely different set of individuals had been tagged [29].
• Node-level metric robustness: Use correlation and regression analyses to assess how node-level network metrics are affected by the proportion of individuals present in the sample [29].
• Node-level confidence intervals: Employ bootstrapping to generate confidence intervals for each node's individual network metric value [29].

This protocol enables statistical comparison of networks under different conditions (e.g., daily and seasonal changes) and guides methodological decisions in animal social network research [29].

Protocol 2: Accelerometer Data Processing for Behavioral and Energetic Metrics The analysis of tri-axial accelerometer data follows a standardized workflow from raw data collection to behavioral and energetic inference:

• Data Collection: Accelerometers measure body acceleration across three spatial axes at high temporal resolutions (typically ≥10 Hz), capturing both static (body orientation) and dynamic (body movement) acceleration components [27].
• Behavioral Classification: Through calibration studies (often with captive animals), distinct behavioral categories are defined based on ODBA (Overall Dynamic Body Acceleration) threshold values [28]. For example, in gray wolves, five behavioral categories were defined: resting (<0.1 g), stationary (0.1<0.25 g), walking (0.25<0.75 g), highly active (0.75<1 g), and running (≥1 g) [28].
• Energy Expenditure Estimation: ODBA values are calibrated against direct measures of oxygen consumption to estimate energy expenditure [27] [28].
• Path Segmentation and Analysis: GPS tracks are segmented into biologically meaningful phases (e.g., commuting vs. foraging) using both spatial coordinates and accelerometer-derived behavioral states [27].
• Environmental Covariate Integration: Environmental variables (e.g., temperature, wind, solar radiation, snow depth) are extracted for each track segment to analyze their influence on movement decisions [27] [28].

Figure 2: Accelerometer data processing workflow from raw data collection to ecological inference.

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 2: Essential Research Materials and Technologies for Movement Ecology Studies

Research Tool	Specifications & Functions	Example Applications in Movement Ecology
GPS Datalogger	GiPSy-5 model (Technosmart); Provides high-resolution location data; Determines movement paths and space use	Tracking commuting routes of lesser kestrels between colony and foraging patches; Quantifying gray wolf home ranges from 500 to 8300 km² [27] [28]
Tri-axial Accelerometer	Axy-3 model (Technosmart); Measures body acceleration in 3 axes; Classifies behaviors and estimates energy expenditure via ODBA	Discriminating between flapping and soaring flights in kestrels; Identifying resting, walking, and running behaviors in wolves [27] [28]
Attachment System	Carbon fiber plate with 4mm wide Teflon ribbon harness; Secures devices to animals while minimizing impact	Deployment on lesser kestrels during breeding season; Total equipment mass maintained below 5% of body mass [27]
Data Processing Software	R package aniSNA; Implements specialized protocols for assessing social network robustness from tracking data	Analyzing bias and uncertainty in social network metrics for ungulate species; Implementing 5-step workflow for data validation [29]
Calibration Equipment	Treadmill with oxygen consumption monitoring; Captive animal facilities for behavioral reference	Establishing ODBA thresholds for wolf behaviors (resting: <0.1g, walking: 0.25-0.75g, running: ≥1g) [28]
Benz[a]azulene	| Benz[a]azulene | | For Research	| High-purity Benz[a]azulene for research (RUO). A non-benzenoid polycyclic aromatic hydrocarbon for organic electronics & photochemistry studies.
Diphenicillin	Diphenicillin, CAS:304-43-8, MF:C21H20N2O4S, MW:396.5 g/mol	Chemical Reagent

The Movement Ecology Paradigm has fundamentally transformed the study of organismal movement by providing a unified framework that integrates internal state, motion capacity, navigation capacity, and external factors. The integration of GPS and accelerometer technologies has been particularly instrumental in advancing this field, enabling researchers to move beyond simple descriptive studies of movement paths to mechanistic understanding of how and why organisms move. As demonstrated by the case studies on lesser kestrels and gray wolves, this approach reveals how animals adjust their movement strategies in response to environmental conditions while balancing energy and time budgets.

Future developments in movement ecology will likely focus on refining analytical protocols for assessing data robustness, particularly when working with partially sampled populations [29]. Additionally, as tracking technologies continue to miniaturize while collecting higher-resolution data, opportunities will expand to apply the Movement Ecology Paradigm across a broader range of species and ecological questions. This progress will further enhance our ability to address critical challenges in conservation, wildlife management, and predicting species responses to environmental change.

From Raw Data to Biological Insight: Analytical Methods and Workflows

Machine Learning Classification of Behavioral States (e.g., Random Forest)

The accurate classification of behavioral states is a cornerstone of behavioral neuroscience, ecology, and precision livestock management. Traditional methods, which often rely on manual scoring or subjective cutoffs, are time-intensive, prone to low inter-rater reliability, and impractical for large datasets [30] [31]. Modern research leverages data from sensors like accelerometers and GPS collars, generating complex datasets that are ideally suited for machine learning (ML) analysis. This protocol details the application of ML models, including Random Forest, for classifying behavioral states from animal sensor data, providing a framework for researchers in drug development and related fields to obtain objective, high-throughput behavioral classifications.

Key Research Reagent Solutions

The following table catalogues essential hardware and software reagents for implementing a behavioral classification pipeline.

Table 1: Essential Research Reagents and Materials for Behavioral Classification Studies

Reagent/Material	Specification/Function
GPS & Accelerometer Collars	Combined sensor units (e.g., LiteTrack Iridium 750+) that collect location (latitude, longitude) and tri-axial movement data (X, Y, Z axes) simultaneously [32].
Field Cameras	Provides ground truth data for labeling behavioral states (e.g., grazing, resting) to train and validate supervised ML models [32].
Computational Environment	Freely available software platforms such as R or Python for implementing ML algorithms and performing statistical analysis [30].
Machine Learning Libraries	R: randomForest, xgboost. Python: scikit-learn, TensorFlow, Keras. Provide pre-built functions for model implementation and training [30] [32].

Quantitative Performance of Classifiers

Multiple ML models have been evaluated for classifying behavioral states. The table below summarizes the performance of various algorithms as reported in recent studies, with Random Forest and XGBoost often demonstrating high accuracy.

Table 2: Classifier Performance on Behavioral State Data

Behavioral Classification Task	Best-Performing Model(s)	Reported Accuracy	Key Predictors
General Foraging Behaviors (Grazing, Ruminating, etc.)	Random Forest (RF), XGBoost (XGB)	RF: Up to 83.9% (Posture); XGB: Up to 74.5% (Activity State) [32]	Speed, Actindex, accelerometer axes (X, Z) [32]
Active vs. Static States	XGBoost (XGB)	74.5% (RTS), 74.2% (CV) [32]	Movement-derived metrics
Posture States (Standing vs. Lying)	Random Forest (RF)	83.9% (CV), 79.4% (RTS) [32]	Accelerometer data (orientation)
Brain States (Slow Oscillation, Microarousal, etc.)	Convolutional Neural Network (CNN)	Up to 97% for high-confidence samples [33]	LFP/EEG features (amplitude, frequency, power spectral density) [33]
Behavioral Phenotyping (Sign-tracking vs. Goal-tracking)	k-Means Clustering, Derivative Method	N/A (Method addresses subjective cutoffs) [31]	Pavlovian Conditioning Approach (PavCA) Index scores [31]

Experimental Protocol for Behavioral Classification

This protocol outlines the key steps for developing an ML pipeline to classify behavioral states from accelerometer and GPS data, using the classification of cattle foraging behavior as a model [32].

Data Acquisition and Preprocessing

• Sensor Deployment: Fit subjects (e.g., cows) with GPS collars integrated with tri-axial accelerometers. Ensure collars are securely fastened with a consistent orientation to maintain data consistency across axes [32].
• Data Collection: Program collars to record GPS locations and accelerometer readings at regular intervals (e.g., every 5 minutes). Concurrently, record video footage of the subjects to serve as ground truth for behavioral labeling [32].
• Behavioral Labeling: Manually review video recordings and label each time interval with the corresponding behavioral state. Common states include:
  • Grazing (GR): Actively consuming forage.
  • Ruminating (RU): Regurgitating and re-chewing food.
  • Walking (W): Purposeful locomotion.
  • Resting (RE): Inactive while standing or lying down.
  • Posture States: Standing up (SU) or Lying down (LD) [32].
• Feature Engineering: Synchronize sensor data with behavioral labels and extract relevant features. Key features often include:
  • Movement Metrics: Speed, distance traveled, heading (from GPS).
  • Activity Index: A composite measure derived from accelerometer data.
  • Accelerometer Axes: Mean, variance, and other statistical measures of the X (left-right), Y (forward-backward), and Z (up-down) axes [32].

Model Training and Evaluation

• Data Partitioning: Split the labeled dataset into training and testing sets. Use either a Random Test-Split (RTS) or Cross-Validation (CV). CV is generally preferred for smaller datasets as it provides a more robust estimate of model performance by repeatedly splitting the data [32].
• Model Selection and Training: Train multiple supervised learning models on the training set for comparison. Common algorithms include:
  • Random Forest (RF): An ensemble of decision trees, robust against overfitting.
  • XGBoost (XGB): A gradient-boosting algorithm known for high performance.
  • Support Vector Machines, Logistic Regression, k-Nearest Neighbors [32].
• Model Evaluation: Use the held-out test set to evaluate model performance. Key metrics include:
  • Overall Accuracy: The percentage of correctly classified instances.
  • Class-specific Accuracy: Accuracy for each behavioral state (e.g., accuracy for classifying "Grazing").
  • Confusion Matrix: To visualize where misclassifications occur [32].

The following workflow diagram illustrates the complete experimental pipeline.

Advanced Applications: Brain State Classification

The principles of ML classification extend beyond gross motor behavior to finer-scale brain states, which is highly relevant for pharmacological and neuroscience research. The following diagram and protocol detail a dual-model approach for classifying brain states from local field potential (LFP) recordings.

Protocol for Brain State Classification using a Dual-Model CNN

This methodology classifies brain states (e.g., during anesthesia) with high confidence [33].

• Signal Acquisition and Feature Extraction: Collect chronic LFP recordings from subjects. Compute dynamic features such as power spectral density (PSD), and characteristics of slow oscillations and microarousals [33].
• Primary Classification with CNN:
  • Process data through two Convolutional Neural Network (CNN) models.
  • CNN Model 1 classifies major states: Awake (AW), Slow Oscillation (SO), and Microarousal (MA).
  • CNN Model 2 further classifies MA into sub-states: slow MA and asynchronous MA [33].
• Confidence-Based Filtering: Assign a confidence level (e.g., 90%) to the CNN predictions. Samples with prediction probability at or above this threshold are automatically assigned the classified state [33].
• Secondary Clustering of Ambiguous Samples: Samples below the confidence threshold are labeled "unknown" and processed by a self-supervised autoencoder-based clustering algorithm. This step reconstructs the samples and clusters them in the frequency domain to provide a final prediction, which is particularly useful for identifying transitions between states [33].

Step Selection Functions (SSFs) and Integrated Step Selection Analysis (iSSA)

Step-selection functions (SSFs) are powerful statistical tools developed to study resource selection and movement decisions of animals by linking sequential spatial data to environmental features. SSFs compare the environmental attributes of observed steps (the linear segments between two consecutive tracked locations) with those of alternative, random steps that an animal could have taken but did not [34]. This matched-case conditional approach allows researchers to model how animals respond to their environment while accounting for inherent movement constraints. The foundational SSF model takes the form w(x) = exp(βx), where the function is proportional to the probability of selecting a step given its environmental characteristics x and selection coefficients β [34].

Integrated step-selection analysis (iSSA) represents a significant methodological advancement that simultaneously estimates movement parameters and habitat selection coefficients within a unified framework [35]. Unlike traditional SSFs that treat movement and habitat selection as separate processes, iSSA incorporates movement characteristics (e.g., step lengths and turning angles) directly into the selection function, thereby bridging the gap between movement mechanics and environmental selection [36] [35]. This integration allows for more biologically realistic models that can simulate space use under novel environmental conditions and quantify landscape resistance [37].

Table 1: Key Components of Step-Selection Analyses

Component	Description	Role in Analysis
Step	Straight-line segment connecting two consecutive observed locations	Fundamental unit of analysis representing a single movement decision
Random Steps	Hypothetical alternative steps generated from movement distributions	Define the "availability" domain and serve as controls for statistical comparison
Movement Kernel	Probability distribution of step lengths and turning angles in neutral landscape	Models intrinsic movement capacity without habitat selection
Selection Kernel	Function modeling environmental preference	Quantifies how habitat features influence movement choices
Integrated SSF	Combined function of movement kernel and selection kernel	Jointly estimates movement and selection parameters [35]

Theoretical Foundations and Methodological Framework

Statistical Foundations of SSFs and iSSA

The theoretical foundation of step-selection analysis rests on weighted distribution theory and the inhomogeneous Poisson point process [36]. In this framework, the probability of observing an animal at a particular location depends on both its movement capabilities and its environmental preferences. The integrated step-selection function takes the form:

u(s_{t+1}) = [φ(s_{t+1}, s_t, s_{t-1}; γ) × w(x(s_{t+1}); β)] / [∫_{s ∈ G} φ(s, s_t, s_{t-1}; γ) w(x(s); β) ds]

where u(s_{t+1}) represents the probability of finding an individual at location s at time t+1, φ is the movement kernel with parameters γ, and w is the habitat-selection function with parameters β [37]. The denominator normalizes this probability to ensure it integrates to 1 over the spatial domain G.

The movement kernel φ is typically composed of distributions for step lengths (distance between consecutive locations) and turning angles (direction changes between successive steps). Commonly used distributions include gamma or exponential distributions for step lengths and von Mises or wrapped Cauchy distributions for turning angles [34]. Recent research has shown that ecological diffusion theory implies a Rayleigh step-length distribution with uniform turning angles, which may be particularly suitable for data collected at irregular time intervals [38].

Integrated Step-Selection Analysis Workflow

The following diagram illustrates the comprehensive workflow for conducting an integrated step-selection analysis:

The iSSA workflow begins with data preparation, where GPS locations are processed to calculate step lengths and turning angles, while simultaneously extracting environmental covariates for each location [36]. Preliminary movement parameters are estimated by fitting distributions to observed step lengths and turning angles, which inform the generation of random steps [35]. The core analytical step involves fitting a conditional logistic regression model where each observed step is matched with multiple random steps, and movement characteristics (e.g., log step length, cosine of turning angle) are included as covariates alongside environmental variables [36] [35]. The coefficients from this model are then used to update the movement parameters, completing the integration of movement and habitat selection.

Practical Implementation Protocols

Data Requirements and Preparation

GPS Data Collection: Modern step-selection analyses typically require high-frequency GPS data, with fix intervals ranging from 15 minutes to 24 hours depending on the research question and species' movement ecology [34]. Data should be collected for a sufficient number of individuals and time periods to capture relevant biological variation. For irregular data resulting from missed fixes, recent methodological advances provide approaches to leverage these data rather than discarding them [37].

Environmental Covariates: Researchers must select appropriate environmental covariates based on ecological hypotheses and species biology. These can include categorical variables (e.g., vegetation type), continuous variables (e.g., elevation, canopy cover), or distance-based measures (e.g., distance to roads or water sources) [34] [39]. Covariates should be prepared as GIS raster layers at resolutions appropriate to the study scale and species' perceptual range.

Table 2: Essential Research Tools for iSSA Implementation

Tool Category	Specific Tools/Software	Application in iSSA
Tracking Technology	GPS loggers, GPS collars	Collect animal movement data at specified intervals
Environmental Data	Remote sensing imagery (Landsat, Sentinel), Digital Elevation Models	Characterize environmental conditions and habitat features
Spatial Analysis	Raster GIS (ArcGIS, QGIS), Spatial processing packages	Process and extract spatial covariates for animal locations
Statistical Analysis	R with amt package, Python with movement libraries	Implement iSSA models and estimate parameters [36]
Movement Visualization	GIS software, R visualization packages (ggplot2, sf)	Visualize movement paths and spatial selection patterns

Step-by-Step Analytical Protocol

Protocol 1: Integrated Step-Selection Analysis Implementation

• Data Preparation and Cleaning

  • Import GPS data and convert to regular time series using functions like track_resample() in the amt package [37]
  • Filter out unrealistic movements and classify into behavioral states if applicable
  • Extract environmental covariates at each GPS location

• Preliminary Movement Analysis

  • Calculate step lengths and turning angles from the cleaned trajectory
  • Fit tentative distributions to step lengths (typically gamma distribution) and turning angles (typically von Mises distribution)
  • Store these distributions as the initial movement kernel φ(s; γ)

• Random Step Generation

  • For each observed step, generate K random steps (typically 10-100) [34]
  • Sample random step lengths and turning angles from the distributions estimated in Step 2
  • Ensure random steps maintain the same starting point as the observed step

• Covariate Extraction

  • Extract environmental covariates for the ending points of both observed and random steps
  • Calculate movement characteristics (log step length, cosine of turning angle) for all steps

• Model Fitting

  • Prepare data for conditional logistic regression with strata defined by each step and its associated random steps
  • Fit model with environmental covariates and movement characteristics using the following general form: w × φ = exp(β₁X₁ + ... + βₖXₖ + α₁log(SL) + α₂cos(TA) + ...)
  • Include interactions between movement parameters and environmental covariates if testing context-dependent movement

• Parameter Interpretation and Model Validation

  • Interpret exponentiated coefficients as relative selection strengths [36]
  • Validate model using cross-validation or by simulating movements and comparing to observed patterns
  • Use fitted model for prediction, such as generating utilization distributions or identifying movement corridors

Protocol 2: Handling Irregular Data in iSSA

For datasets with missing fixes or irregular sampling intervals:

• Approach Selection

  • Choose among four alternative approaches: imputation, naïve scaling, dynamic modeling, or hybrid methods [37]
  • Consider the proportion of missing data and potential biases when selecting an approach

• Imputation Method

  • Fit continuous-time correlated random walk movement model to the observed data
  • Use the fitted model (e.g., via the crawl package in R) to impute missing locations [37]
  • Proceed with standard iSSA on the completed regular trajectory

• Naïve Scaling Approach

  • For each observed step with duration Δt, generate random steps by sampling step speeds and turning angles
  • Calculate random step lengths as speed × Δt
  • Include step duration as a covariate in the iSSA model to account for temporal irregularity

• Dynamic Modeling

  • Group steps by their duration and fit separate movement distributions to each group
  • Sample random steps from the appropriate duration-specific distribution
  • This approach acknowledges potentially non-linear relationships between step duration and movement parameters

Advanced Applications and Extensions

Modeling Individual Variability

Advanced iSSA implementations can incorporate random effects to account for individual variability in both movement parameters and habitat selection [40]. This approach recognizes that animals within a population may exhibit different movement strategies and habitat preferences due to factors such as personality, experience, or competitive status. Implementing random effects in iSSA requires specialized software or custom programming, but provides more robust inference about population-level processes while accounting for inter-individual differences [40].

Applications in Disease Ecology

Step-selection analysis has been successfully adapted for studying human movement in infectious disease epidemiology. A recent study on leptospirosis in urban slums of Brazil used SSFs to analyze how fine-scale movements influence exposure to environmental pathogens [39]. Researchers collected GPS data from 128 participants with locations recorded every 35 seconds during active daytime hours, then used SSFs to estimate selection coefficients for environmental features like open sewers and domestic rubbish piles. The analysis revealed gender-based differences in movement patterns, with women moving closer to central streams but farther from open sewers compared to men [39].

Multi-Scale and Behavioral State Applications

iSSA can be extended to analyze resource selection at multiple spatial and temporal scales, and to incorporate behavioral state classification [34]. By including interactions between movement parameters and environmental covariates, researchers can model how animals adjust their movement strategies in response to landscape features. Additionally, iSSA can be integrated with state-space models to classify behavioral states (e.g., foraging, resting, traveling) when estimating selection parameters, providing more mechanistic understanding of animal decision-making [34].

Table 3: Methodological Considerations in iSSA Implementation

Analytical Decision	Options	Recommendations
Number of Random Steps	2 to 200 per observed step [34]	10-20 provides good balance between computational efficiency and statistical power
Temporal Resolution	15 minutes to 24 hours [34]	Match to natural decision-making rhythm of study species and research question
Covariate Measurement	Endpoint vs. along-step assessment [34]	Endpoint sufficient for most applications; along-step for linear features or detailed path selection
Handling Irregular Data	Burst filtering, imputation, scaling, dynamic modeling [37]	Dynamic modeling preferred when sufficient data; imputation for low to moderate missingness
Random Effects Structure	Random intercepts vs. random slopes [40]	Include random slopes for key habitat covariates when individuals show differential selection

Analytical Framework and Interpretation

The following diagram illustrates the conceptual framework of integrated step-selection analysis, showing how movement mechanisms and habitat selection interact to shape space use patterns:

Parameter Interpretation: In iSSA, parameters are interpreted as relative selection strengths when exponentiated [36]. For continuous environmental covariates, exp(β) represents how many times more likely an animal is to select a location with a one-unit increase in that covariate, assuming all other factors are equal. For categorical covariates, exp(β) indicates the relative selection strength for that category compared to the reference category. Movement parameters (e.g., coefficients on log step length or cosine of turning angle) describe how movement characteristics influence transition probabilities between locations [36] [35].

Model Validation: Essential validation procedures include cross-validation to assess predictive performance, residual analysis to check model assumptions, and simulation-based validation where movements are simulated from the fitted model and compared to observed patterns [35]. When working with iSSAs, it's particularly important to validate that the integrated model can reproduce key features of the observed movement trajectories and space use patterns.

Integrated step-selection analysis represents a sophisticated framework for understanding animal movement and habitat selection as integrated processes. By simultaneously modeling movement mechanisms and environmental selection, iSSA provides a powerful approach for addressing fundamental questions in movement ecology and generating realistic predictions of space use under changing environmental conditions.

Spatio-Temporal Point Process Models (ST-PPMs) for Habitat Use

Spatio-Temporal Point Process Models (ST-PPMs) represent a sophisticated statistical framework for analyzing animal tracking data to infer habitat selection and space use patterns. These models naturally integrate spatial and temporal autocorrelation structures inherent in movement data while rigorously accounting for observer effort and habitat availability [41] [42]. Within the broader context of GPS tracking and accelerometer data analysis research, ST-PPMs provide a powerful approach for quantifying how internal and external factors influence animal movement decisions across multiple scales. This protocol details the implementation of ST-PPMs for habitat use studies, including data requirements, model specifications, validation procedures, and interpretation guidelines tailored for researchers in movement ecology and wildlife conservation.

The analysis of animal movement has been transformed by advanced biologging technologies that generate high-throughput GPS and accelerometer data [8] [43]. Spatio-Temporal Point Process Models (ST-PPMs) have emerged as a statistically robust method for analyzing such data, particularly for inferring habitat selection and large-scale attraction/avoidance behaviors [41] [42]. Unlike traditional methods that treat autocorrelation as a nuisance, ST-PPMs explicitly incorporate spatio-temporal dependencies, providing more accurate estimates of resource selection [42].

ST-PPMs belong to a class of methods that utilize "pseudo-absences" or "dummy points" to quantify habitat availability, but they provide a mathematical foundation for determining the optimal number and location of these points [42]. This framework generalizes many earlier approaches and can be implemented using standard generalized linear modeling software while appropriately accounting for autocorrelation structures [42]. Comparative studies have demonstrated that ST-PPMs maintain nominal Type I error rates across various scenarios, outperforming spatial logistic regression models (SLRMs) and showing comparable performance to integrated step selection models (iSSMs) [41] [42].

Theoretical Foundations

Mathematical Framework

Spatio-Temporal Point Process Models characterize the intensity function λ(s,t) representing the expected number of animal locations per unit area per unit time at spatial coordinate s and time t [42] [44]. For habitat selection studies, this intensity is typically modeled as:

λ(s,t) = exp[β₀ + ΣβᵢXᵢ(s,t) + ε(s,t)]

where Xᵢ(s,t) are spatio-temporal covariates, βᵢ are selection coefficients, and ε(s,t) represents spatio-temporal random effects that capture autocorrelation [42].

The likelihood function for ST-PPMs is formulated as:

L(β) = Πλ(sᵢ,tᵢ) exp[-∫∫λ(s,t)dsdt]

This likelihood can be approximated using a Poisson regression with sufficiently many dummy points, enabling implementation with standard statistical software [42] [44].

Comparative Method Performance

ST-PPMs demonstrate distinct advantages over alternative approaches for analyzing animal tracking data:

Table 1: Comparison of Statistical Methods for Animal Tracking Data Analysis

Method	Type I Error Rate	Statistical Power	Handling of Autocorrelation	Implementation Complexity
ST-PPM	Nominal [42]	Moderate to High [42]	Explicit modeling [42]	Moderate [42]
iSSM	Nominal [42]	High [42]	Used in stratification [42]	High [42]
SSM	Slightly inflated [42]	Moderate [42]	Used in stratification [42]	Moderate [42]
SLRM	Frequently exceeded [42]	Variable [42]	Often neglected [42]	Low [42]

Experimental Protocols

Data Collection Requirements

Implementing ST-PPMs requires carefully collected animal tracking data with the following specifications:

Table 2: Data Collection Specifications for ST-PPM Analysis

Parameter	Minimum Specification	Optimal Specification	Notes
GPS Fix Rate	Every 5 minutes [21]	1-30 seconds [43]	Balance battery life with resolution [21]
Accelerometer Sampling	10 Hz [21]	25 Hz [45]	For behavior classification [21] [45]
Tracking Duration	2 weeks [21]	Several months [45]	Capture relevant behavioral cycles
Number of Individuals	10-15 [21]	30+ [21]	Account for individual variation
Location Error	<10m [21]	<5m [21]	GPS DOP threshold of 1 recommended [21]

Pre-processing Pipeline

Raw tracking data must be cleaned and processed before ST-PPM analysis:

• Data Cleaning: Remove positional outliers using speed filters and movement-based algorithms [43]. Automated pipelines like atlastools R package can efficiently process large datasets [43].

• Sensor Integration: Merge GPS locations with accelerometer-derived behaviors (e.g., grazing, ruminating, traveling) using synchronized timestamps [21] [45].

• Covariate Extraction: Extract environmental covariates (vegetation, topography, human infrastructure) at each GPS fix and available locations [41] [42].

• Dummy Point Generation: Create pseudo-absence points following ST-PPM specifications to represent available habitat [42].

ST-PPM Implementation

The core implementation of Spatio-Temporal Point Process Models follows these steps:

• Intensity Function Specification: Define the base intensity function incorporating spatial, temporal, and environmental covariates [42] [44].

• Autocorrelation Structure: Incorporate spatio-temporal random effects using Gaussian processes or spline-based smoothers [42].

• Parameter Estimation: Fit the model using maximum likelihood or Bayesian approaches with appropriate computational techniques [42] [44].

• Model Selection: Use information-theoretic approaches (AIC, BIC) or cross-validation to select optimal covariate combinations [42].

• Validation: Assess model fit using residual analysis and goodness-of-fit tests specific to point processes [42] [44].

The Scientist's Toolkit

Table 3: Essential Research Reagents and Tools for ST-PPM Studies

Tool/Reagent	Specification	Function/Purpose	Example Sources
GPS Collars	3-D accelerometer (10-25 Hz), GPS with 5m accuracy [21]	Movement and location data collection [21]	Digitanimal, Wildbyte Technologies [21] [11]
Data Loggers	MEMS accelerometers, temperature range -40° to 85°C [21]	Fine-scale movement recording [21] [13]	Technosmart, Daily Diary tags [11] [13]
Calibration Equipment	Tilt platforms, motion rate tables [11]	Accelerometer calibration for accuracy [11]	Custom laboratory setups [11]
Tracking Software	ATLAS systems, GPS data loggers [43]	High-throughput data collection [43]	Wadden Sea ATLAS [43]
Analysis Packages	R packages: atlastools, spatstat [43] [42]	Data cleaning and ST-PPM implementation [43] [42]	CRAN, GitHub repositories [43]
Computer Vision Tools	AlphaTracker, DeepLabCut [46]	Multi-animal tracking and behavior analysis [46]	Open-source platforms [46]
Peptide M acetate	Peptide M acetate, MF:C83H145N21O33, MW:1965.2 g/mol	Chemical Reagent	Bench Chemicals
Hemiphroside A	Hemiphroside A, MF:C31H40O16, MW:668.6 g/mol	Chemical Reagent	Bench Chemicals

Applications and Case Studies

Cattle Behavior and Habitat Use

A study monitoring 30 beef cattle equipped with 3-D accelerometers and GPS sensors demonstrated ST-PPM applications for classifying behavioral patterns (grazing, ruminating, laying, steady standing) with high accuracy (0.93 for grazing) [21]. GPS data sampled every 5 minutes was analyzed via k-medoids clustering to track herd spatial scatter, while accelerometer data enabled behavior classification through random forest algorithms [21]. This integrated approach allowed monitoring of sustainable pasture consumption and detection of anomalous events in small and mid-size farms [21].

Wildlife Conservation and Management

ST-PPMs have been applied to estimate effort-corrected space use of endangered Southern Resident Killer Whales by combining multiple datasets [44]. The framework enabled integration of sightings data from citizen scientists with different observation protocols while controlling for unknown observer effort [44]. This application highlights how ST-PPMs can leverage heterogeneous data sources to inform conservation strategies for highly mobile species.

Integration with Accelerometer-Derived Behaviors

A study on Pacific Black Ducks implemented continuous on-board processing of accelerometer data to classify eight distinct behaviors [45]. When combined with GPS data, this behavioral information enabled more nuanced habitat use analysis through ST-PPMs, revealing how specific sites within home ranges satisfied particular behavioral needs (e.g., roosting, foraging) [45]. The continuous behavior records significantly improved accuracy of time-activity budgets compared to interval sampling, particularly for rare behaviors [45].

Advanced Applications

Multi-Scale Habitat Selection

ST-PPMs can simultaneously model habitat selection at different spatial scales, from fine-scale resource selection to landscape-level attraction/avoidance [41] [42]. This multi-scale capability is particularly valuable for understanding how animals respond to environmental features across hierarchical levels of organization.

Population-Level Inference

While originally developed for individual-level analysis, ST-PPMs can be extended to population-level inference by incorporating hierarchical structures that share information across individuals while accounting for individual heterogeneity [42] [44].

Dynamic Habitat Use

Temporal variation in habitat selection can be modeled by including time-varying coefficients in the intensity function, allowing researchers to investigate how habitat use changes with diel cycles, seasons, or in response to anthropogenic disturbances [42] [45].

Spatio-Temporal Point Process Models provide a powerful, statistically robust framework for analyzing animal habitat use from GPS and accelerometer data. Their ability to explicitly incorporate spatio-temporal autocorrelation while rigorously accounting for sampling effort addresses key limitations of earlier methods. As biologging technologies continue to generate increasingly detailed movement data, ST-PPMs offer a flexible approach for understanding the complex interactions between animals and their environments across multiple scales. The protocols outlined in this document provide researchers with practical guidance for implementing ST-PPMs in diverse ecological contexts, from wildlife conservation to livestock management.

Trajectory Segmentation and Behavioral State Identification

The analysis of animal movement through trajectory segmentation and behavioral state identification represents a critical methodology in movement ecology and related fields. Modern tracking technologies, including GPS collars and accelerometers, now provide high-resolution location data, sometimes recorded second-by-second, enabling an unprecedented detailed view of animal movement [47]. This advancement has created a pressing need for sophisticated analytical methods that can parse these complex datasets into biologically meaningful segments and identify underlying behavioral states.

The ability to accurately distinguish between behaviors such as foraging, resting, traveling, and predation events from movement data alone provides powerful insights into animal ecology, energetics, and responses to environmental change [47] [48]. For researchers in drug development and behavioral neuroscience, these methods offer objective, quantitative tools for assessing animal behavior in both field and laboratory settings, with applications ranging from understanding drug effects to evaluating welfare indicators [49] [50].

This document presents application notes and protocols for implementing trajectory segmentation and behavioral state identification, framed within the broader context of GPS tracking and accelerometer data analysis in animal movement research.

Core Concepts and Methodological Approaches

Fundamental Movement Parameters

Animal movement paths are conceptualized as temporal sequences of locations from which fundamental movement parameters are derived. These parameters serve as proxies for inferring behavioral states [51] [52]. The table below summarizes key movement parameters used in behavioral analysis.

Table 1: Fundamental Movement Parameters Derived from Tracking Data

Parameter	Description	Behavioral Significance
Step Length	Straight-line distance between successive locations	Indicates speed of movement; longer steps suggest traveling [53] [52]
Turning Angle	Change in direction between successive movement steps	High variance suggests area-restricted search (e.g., foraging); low variance suggests directed movement [47] [52]
Persistence Velocity	Composite measure incorporating speed and directionality	Distinguishes between tortuous and directed movement [52]
Residence Time	Time spent in a localized area	May indicate foraging, resting, or resource use [53]
Straightness Index	Net displacement divided by total path length	Measures efficiency of movement between points [53]

Path segmentation methods dissect movement trajectories into segments assumed to reflect different underlying behaviors [53]. These methods can be broadly categorized according to their primary analytical approach:

• Change-Point Analysis Methods: Identify significant statistical changes in movement parameters along a trajectory. Behavioral Change Point Analysis (BCPA) explicitly models temporal autocorrelation to detect significant changes in parameters like persistence velocity and turning angle [52].

• State-Space Models: Infer unobserved behavioral states from observed location data while accounting for measurement error. The First-Difference Correlated Random Walk with Switching (DCRWS) uses a Bayesian framework to estimate behavioral states [54].

• Hidden Markov Models (HMMs): Assume movement observations depend on an unobserved Markov process representing behavioral states. The Hidden Markov Movement Model (HMMM) implements this approach using maximum likelihood estimation for rapid model fitting [54].

• Clustering-Based Methods: Group similar path segments according to their movement parameters using clustering algorithms such as k-medoids or hierarchical clustering [10] [52].

Experimental Protocols

Data Collection and Sensor Configuration

Objective: To collect high-resolution movement data suitable for trajectory segmentation and behavioral state identification.

Table 2: Sensor Configuration Protocols for Animal Tracking

Sensor Type	Recommended Specifications	Configuration Notes
GPS Receiver	Fix intervals: 1 second to 5 minutes depending on research question	Shorter intervals for fine-scale behavior; longer intervals for broader patterns [47] [10]
Tri-axial Accelerometer	Sampling frequency: 10-25 Hz; Dynamic range: ±2g to ±8g	Higher frequencies capture more detailed movements; orientation should be consistent [48] [10]
Data Logging	Sufficient memory for entire deployment period; Secure Digital (SD) cards recommended	Raw data storage preferred over pre-processed summaries [10]
Power Management	Battery life adequate for deployment duration; solar options for long-term studies	GPS typically most power-intensive; sampling intervals affect battery life [10]

Procedure:

• Sensor Deployment: Fit sensors securely using collars, harnesses, or attachments appropriate to the study species. Ensure minimal interference with natural behavior.
• Synchronization: Synchronize internal clocks of all sensors to a common time standard.
• Data Validation: Conduct focal animal observations during initial deployment to validate sensor readings against observed behaviors [48].
• Data Retrieval: Download raw data and perform initial quality checks for missing values or sensor malfunctions.

Behavioral Validation Using Captive Observations

Objective: To establish ground-truthed correlations between sensor data and specific behaviors.

Materials:

• Animals in controlled environments (captive facilities or seminatural enclosures)
• Video recording system synchronized with sensor data collection
• Behavioral coding software (e.g., BORIS, DeepLabCut) [50]

Procedure:

• Simultaneous Data Collection: Record video while collecting GPS and accelerometer data from captive subjects.
• Behavioral Coding: Systematically code video recordings to identify specific behaviors (e.g., feeding, resting, locomotion) and their precise timing [48].
• Threshold Determination: For accelerometer data, use statistical methods (e.g., Welch's t-test) to establish activity thresholds distinguishing active vs. inactive states [48].
• Model Training: Use matched sensor-behavior data to train supervised classification models such as Random Forest classifiers [10].

Trajectory Segmentation Using BCPA

Objective: To partition movement paths into segments with homogeneous movement characteristics.

Software Requirements: R statistical environment with bcpa package [52].

Procedure:

• Data Preparation: Import GPS data and calculate movement parameters (step lengths, turning angles, persistence velocity).
• Parameter Selection: Choose the movement parameter of interest for segmentation (persistence velocity recommended for initial analysis).
• Model Fitting: Apply BCPA to identify significant change points in the movement parameter time series.
• Segment Extraction: Divide the trajectory at identified change points.
• Validation: Compare segments with ground-truthed behavioral data when available.

Behavioral State Classification Using HMMs

Objective: To identify discrete behavioral states from movement data.

Software Requirements: R with moveHMM or momentuHMM packages [54].

Procedure:

• Data Preparation: Format movement data as a series of step lengths and turning angles.
• Initialization: Specify initial parameter values for the HMM based on expected behaviors (e.g., short steps with high turning variance for foraging).
• Model Fitting: Estimate model parameters using maximum likelihood estimation.
• State Decoding: Apply the Viterbi algorithm to determine the most likely sequence of behavioral states.
• Interpretation: Label decoded states according to validated behavioral definitions.

Data Analysis and Interpretation

Quantitative Comparison of Segmentation Methods

Table 3: Performance Characteristics of Segmentation Methods

Method	Statistical Approach	Handles Autocorrelation	Computational Demand	Best Application Context
BCPA	Likelihood-ratio based change point detection	Yes [52]	Moderate	Single-parameter segmentation; high-frequency data [52]
HMM	Maximum likelihood or Bayesian estimation	Yes [54]	Low to moderate	Multiple behavioral states; regularly sampled data [54]
State-Space Models	Bayesian estimation (MCMC)	Yes [54]	High	Data with significant measurement error [54]
Clustering-Based	Distance-based clustering of path segments	No	Low	Exploratory analysis; distinct movement modes [10] [52]

Integration of GPS and Accelerometer Data

Objective: To improve behavioral classification by combining location and acceleration data.

Procedure:

• Temporal Alignment: Precisely align GPS and accelerometer data timestamps.
• Feature Extraction: From accelerometer data, extract features in both time and frequency domains (e.g., variance, spectral entropy) [10].
• Data Fusion: Integrate GPS-derived movement parameters with accelerometer features.
• Model Training: Train machine learning classifiers (e.g., Random Forest) on the combined feature set.
• Performance Assessment: Compare classification accuracy with and without accelerometer features.

Research indicates that including accelerometer data can improve correct assignment of behaviorally significant sites (e.g., predation events) by 5-38% compared to GPS data alone [48].

The Scientist's Toolkit

Essential Research Reagents and Solutions

Table 4: Essential Materials for Animal Movement Studies

Item	Specifications	Primary Function
GPS Collars	Lotek 7000 series or similar; with onboard accelerometer capability [48]	Records animal locations at programmed intervals; provides fundamental movement data
Tri-axial Accelerometers	MEMS technology; 10-25 Hz sampling frequency; ±2g to ±8g range [10]	Measures fine-scale movements and body orientation; critical for behavior identification
Data Loggers	Secure Digital (SD) card storage; weatherproof casing [10]	Stores raw sensor data for subsequent analysis
Video Recording System	Multiple cameras; infrared capability for nighttime recording; time-synchronization [50]	Provides ground-truth data for validating automated behavior classifications
R Statistical Software	with packages: adehabitatLT, bcpa, moveHMM, momentuHMM [51] [52]	Primary platform for trajectory analysis and segmentation
Mazdutide	Mazdutide (IBI362)	Mazdutide is a dual GLP-1 and glucagon receptor agonist for research into obesity, diabetes, and metabolic health. For Research Use Only.
2,3-Benzodioxine	2,3-Benzodioxine|CAS 253-57-6|C8H6O2	2,3-Benzodioxine (CAS 253-57-6) is a benzodioxine derivative for research. This product is For Research Use Only (RUO). Not for human or veterinary use.

Workflow Visualization

Experimental Workflow for Behavioral State Identification

Trajectory Segmentation Analytical Process

Trajectory segmentation and behavioral state identification represent powerful approaches for extracting biologically meaningful information from animal movement data. The integration of GPS technology with accelerometer data significantly enhances our ability to distinguish subtle behavioral states, providing insights that were previously inaccessible through direct observation alone [47] [48].

As tracking technology continues to advance, providing ever-higher resolution data, the development of sophisticated analytical methods will remain crucial for fully leveraging these rich datasets. The protocols outlined here provide researchers with practical tools for implementing these methods across diverse study systems and research questions, from basic movement ecology to applied drug development studies assessing behavioral responses to pharmacological interventions.

The fields of ecology, wildlife management, and conservation biology are experiencing a revolution driven by technological advancements in bio-logging and animal tracking. Researchers can now document animal behavior and ecology in unprecedented detail and extent, generating complex datasets that describe movement across multiple spatiotemporal scales [3]. However, this data deluge presents a significant analytical challenge; the ability to fully exploit the rich information contained within tracking datasets often lags behind our capacity to collect it [55]. The sheer volume, variety, veracity, and velocity of this data mean its analysis often exceeds the capacity of conventional methods and systems. This creates a dependency on computational experts, potentially leaving experienced field biologists and wildlife managers without the tools to directly interrogate their own data. Platforms like MoveApps have been developed specifically to bridge this gap, making sophisticated analytical tools accessible to a global community of researchers regardless of their coding expertise [55]. By providing an intuitive, web-based interface for designing and executing analytical workflows, such platforms empower a broader range of scientists to contribute to and benefit from the latest methodological developments in movement ecology.

The MoveApps Platform: Core Architecture and Features

MoveApps is an open analysis platform designed to enable the analysis of animal tracking data through a serverless, no-code cloud computing system. Its core design philosophy is modularity, allowing users to build complex analyses by connecting simple, single-function building blocks called "Apps" [55]. This architecture maximizes flexibility while minimizing the complexity and potential for errors within any single component. The platform is built using widely adopted open-source tools and languages, with the backend programmed in Kotlin and Java. A key technical innovation is the implementation of Apps as Docker containers instead of virtual machines. Containers share an underlying host operating system (Linux GNU), making them faster and less resource-intensive, which is ideal for a platform hosting many small, independently developed analytical modules [55]. This container-based approach, orchestrated by Kubernetes, ensures that each App runs in an isolated environment with its own defined programming language, version, and supporting software packages. This isolation is crucial for long-term reproducibility, as it protects analyses from cascading errors caused by updates or changes in interdependent software libraries.

Table 1: Key Features and Benefits of the MoveApps Platform

Feature	Description	Primary Benefit
Modular Workflow Design	Analysis built by linking single-function Apps in a web-based interface [55]	No coding skills required; intuitive and flexible analysis design
Serverless Cloud Computing	Platform runs on a cloud system independent of user hardware [55]	No local installation; access from anywhere; scalable computing power
Containerized Apps (Docker)	Each App runs in an isolated software environment [55]	Long-term reproducibility and stability of analyses
Integration with Movebank	Directly accesses and utilizes data from the Movebank repository [55]	Streamlined data management and seamless analysis of existing datasets
Open & Shareable Workflows	Workflows can be shared, published, and archived with DOIs [55]	Promotes collaboration, open science, and methodological replication

The process of using MoveApps is designed to be intuitive. Users can browse a library of available Apps, drag and drop them to create a workflow, customize parameters, execute the analysis, and access results entirely through a web browser. This process democratizes access to advanced methods that would otherwise require significant programming proficiency in languages like R. Furthermore, the platform fosters a collaborative ecosystem. Developers can contribute new Apps via public Git repositories, making their analytical code available to the entire community. By bringing together analytical experts who develop methods and the data collectors who need them, MoveApps increases the pace of knowledge generation to match the rapid growth in bio-logging data acquisition [55].

Protocols for Integrated GPS and Accelerometer Data Analysis

The integration of GPS and accelerometer data is a powerful approach in movement ecology, as the strengths of one sensor can compensate for the weaknesses of the other. The following protocols outline methodologies for processing and integrating these complementary data streams.

Protocol 1: Processing Accelerometer Data for Dynamic Displacement

Objective: To calculate high-frequency dynamic displacement from raw accelerometer data, overcoming the inherent drift and noise of inertial sensors.

Materials & Methods:

• Input Data: Tri-axial accelerometer data collected at a high frequency (e.g., 50 Hz).
• Processing Software: A computational environment capable of signal processing (e.g., R, Python).
• Key Steps:
  • Noise Filtering: Remove high-frequency noise and low-frequency drift using a band-pass filter. The Finite Impulse Response (FIR) filter is a widely recommended method for this purpose due to its stability and simple structure [56]. The filter's passband should be selected based on the expected frequency range of the animal's behaviors of interest.
  • Double Integration: Convert the filtered acceleration-time data to displacement-time data via double integration using the formula: s(t) = s₀ + v₀ × t + ∫₀ᵗ (∫₀ᵗ a(t)dt)dt where s(t) is displacement at time t, a(t) is acceleration, sâ‚€ is the initial position, and vâ‚€ is the initial velocity [56].
  • Initial Value Problem Resolution:
    • The initial position (sâ‚€) is typically set to zero for dynamic displacement, with the resulting time series later adjusted to align with static or quasi-static displacements measured by GPS.
    • The initial velocity (vâ‚€) is often set to zero, and any resulting linear trend in the displacement time series is subsequently estimated and removed (detrending) [56].

Considerations: This method yields precise dynamic displacements but is generally unsuitable for measuring static or quasi-static displacements, as these are removed during the high-pass filtering stage [56].

Protocol 2: Sensor Fusion for Comprehensive Movement Analysis

Objective: To synergistically combine GPS and accelerometer data to obtain a complete picture of animal movement, including static, quasi-static, and dynamic displacements.

Materials & Methods:

• Input Data: GPS location data and tri-axial accelerometer data, preferably time-synchronized.
• Processing Software: A platform capable of handling both data types (e.g., R, MoveApps).
• Key Steps:
  • Data Synchronization and Alignment: Precisely align the GPS and accelerometer data streams using their timestamps.
  • Frequency-Based Data Extraction: Fuse the data by leveraging the strengths of each sensor.
    • Use the GPS track to define the static and quasi-static displacements of the animal's trajectory. GPS is well-suited for this as it provides absolute positional information, though it can be noisy at high frequencies [56].
    • Use the accelerometer-derived dynamic displacements to add high-frequency, fine-scale movements (e.g., wingbeats, head movements, step-by-step gait dynamics) to the GPS-defined path [56] [57].
  • Behavioral Classification: Use the combined data to improve activity recognition. For example, GPS-provided speed and elevation change (from a Digital Elevation Model) can be input into a machine learning classifier (e.g., Random Forest) alongside accelerometer features to distinguish between behaviors like level walking, uphill walking, and running with greater accuracy than using accelerometer data alone [57].

Considerations: This integrated approach provides a more robust and comprehensive analysis of animal movement, enabling researchers to connect fine-scale behaviors with larger-scale movement paths.

Essential Research Reagents and Tools for Movement Analysis

A modern movement ecology study relies on a suite of hardware and software "reagents" to collect, manage, and analyze data. The table below details key solutions required for a successful research program.

Table 2: Key Research Reagent Solutions for Animal Movement Analysis

Tool / Solution	Type	Primary Function	Example/Note
GPS Bio-logger	Hardware	Records animal location and trajectory over time [3].	Tags (e.g., GPS/Argos) vary in weight, accuracy, and data retrieval method (store-on-board vs. satellite transmit).
Tri-axial Accelerometer	Hardware	Senses body posture, motion, and behavior-specific signatures [57].	Often integrated into bio-loggers; samples at high frequencies (e.g., 50 Hz) to capture fine-scale movement.
Movebank	Data Repository	Stores, manages, and standardizes animal tracking data from many sources [55].	Provides vital data curation, harmonization, and sharing protocols.
MoveApps	Analysis Platform	No-code, cloud-based platform for designing and executing analytical workflows [55].	Enables reproducible analysis via a library of containerized Apps.
R move/move2	Software Package	Open-source R packages for the statistical analysis of animal movement data [58].	Provides a flexible, code-based environment for advanced statistical modeling and analysis.
Digital Elevation Model (DEM)	Data Layer	Provides topographic (elevation) information for the landscape [57].	Used to derive elevation change from GPS tracks, aiding behavioral classification.

The development of accessible analytical platforms like MoveApps represents a pivotal advancement for movement ecology and related fields. By abstracting away complex computational infrastructure and providing an intuitive interface for state-of-the-art methods, these platforms directly address the critical bottleneck between data collection and knowledge extraction. They empower a broader community of researchers and wildlife managers to engage in sophisticated data analysis, thereby accelerating the pace of discovery. When combined with robust protocols for processing and fusing multi-sensor data—such as integrating GPS for macro-scale movement and accelerometry for micro-scale behavior—these tools enable a more holistic and mechanistic understanding of animal movement. As the volume and complexity of bio-logging data continue to grow, the role of such integrated, reproducible, and accessible analysis platforms will only become more central to ecological research, wildlife conservation, and the study of animal behavior.

Navigating Pitfalls: Ensuring Data Accuracy and Methodological Rigor

Critical Protocols for Accelerometer Calibration and Validation

Within the field of animal movement analysis, accelerometer data has become a cornerstone for quantifying behaviour, energy expenditure, and movement ecology [13]. However, the scientific robustness of inferences drawn from this data is entirely contingent upon rigorous sensor calibration and behavioural validation protocols. Uncalibrated sensors and unvalidated behavioural predictions can introduce significant error, leading to erroneous ecological conclusions [11]. These Application Notes provide detailed methodologies to ensure the data quality and interpretive accuracy essential for research in ecology, conservation, and related life sciences.

The Critical Need for Calibration and Validation

Accelerometers measure proper acceleration, comprising static (gravity) and dynamic (movement) components. Inaccurate calibration directly impacts derived metrics such as Dynamic Body Acceleration (DBA), a common proxy for energy expenditure, with errors of up to 5% reported between calibrated and uncalibrated tags [11]. Furthermore, the placement of the tag on the animal's body (e.g., back versus tail) can cause variation in DBA values exceeding 10%, which can be misconstrued as biological variation [11].

Validation is equally critical when classifying animal behaviour from acceleration signals. Machine learning models, such as Random Forest classifiers, can achieve high overall accuracy (F-measure up to 0.96), but their performance is highly dependent on the quality of the training data and the specific behaviours being identified [59]. For instance, slow, aperiodic behaviours like grooming are often misidentified, whereas locomotory behaviours are classified with higher reliability [59]. Without field validation, these misclassifications can go undetected, compromising the study's findings.

Accelerometer Calibration Protocols

The Six-Orientation Static Calibration (6-O) Method

This method calibrates the sensor against the known constant of Earth's gravity.

Experimental Principle and Workflow

The core principle is to collect static acceleration readings with each of the three sensor axes aligned parallel and anti-parallel to the gravitational field. The vector sum of the three acceleration axes for a perfectly calibrated sensor at rest will be 1.0 g [11]. Deviations from this value indicate measurement error that requires correction.

The following diagram illustrates the experimental workflow:

Required Materials

Table 1: Research Reagent Solutions for Accelerometer Calibration

Item	Specification	Function
Tri-axial Accelerometer	Animal-borne data logger (e.g., Daily Diary tag)	The sensor unit to be calibrated.
Calibration Platform	A level, stable surface	Provides a reference plane aligned with gravity.
Data Logging System	Microcontroller (e.g., Arduino) with serial output	Records raw accelerometer measurements [60].
Calibration Software	Custom script (e.g., record-data.py [60])	Automates data collection and computes calibration parameters.

Step-by-Step Protocol

• Mount the Sensor: Securely fix the accelerometer to a rigid plate or breadboard. The squared edges of the mounting platform provide reference planes for consistent orientation [61].
• Initialize Data Logging: Execute the data logging script (e.g., record-data.py) to begin capturing raw, comma-separated accelerometer measurements from the serial port [60].
• Collect Orientation Data:
  • Place the mounted sensor in the first orientation (e.g., with the X-axis pointing straight down). Ensure it is perfectly stationary.
  • In the logging software, trigger a measurement. Hold the sensor steady for at least 10 seconds to collect a sufficient static sample [11].
  • Repeat this process for all six orientations, ensuring that for each axis, data is collected for both the +1g and -1g positions.
• Compute Calibration Parameters: Use the collected data to calculate two correction factors for each axis [11]:
  • Bias Correction: Adjusts the raw values so that the minimum and maximum for each axis are symmetric around zero.
  • Gain Correction: Scales the bias-corrected values so that the vector sum of the three axes equals 1.0 g when stationary.
• Apply and Visualize: Integrate the final calibration parameters (the A^-1 matrix and bias vector) into your data processing pipeline. Use plotting scripts to visualize the calibrated data against the uncalibrated data to confirm the correction [60].

Behavioural Validation Protocols

Creating a Labelled Dataset for Machine Learning

The gold standard for validating accelerometer-based behaviour classification is to match the acceleration signal with directly observed behaviour.

Experimental Principle and Workflow

Video recordings of tagged animals are synchronized with the accelerometer data stream. An observer then annotates the video, labelling the behaviour at each point in time. This labelled dataset is used to train and test a machine learning model, which can then predict behaviours from acceleration data alone [62].

The workflow for this process is detailed below:

Required Materials

Table 2: Research Reagent Solutions for Behavioural Validation

Item	Specification	Function
Tri-axial Accelerometer	High-frequency capable (e.g., >40 Hz)	Logs the movement data of the subject.
Video Recording System	High-speed camera (e.g., 90 fps)	Captures ground-truth behaviour for labelling [63].
Synchronization System	Custom electronics or timestamps	Ensures precise alignment of video and accelerometer data [63].
Annotation Software	e.g., Framework4 [62]	Allows for precise labelling of behaviours in the video footage.
Data Processing Software	e.g., R or Python with ML libraries	For feature extraction and model training [62].

Step-by-Step Protocol

• Data Collection: Equip the subject with an accelerometer. For captive or habituated wild animals, simultaneously record high-quality video from multiple angles if necessary. For the accelerometer, select an appropriate sampling frequency. For short-burst behaviours (e.g., swallowing in birds), a high frequency of 100 Hz or more may be required. For sustained, rhythmic behaviours (e.g., flight), a lower frequency of 12.5 Hz may be sufficient [63].
• Synchronization: Synchronize the video and accelerometer data streams to millisecond accuracy. This can be achieved by using a shared timestamp or by creating a distinct synchronous event recorded by both systems (e.g., a sharp tap on the sensor) [62].
• Behavioural Annotation: Using annotation software, label the observed behaviours in the video at a defined frequency (e.g., 1 Hz). It is crucial to define behaviours clearly and consistently. Discard rare behaviours with insufficient observational data (e.g., less than 100 seconds in total) to avoid training on poor examples [62].
• Feature Extraction: From the raw, calibrated accelerometer data, calculate a suite of descriptive variables for the same time windows used in annotation. Key variables include [62] [59]:
  • Static acceleration (posture)
  • Dynamic Body Acceleration (DBA) and Vectorial DBA (VeDBA) (movement intensity)
  • Pitch and Roll
  • Spectral properties (e.g., dominant frequency and amplitude)
• Model Training and Testing: Use a supervised machine learning algorithm, such as a Random Forest, to train a model on a portion (e.g., 70-80%) of the labelled data. Validate the model's predictive accuracy against the remaining, held-out test data. Report performance metrics like precision, recall, and F-measure for each behaviour [59].

Key Considerations for Optimizing Validation

• Standardize Behaviour Durations: Balance the training dataset so that each behaviour class is represented for a similar total duration. This prevents the model from being biased towards predicting the most common behaviours and improves the identification of rarer ones [59].
• Field Validation: For wild studies, where possible, validate the model's predictions on a subset of field data through direct observation. This confirms that the model generalizes beyond the initial training conditions [59].

Technical Specifications and Data Collection

Adhering to technical best practices during data acquisition is a foundational form of validation.

Table 3: Summary of Technical Specifications for Accelerometer Data Collection

Parameter	Consideration	Impact on Data & Analysis
Sampling Frequency	Must obey Nyquist-Shannon theorem [63].	Too Low: Aliasing, loss of high-frequency behavioural information (e.g., swallows at 28 Hz) [63].
	Short-burst behaviours: ≥ 2x Nyquist frequency (e.g., 100 Hz) [63].	Too High: Rapid battery drain, large data files, computationally intensive processing [63].
	Long-endurance behaviours: Can use lower frequencies (e.g., 12.5 Hz) [63].
Device Placement	Standardize position and orientation on the body [11].	Different placements (back vs. tail) yield different signal amplitudes, confounding energy expenditure estimates and behaviour classification [11].
Attachment Method	Use a secure, consistent attachment (e.g., leg-loop harness, collar) [63] [62].	Loose attachments create sensor noise and motion artefacts, reducing classification accuracy.

Rigorous calibration and comprehensive validation are not optional steps but are fundamental prerequisites for generating scientifically sound data from animal-borne accelerometers. The protocols outlined herein—from the 6-O static calibration to the creation of video-verified labelled datasets—provide a framework to minimize measurement error and maximize behavioural classification accuracy. By integrating these practices, researchers can ensure that their inferences about animal movement, behaviour, and energetics are built upon a reliable and valid foundation.

Impact of Sensor Placement and Attachment on Data Quality

In the field of animal movement ecology, biologging devices that combine GPS tracking and accelerometry have become indispensable tools for remotely observing behavior, migration patterns, and energy expenditure [64] [65] [11]. The data quality from these sensors is paramount, as it directly influences the validity of ecological inferences. However, this quality is not solely a function of sensor specifications; it is critically dependent on sensor placement on the animal's body and the method of attachment [66] [11]. Variations in these factors can alter the recorded signal, introducing noise or bias that can be misinterpreted as biological phenomenon [11]. This application note, framed within a broader thesis on animal movement analysis, outlines the impacts of placement and attachment on data quality and provides detailed protocols to mitigate these issues for researchers and scientists.

The Impact of Sensor Placement

The location of a sensor on an animal's body determines which movements are captured and amplified. A placement optimal for measuring one type of behavior might be unsuitable for another, and this varies significantly across taxa.

Key Considerations for Placement

• Body Segment Dynamics: The acceleration of one body segment may not represent the movement of the whole body. For instance, during cycling in humans, or during the wingbeat cycle in birds, the trunk experiences pitch changes that affect acceleration measurements depending on the sensor's precise position [67] [11].
• Research Question Alignment: The most appropriate sensor position depends heavily on the study's aim. Research on head-specific behaviors like foraging should prioritize neck mounting, whereas studies focused on overall locomotion energy expenditure might be better served with a placement near the center of mass, such as the back [66] [68].

Empirical Evidence from Animal Studies

Studies across species have quantified the effect of sensor placement on acceleration data.

Table 1: Impact of Sensor Placement on Acceleration Metrics in Different Species

Species	Compared Placements	Key Finding	Effect Size
Canada Goose [66]	Neckband vs. Backpack	Behaviors performed by the head (e.g., foraging, vigilance) were better detected by neckbands. Behaviors like resting and walking were more successfully identified by backpacks.	Classification success varied by behavior.
Pigeon [11]	Upper back vs. Lower back	Variation in Dynamic Body Acceleration (DBA), a proxy for energy expenditure.	~9% difference in VeDBA
Black-legged Kittiwake [11]	Back vs. Tail	Variation in Dynamic Body Acceleration (DBA).	~13% difference in VeDBA
Domestic Dog [68]	Neck, Sternum, Pelvis, Knee	The pelvis and knee showed the highest acceleration peaks. The sternum and pelvis offered the most consistent signals for gait analysis in larger dogs.	Significant differences in acceleration peaks between body regions.

The findings from these studies underscore that there is no universally "best" location. The choice is a trade-off that must be deliberately made based on the target behaviors [66].

The Impact of Sensor Attachment

The method used to attach the sensor to the animal influences both the animal's welfare and the fidelity of the data collected. An improper attachment can lead to device loss, injury, or compromised data.

Attachment Methods and Their Consequences

• Collars: Commonly used for mammals with a neck larger than the head (e.g., primates, large cats) [65]. They must be fitted correctly to avoid injuries, as evidenced by a study on mantled howler monkeys where collars caused deep neck lacerations [65].
• Harnesses: Used for animals where neck diameter exceeds that of the head (e.g., pigs, Tasmanian devils) or for large birds [65]. However, some goose species are known to quickly destroy backpack harnesses, affecting tag retention [66].
• Direct Attachment: Used for birds, reptiles, and marine mammals. Devices are glued or taped to the skin, feathers, or carapace, and are designed to fall off during molting or are retrieved after a set time using a release timer [65].
• Implantation: Used for species where external attachment is not feasible (e.g., rhinoceros horn implants, snake implants). This method may suffer from a reduced transmission range as the animal's body tissue can absorb signal power [65].

Tightness of Fit and Calibration

A snug sensor fit is crucial for data quality. A loose fit can result in sensor rotation and reduced output amplitude, as the movement of clothes or fur introduces interference [11] [69]. Furthermore, the fabrication process of loggers, which involves soldering, can alter the accelerometer's output, making pre-deployment calibration essential [11]. One study found that uncalibrated tags resulted in DBA differences of up to 5% in humans walking at various speeds [11].

Detailed Experimental Protocols

To ensure data quality and cross-study comparability, standardized protocols for calibration and placement are indispensable.

Protocol 1: Accelerometer Calibration (6-O Method)

This protocol, adapted for field conditions, corrects for sensor inaccuracies and should be performed prior to every deployment [11].

Objective: To derive correction factors for the gain and offset of each axis of a tri-axial accelerometer. Materials: Data logger, flat, stable surface, data recording system.

• Preparation: Power on the data logger and start recording.
• Orientation Sequence: Place the motionless logger sequentially in six predefined orientations, ensuring one axis is perpendicular to gravity each time. Hold each position for ~10 seconds.
  • Front (X-axis up)
  • Back (X-axis down)
  • Left (Y-axis up)
  • Right (Y-axis down)
  • Up (Z-axis up)
  • Down (Z-axis down)
• Data Analysis: For each orientation period, calculate the vector sum of the raw acceleration: ||a|| = √(x² + y² + z²). The six maxima of this sum correspond to the static readings for each axis direction.
• Correction Factor Calculation: For each axis, calculate a correction factor to ensure the two maxima (positive and negative) are equal and normalized to 1.0 g. Apply these factors as a gain to all subsequent data from that device.

Protocol 2: Evaluating Placement and Attachment in Captive Trials

This protocol uses a controlled setting to assess the impact of tag design on animal behavior and data quality before field deployment [66].

Objective: To quantify the effects of tag attachment type on animal behavior, GPS accuracy, and accelerometer-based behavior classification. Materials: Multiple tag types (e.g., neckband, backpack), captive animal group, video recording system, high-precision GPS reference (e.g., DGPS).

• Experimental Design: Select a group of captive animals (e.g., geese). Use a rotational design where individuals are assigned to different tag treatments (e.g., neckband, backpack, control) over successive trials.
• Data Collection:
  • Behavioral Observation: Record animals live (e.g., using Observer XT software) for predefined periods, coding for key behaviors (e.g., feeding, resting, vigilance, preening).
  • GPS Accuracy Assessment: Simultaneously, log the position of the experimental animal using a high-precision Total Station or DGPS to establish a ground truth. Compare this with the GPS coordinates from the animal-borne tag.
  • Accelerometer Data Collection: Program tags to record high-frequency (e.g., 50 Hz) tri-axial acceleration data.
• Data Analysis:
  • Compare the duration and frequency of behaviors between tagged and control individuals.
  • Calculate the deviation of the tag GPS positions from the ground truth reference.
  • Train machine learning classifiers to identify behaviors from the accelerometer data and compare the success rates between different tag attachments.

Diagram Title: Experimental workflow for sensor calibration and evaluation.

The Scientist's Toolkit: Research Reagent Solutions

Table 2: Essential Materials for Animal-Borne Sensor Studies

Item / Solution	Function / Explanation
GPS/GSM Data Loggers	Core tracking unit. Acquires location data and transmits it via cellular (GSM) or satellite networks (e.g., Argos) [64] [65] [70].
Tri-axial Accelerometer	Sensor measuring acceleration in three perpendicular axes (X, Y, Z). The foundation for behavior and energy expenditure analysis [67] [71].
Programmable Release Mechanism	Enables non-recapture retrieval. Can be timer-based or remotely triggered via radio signal [65].
Helical Antenna	An omnidirectional antenna design that can improve signal reception reliability compared to patch antennas, especially when tag orientation is variable [66].
Biocompatible Adhesives	For direct attachment to animals (e.g., birds, marine mammals). Must be strong enough for deployment duration but allow release during molting [65].
Customized Harnesses/Collars	Species-specific attachment systems using materials like silicone rubber or nylon webbing. Design must minimize abrasion and maximize animal welfare [65] [66].
Differential GPS (DGPS)	Provides high-precision ground truth location data for validating the accuracy of animal-borne GPS units [66].
Animal-borne Video Camera	Provides contextual validation, allowing researchers to match specific accelerometer signatures directly to observed behaviors [71].
4H-1,4-Oxazine	4H-1,4-Oxazine|CAS 290-47-1|Research Chemical
17(R)-Protectin D1	17(R)-Protectin D1

The integrity of data in animal movement research is profoundly affected by seemingly mundane methodological choices regarding sensor placement and attachment. As evidenced, these choices can introduce variation in key metrics like DBA that is comparable to the magnitude of biological effects being studied. Therefore, it is not sufficient to select a device based solely on its specifications. Researchers must:

• Calibrate devices before deployment.
• Empirically validate the chosen placement and attachment method for their specific study species and question, ideally through controlled captive trials.
• Document and report all methodological details, including calibration procedures, attachment type, and precise sensor position, to ensure reproducibility and enable meaningful meta-analyses.

Adhering to these rigorous protocols will minimize confounding technical artifacts and ensure that the observed signals truly reflect the fascinating biology of the study animals.

Diagram Title: Decision logic for sensor placement and attachment.

The analysis of animal movement via GPS tracking and accelerometer data is a cornerstone of modern movement ecology. However, the integrity and continuity of the collected data are frequently compromised by three interconnected challenges: limited battery life, GPS signal dropout, and suboptimal sampling interval selection. These factors can introduce significant gaps and inaccuracies in movement paths, ultimately biasing ecological inference [53] [72]. This document provides detailed application notes and experimental protocols to manage these issues, ensuring the collection of high-quality data for a broader thesis on animal movement analysis. The guidance is structured to assist researchers in making informed decisions from experimental design to data validation.

The following tables synthesize key quantitative relationships from empirical studies to inform device configuration.

Table 1: GPS Fix Interval Trade-Offs

Fix Interval	Horizontal Accuracy (Mean Error)	Vertical Accuracy (Mean Error)	Effect on Track Length Estimation	Battery & Longevity Impact
1 second	3.4 m [72]	4.9 m [72]	Most accurate for fine-scale behavior [73]	Highest battery drain; shorter study duration
1 minute	5.1 m [72]	7.2 m [72]	Suitable for high-resolution foraging tracks [73]	High battery drain
60 minutes	6.5 m [72]	9.7 m [72]	Can underestimate actual track length by up to 50% [73]	Lowest battery drain; enables long-term studies

Table 2: Energy and Accuracy Configuration Guide

Setting	Impact on Battery Life	Impact on Data Accuracy/Completeness	Recommended Use Case
High GPS Accuracy	Very High	Highest positional accuracy (within meters) [74]	Critical navigation or fine-scale habitat use [74]
Balanced/Power Saving Mode	Moderate	Good accuracy (100-500m for power saving) [74]	Most general tracking applications [74]
Frequent Location Updates (e.g., 10 sec)	Very High	High path resolution [74]	Capturing detailed movement kinematics [74]
Infrequent Updates (e.g., 5-10 min)	Low	Coarse path resolution; may miss behaviors [74] [21]	Long-term home range or migration studies [74]
GSM/GPRS Data Transmission	High	Enables remote data access without retrieval [72]	Animals that are difficult to recapture [72]
Archival (Data Logging)	Low	Data recovery requires device retrieval [75] [73]	Short-term studies or species with high site fidelity [75]

Experimental Protocols

Protocol: Pre-Deployment GPS Device Validation

This protocol is designed to quantify the baseline accuracy and precision of GPS devices before deployment on animals [72].

1. Objectives:

• To determine the horizontal and vertical accuracy of GPS loggers under controlled, stationary conditions.
• To establish the relationship between fix interval and positional error.
• To assess the effectiveness of built-in error metrics (e.g., GPS-Error, DOP) for identifying inaccurate fixes.

2. Materials:

• GPS tracking devices ready for deployment (n ≥ 10 recommended for statistical power) [72].
• A geodetic survey marker or a known location with high-precision coordinates (e.g., via differential GPS).
• A secure, open-sky test site free from multipath interference (e.g., large open field).
• Data sheets or software for recording reference coordinates and device outputs.

3. Procedure:

• Step 1: Site and Device Setup. Place all devices centrally on the known reference point. Ensure they are secured in their deployment housings to replicate field conditions [72].
• Step 2: Data Collection. Program devices to simultaneously collect fixes at multiple intervals (e.g., 1s, 10s, 1min, 10min). Log data for a minimum of 24 hours to capture various satellite geometries.
• Step 3: Data Analysis.
  • Accuracy: For each fix, calculate the horizontal error (2D distance from recorded position to the known reference) and vertical error (altitude difference).
  • Precision: For each device and fix interval, calculate the standard deviation of all recorded positions around their mean position.
  • Error Metric Validation: Compare the recorded GPS-Error or HDOP values against the measured accuracy to determine a threshold for filtering low-quality fixes post-deployment [72].

Protocol: Integrated GPS-Accelerometer Behaviour Classification

This protocol outlines a method for using accelerometer data to classify behaviour, which can be used to infer activities during GPS signal dropouts [21] [13].

1. Objectives:

• To train a machine learning model to classify specific animal behaviours from accelerometer signals.
• To create a validated ethogram that can be applied to periods with missing GPS data.

2. Materials:

• Tri-axial accelerometers (sampling at ≥10 Hz) collocated with GPS devices [21] [13].
• Video recording equipment for ground-truthing.
• Computing environment with machine learning libraries (e.g., R, Python).

3. Procedure:

• Step 1: Data Collection. Deploy devices on study animals. Simultaneously record high-quality video of the animals' activities for a subset of the deployment period to create a labeled dataset [21].
• Step 2: Data Preprocessing & Feature Engineering. Synchronize accelerometer data with video observations. For each annotated behaviour (e.g., grazing, ruminating, walking), extract ~100 features from the raw accelerometer signals (X, Y, Z axes) in both time and frequency domains (e.g., mean, variance, dominant frequency, signal entropy) [21].
• Step 3: Model Training. Use a Random Forest classifier, trained on the labeled features, to predict behaviour from accelerometer data alone. Validate model performance using k-fold cross-validation [21].
• Step 4: Application. Apply the trained model to the entire accelerometer dataset. The classified behaviours can now be used to contextualize movement paths and infer likely locations during GPS dropouts based on known behaviour-habitat relationships.

Workflow Visualization

The following diagram illustrates the integrated decision-making process for managing GPS data loss, from device configuration to data analysis and gap mitigation.

GPS Data Management Workflow

The Scientist's Toolkit: Research Reagent Solutions

Table 3: Essential Materials and Technologies for GPS Tracking Studies

Item	Specification / Example	Primary Function in Research
GPS/GPRS Tracking Device	e.g., Movetech Telemetry Flyways-50; solar-powered, archival & remote data transmission [72].	Core unit for obtaining animal location (fix) data. Key parameters are weight, fix interval, and power source.
Tri-axial Accelerometer	MEMS-based sensor, sampling ≥10 Hz, dynamic range ±2g, integrated into device collar [21] [13].	Records fine-scale movement and body posture for behavioral classification and energy expenditure estimation.
Data Logging / Transmission Module	Archival (SD card) or GSM/GPRS for remote transfer [72].	Manages data storage and recovery. Remote transmission is vital for animals that are difficult to recapture.
Machine Learning Classifier	Random Forest algorithm (e.g., in R or Python) [21].	Classifies raw accelerometer data into distinct behavioral states (e.g., grazing, ruminating).
Path Segmentation Software/Methods	Hidden Markov Models (HMMs) or Change-point Analysis [53].	Partitions movement paths into discrete segments representing potential behavioral states.
Stationary Test Kit	Geodetic survey marker, open-sky test site, external power banks.	Provides a known location for pre-deployment validation of GPS device accuracy and precision [72].

Addressing Spatial and Temporal Autocorrelation in Tracking Data

In animal movement ecology, autocorrelation refers to the statistical dependence between consecutive location estimates in a tracking dataset [76]. Spatial autocorrelation exists when nearby locations are more similar than distant ones, while temporal autocorrelation occurs when measurements taken close in time are more similar than those taken far apart [77]. This autocorrelation presents a fundamental challenge for statistical analysis because most conventional statistical methods assume independence of data points. When tracking technologies record animal positions at fine temporal scales (from hourly to second-by-second), successive locations are inherently non-independent, creating a signature of autocorrelation that must be explicitly addressed to avoid biased ecological inferences [47] [76] [78].

The proliferation of high-resolution tracking technologies, including GPS telemetry and accelerometers, has dramatically increased the quantity and quality of animal movement data [47] [13]. While this detailed data provides unprecedented insight into animal behavior, it also intensifies the challenges associated with autocorrelation. Properly addressing these dependencies is crucial for accurate home range estimation, resource selection analysis, and behavioral classification [76] [78]. This protocol outlines comprehensive methods for identifying, quantifying, and accounting for spatial and temporal autocorrelation in animal tracking data, with specific application notes for researchers working with GPS and accelerometer data within a broader movement ecology framework.

Theoretical Foundation

The Nature of Autocorrelation in Tracking Data

Autocorrelation in animal tracking data arises from the fundamental nature of animal movement itself. Animals do not move randomly through their environment but instead exhibit directional persistence and behavioral states that create predictable patterns in their movement trajectories [78]. The strength and scale of this autocorrelation are influenced by both internal factors (e.g., hunger, reproductive state, species-specific movement capacities) and external factors (e.g., resource distribution, predation risk, landscape heterogeneity) [78] [13].

Temporal autocorrelation manifests at multiple scales, including diurnal cycles (activity-rest patterns), seasonal patterns (migration, seasonal resource use), and behavioral sequences (foraging bouts, territorial patrols) [78]. Spatial autocorrelation emerges from the fact that an animal's location at time t+1 is physically constrained by its location at time t, with the strength of this constraint inversely related to the time between observations [76]. Understanding this multi-scale nature of autocorrelation is essential for selecting appropriate analytical techniques.

Implications for Ecological Inference

Failure to properly account for autocorrelation can lead to several analytical problems, including pseudoreplication, inflated sample sizes in statistical tests, biased parameter estimates, and overly narrow confidence intervals [76] [78]. In home range analysis, autocorrelation can cause underestimation of home range size when sampling intervals are too short to capture the full extent of movement [76] [79]. In resource selection studies, autocorrelation can create spurious associations with environmental variables that merely correlate with an animal's movement path rather than truly representing selection [76].

Importantly, autocorrelation is not merely a statistical nuisance—it also contains valuable biological information about movement processes [78]. The autocorrelation structure of movement paths can reveal behavioral modes, energy expenditure patterns, and responses to environmental stimuli [78] [13]. Thus, the goal of addressing autocorrelation is not simply to remove it, but to properly model it to extract meaningful biological insight while maintaining statistical validity.

Quantification Methods

Temporal Autocorrelation Analysis

Table 1: Methods for Quantifying Temporal Autocorrelation in Movement Data

Method	Application	Key Outputs	Biological Interpretation
Variogram Analysis [79]	Assessing range residency and effective sample size	Semivariance vs. time lag plot; Asymptotic variance; Range crossing time	Indicates whether animal is range resident; Informs sampling interval selection
Fourier Analysis [78]	Identifying periodic patterns in movement	Periodogram showing variance explained by different frequencies	Reveals diurnal/seasonal cycles; Identifies dominant behavioral rhythms
Wavelet Analysis [78]	Detecting non-stationary periodic patterns	Scalogram showing frequency power across time	Locates temporal shifts in behavioral patterns; Identifies transient cyclic behaviors
Autoregressive (AR) Modeling [78]	Characterizing dependency structure	AR coefficients for different time lags; Optimal model order (p)	Quantifies persistence in movement parameters; Models memory in movement process

Variogram Protocol

Variogram analysis assesses the dependence between observations as a function of the time separation between them [79]. The protocol involves:

• Data Preparation: Calculate step lengths (distances between consecutive locations) and time intervals for the entire tracking series.

• Semivariance Calculation: For each possible time lag (Ï„), compute the semivariance using the formula: γ(Ï„) = ½Var[Z(t+Ï„) - Z(t)], where Z(t) represents the animal's location at time t.

• Variogram Plotting: Create a plot of semivariance versus time lag. For range-resident animals, this plot will show an increasing semivariance that eventually reaches an asymptote (the sill variance) at the time lag where locations become independent (the range) [79].

• Parameter Estimation: Fit a theoretical model (e.g., spherical, exponential) to the empirical variogram to estimate the range (time to independence) and sill (asymptotic variance).

• Effective Sample Size Calculation: Compute the effective sample size as n' = n / (1 + 2Σ(1 - k/N)ρₖ), where n is the total number of locations, N is the maximum lag, and ρₖ is the autocorrelation at lag k.

Fourier Analysis Protocol

Fourier analysis decomposes the movement time series into constituent frequencies to identify periodic patterns [78]:

• Data Transformation: Convert the sequence of step lengths into a standardized time series, optionally applying a transformation (e.g., log) to stabilize variance.

• Periodogram Calculation: Compute the Fourier periodogram using the fast Fourier transform (FFT) algorithm: I(ω) = (1/n)|Σxâ‚œe^(-iωt)|², where xâ‚œ represents the step length at time t, and ω represents angular frequency.

• Significance Testing: Compare the observed periodogram against a theoretical red-noise spectrum using a chi-square test or bootstrap methods to identify statistically significant frequencies.

• Biological Interpretation: Identify the biological correlates of significant frequencies (e.g., 24-hour period = diurnal cycle; 12-hour period = bimodal activity pattern).

Spatial Autocorrelation Analysis

Table 2: Methods for Quantifying Spatial Autocorrelation in Movement Data

Method	Application	Key Outputs	Data Requirements
Mantel Test [76]	Testing correlation between distance matrices	Mantel statistic (r); Significance (p-value)	Paired spatial and temporal distance matrices
Spatial Autocorrelogram [76]	Measuring autocorrelation at different distance classes	Moran's I or Geary's C for distance classes	Regular sampling grid or interpolated data
Autocorrelated Kernel Density Estimation (AKDE) [79]	Home range estimation with autocorrelation correction	Utilization distribution; Home range contours	Continuous-time movement model
Behavioral Change Point Analysis [47]	Identifying shifts in movement behavior	Change point locations; Behavioral segmentation	High-resolution movement data (e.g., ≤1 second)

AKDE Home Range Estimation Protocol

Autocorrelated Kernel Density Estimation (AKDE) explicitly models the autocorrelation structure in tracking data to produce unbiased home range estimates [79]:

• Movement Model Selection: Fit a continuous-time movement model (e.g., integrated Ornstein-Uhlenbeck process) to the tracking data using maximum likelihood estimation.

• Autocorrelation Modeling: Estimate the autocorrelation structure from the model residuals using the variogram approach described in section 3.1.1.

• Bandwidth Selection: Calculate the autocorrelation-adjusted bandwidth parameter using the formula: H = (1/n)ΣC(táµ¢ - tâ±¼), where C is the covariance function estimated from the movement model.

• Utilization Distribution Estimation: Apply the Gaussian kernel density estimator with the adjusted bandwidth: f(x) = (1/n)ΣKâ‚•(x - Xáµ¢), where Kâ‚• is the Gaussian kernel with bandwidth h.

• Home Range Contour Delineation: Calculate the 95% and 50% utilization distributions to represent the home range and core area, respectively.

Experimental Protocols

Integrated Workflow for Addressing Autocorrelation

The following diagram illustrates the comprehensive workflow for addressing autocorrelation in animal tracking data, incorporating both spatial and temporal aspects:

Behavioral Segmentation Using Information Theory

For high-resolution tracking data (≥1 Hz), fine-scale behavioral segmentation can be achieved using information theory concepts [47]. The following protocol enables identification of canonical activity modes (CAMs) from raw movement tracks:

• Data Preparation: Resample tracking data to consistent time intervals (e.g., 1 second). Calculate step lengths and turning angles between consecutive locations.

• StaMEs Definition: Define the smallest viable statistical movement elements (StaMEs) as sequences of μ steps. Cluster these StaMEs into distinct types using k-means or hierarchical clustering based on step length and turning angle distributions [47].

• Word Formation: Construct "words" by concatenating sequences of m StaMEs. These words represent behavioral sequences at a higher organizational level.

• CAMs Identification: Apply cluster analysis to the words to identify centroids representing canonical activity modes (e.g., foraging, resting, directed movement) [47].

• Entropy Calculation: Compute the Shannon entropy of the movement path using the formula: H = -Σpáµ¢logâ‚‚páµ¢, where páµ¢ represents the probability of each CAM.

• Validation: Validate CAM assignments against independent behavioral observations from accelerometer data or direct observation where available [13].

Accelerometer Integration Protocol

Integrating accelerometer data with GPS tracking provides a powerful approach for behavioral classification and energy expenditure estimation while addressing autocorrelation [13]:

• Sensor Configuration: Deploy tri-axial accelerometers synchronized with GPS units, sampling at ≥10 Hz for accelerometry and ≥1 Hz for GPS [13].

• Data Synchronization: Align accelerometer and GPS data streams using internal clocks and common time stamps.

• Static Acceleration Separation: Use a high-pass filter (e.g., Butterworth filter) to separate dynamic acceleration (movement) from static acceleration (posture) [13].

• Behavioral Classification: Apply machine learning classifiers (e.g., random forest, hidden Markov models) to accelerometer waveforms to identify specific behaviors (e.g., feeding, walking, resting) [13].

• Movement Validation: Use classified behaviors to validate and interpret movement modes identified from GPS tracking alone.

• Energy Expenditure Estimation: Correlate the overall dynamic body acceleration (ODBA) with energy expenditure measures from doubly labeled water studies for the target species [13].

The Scientist's Toolkit

Table 3: Essential Research Reagent Solutions for Movement Ecology Studies

Tool/Category	Specific Examples	Function/Purpose	Application Notes
Tracking Technologies	GPS collars with accelerometers [13]; Cellular-enhanced GPS [79]	Position and movement data acquisition	Cellular-enhanced GPS improves urban tracking [79]; Accelerometers sample at 10+ Hz for detailed behavior [13]
Data Management Platforms	Movebank; Wildlife Desktop	Centralized data storage and management	Facilitates data sharing and collaboration; Provides basic visualization tools
Statistical Software	R with ctmm package [79]; adehabitat; momentuHMM	Autocorrelation-informed analysis	ctmm implements AKDE for home range estimation [79]; momentuHMM handles behavioral state modeling
Programming Languages	R; Python; MATLAB	Custom analysis and visualization	R has most comprehensive movement ecology packages; Python offers machine learning capabilities
Field Equipment	GPS base stations; Data download kits; Battery testers	Field deployment and maintenance	Regular battery testing essential for long-term deployments; Base stations improve GPS accuracy
Validation Tools	Camera traps; Direct observation logs; Physiological sensors	Independent behavioral validation	Critical for ground-truthing behavioral classifications from movement data

Application Notes

Case Study: Urban Carnivore Tracking

In a proof-of-concept study tracking urban raccoons using cellular phone-enhanced GPS technology, researchers achieved a median positional accuracy of 140 meters, with over 30% of fixes achieving <50 meters error [79]. Key findings and recommendations include:

• Sampling Regime: A 1-hour sampling interval yielded sufficient data for home range analysis using AKDE, with effective sample sizes averaging 34.1 (range: 2.2-133.9) [79].

• Battery Considerations: Continuous operation with hourly sampling resulted in an average battery life of 23.0 days, substantially shorter than the projected 90 days, highlighting the importance of field testing power management strategies [79].

• Home Range Estimation: AKDE produced home range estimates that were on average 2.7 times larger than conventional KDE estimates, demonstrating the substantial bias that can result from failing to account for autocorrelation [79].

Case Study: Elephant Movement Periodicity

Research on African elephant movement using Fourier and wavelet analysis revealed strong diurnal cycles in step lengths, with autocorrelation strength varying seasonally and socially [78]:

• Seasonal Patterns: Autocorrelation was significantly stronger during the dry season (rs = -0.789, P < 0.001), when resource distribution compelled more predictable movement patterns between water and forage [78].

• Social Influences: Socially dominant individuals maintained more consistent autocorrelation patterns across seasons, while subordinate individuals showed distinct dry-season divergence, likely reflecting competitive exclusion [78].

• Risk Effects: Diurnal movement correlation was more common within protected areas, while multiday movement correlations among lower-ranked individuals typically occurred outside protected areas where predation risks were greater [78].

Implementation Considerations

When designing tracking studies to address autocorrelation, several practical considerations emerge:

• Sampling Frequency: The optimal sampling rate depends on the research question and species movement characteristics. For behavioral classification, high-frequency sampling (≤1 second) may be necessary [47], while for home range analysis, longer intervals (1-12 hours) may suffice [79].

• Study Duration: Tracking duration should exceed the range crossing time (time for an animal to cross its home range) to ensure representative sampling of space use. Variogram analysis can determine whether this threshold has been met [79].

• Data Volume Management: High-frequency tracking generates large datasets requiring efficient computational strategies. Cloud computing platforms and specialized movement databases can facilitate storage and analysis.

Addressing spatial and temporal autocorrelation is not merely a statistical necessity in animal movement analysis—it is an opportunity to extract deeper biological insight from tracking data. The protocols outlined here provide a comprehensive framework for quantifying, interpreting, and modeling autocorrelation structures across different temporal and spatial scales. By explicitly incorporating autocorrelation into analytical models, researchers can produce more accurate home range estimates, more realistic resource selection functions, and more meaningful behavioral classifications. As tracking technologies continue to evolve, providing ever-higher resolution data, these autocorrelation-informed approaches will become increasingly essential for advancing our understanding of animal movement ecology.

Optimizing Data Processing for Large, Complex Bio-logging Datasets

In the field of movement ecology, researchers are increasingly confronted with the challenges posed by large, complex bio-logging datasets. Modern tracking technologies, such as GPS sensors and tri-axial accelerometers, now generate high-resolution data at sub-second intervals, smashing the decades-old limits of observational studies [47] [13]. Where early tracking might provide an animal's location hourly, current technologies can record position and dynamic acceleration many times per second, creating dense data streams that require sophisticated processing approaches [47] [10]. This data explosion presents both unprecedented opportunities and significant computational challenges for researchers studying animal behavior across multiple scales—from individual foraging decisions to population-level space-use patterns [8].

The complexity of bio-logging data necessitates robust processing frameworks that can integrate multiple data types while accounting for the unique characteristics of animal-borne sensor data. As noted in research on cattle behavior monitoring, "accelerometer signals were sampled at 10 Hz, and data from each axis was independently processed to extract 108 features in the time and frequency domains" [10]. Similarly, GPS data requires specialized handling to balance battery consumption with spatial accuracy, often employing sampling intervals of 5 minutes or more to extend deployment duration while maintaining ecological relevance [10]. This protocol outlines comprehensive strategies for optimizing the processing of such datasets, with particular emphasis on feature extraction, behavioral classification, and spatial analysis.

Experimental Protocols and Workflows

Data Acquisition Hardware Specifications

Electronic monitoring devices for bio-logging typically integrate multiple sensors within a single, weatherproof unit attached to animals via collars, harnesses, or other mounting systems. The specifications and configuration of these devices fundamentally shape subsequent data processing requirements and possibilities.

Accelerometer Configuration: Research-grade accelerometers are typically Micro Electro Mechanical System (MEMS) based triaxial sensors that measure acceleration in three orthogonal directions (surge, heave, and sway) [10] [13]. These sensors capture both DC acceleration (earth's gravity, providing orientation data) and dynamic inertial acceleration due to movement. For behavioral studies, a sampling frequency of 10 Hz is commonly employed, sufficient to capture most gross motor behaviors while managing data volume [10]. The dynamic range is typically set at ±2g for large animal studies, though this may be adjusted for species with more explosive movement patterns [10].

GPS Configuration: To optimize battery consumption in commercial monitoring devices, GPS sampling intervals are typically set wider than accelerometer sampling—often at 5-minute intervals [10]. Configuration should aim for a maximum Dilution of Precision (DOP) threshold of 1, with signal reception from a minimum of 7 different satellites to ensure spatial accuracy. With proper configuration, the estimated average measurement error can be as low as 1.7 meters, with 90% of measurements presenting errors below 5.2 meters [10].

Device Considerations: Commercial bio-logging devices are designed for extended deployment (typically 2-3 months) and must balance data resolution with battery life and storage capacity [10]. Modern units can weigh as little as 0.7g without battery, making them suitable for a wide range of species [13]. Data can be stored onboard in SD memory cards or transmitted via ultra-high frequency technology similar to cellular phones, enabling download from distances up to 500 meters without physical retrieval [13].

Data Preprocessing Pipeline

Raw bio-logging data requires substantial preprocessing before analysis to ensure data quality and extract meaningful features. The workflow below illustrates this comprehensive preprocessing pipeline.

Data Validation and Quality Control: The initial processing stage involves rigorous quality assessment of raw sensor data. For GPS data, this includes evaluating DOP values, satellite counts, and identifying periods of signal loss that may occur "in certain shadow regions on the farm, or that transmitted data do not successfully arrive at the server, due to propagation issues, network problems or other causes" [10]. Accelerometer data requires checks for signal integrity, sampling gaps, and sensor range violations. Research indicates that "accelerometer measurements are typically collected in three dimensions of movement at very high resolution (>10 Hz)" [13], making comprehensive quality control essential.

Sensor Fusion and Synchronization: Integrating data streams from multiple sensors requires careful temporal alignment. GPS positions sampled at 5-minute intervals must be synchronized with 10 Hz accelerometer data and any video validation recordings [10]. Timestamp alignment should account for internal clock drift across devices and potential latency in data recording initiation.

Coordinate System Alignment: Raw accelerometer data requires transformation to an animal-centric reference frame to ensure consistent interpretation across individuals and deployments. This process involves rotating the sensor-based coordinate system to align with the animal's anatomical planes (sagittal, frontal, and transverse), correcting for variations in device orientation and attachment [13].

Noise Filtering and Smoothing: Acceleration signals benefit from appropriate filtering to reduce high-frequency noise while preserving biologically meaningful signals. Digital low-pass filters with cutoff frequencies between 3-5 Hz are commonly employed for large animal studies, effectively capturing body movements while eliminating high-frequency vibration artifacts [10] [13].

Feature Extraction: The filtered acceleration data serves as the foundation for extracting features in both time and frequency domains. Research on cattle behavior classification demonstrates that "108 features in the time and frequency domains" can be derived from tri-axial accelerometer data [10]. These typically include measures of variability, periodicity, and intensity across all three movement dimensions.

Table 1: Essential Feature Categories for Accelerometer Data Analysis

Domain	Feature Category	Specific Metrics	Biological Relevance
Time Domain	Statistical Moments	Mean, variance, skewness, kurtosis for each axis	Movement intensity and distribution
	Body Posture	Static acceleration components	Animal orientation and posture
	Dynamic Motion	Overall Dynamic Body Acceleration (ODBA)	Energy expenditure estimation [8]
Frequency Domain	Spectral Features	Dominant frequencies, spectral power	Periodicity of repetitive behaviors
	Entropy Measures	Sample entropy, spectral entropy	Behavioral complexity and predictability
Movement Geometry	Step-wise Characteristics	Step length, turning angles [8]	Path tortuosity and direction changes
	Vectorial	Trajectory straightness, net squared displacement	Movement efficiency and directionality

Behavioral Classification Methodology

Supervised machine learning approaches have proven highly effective for classifying animal behavior from accelerometer data. The following workflow outlines the complete process from data preparation to model deployment.

Reference Data Collection: For supervised classification, researchers must collect high-quality reference behavioral observations synchronized with sensor data. In cattle behavior studies, "a total of 238 activity patterns, corresponding to four different classes (grazing, ruminating, laying and steady standing), with duration ranging from few seconds to several minutes, were recorded on video and matched to accelerometer raw data to train a random forest machine learning classifier" [10]. This video-sensor synchronization enables the creation of labeled datasets essential for training accurate classification models.

Model Selection and Training: The Random Forest algorithm has demonstrated particular effectiveness for behavioral classification, with studies reporting "best accuracy (0.93) for grazing" in cattle [10]. Consistent with findings from systematic methodology assessments like the DREAM Challenge, ensemble methods often produce more robust results, and "simple methods can often perform remarkably well, with linear models like elastic net regression providing a strong baseline" [80]. Feature selection should prioritize biologically interpretable metrics while reducing redundancy to prevent overfitting.

Validation and Performance Assessment: Model performance should be evaluated using appropriate metrics such as overall accuracy, per-class precision and recall, and confusion matrix analysis. Cross-validation approaches that maintain temporal dependency in the data are essential, as random splitting of time-series data can produce overly optimistic performance estimates. Independent validation on completely withheld datasets provides the most reliable estimate of real-world performance.

Analytical Frameworks for Movement Ecology

Movement Trajectory Analysis

GPS tracking data enables the computation of fundamental movement metrics that characterize how animals navigate their environments. These metrics provide insight into movement strategies, resource selection, and behavioral states.

Table 2: Key Movement Metrics for GPS Trajectory Analysis

Metric Category	Specific Metric	Calculation Method	Ecological Interpretation
Path Geometry	Step Length	Straight-line distance between consecutive locations	Movement scale and intensity
	Turning Angle	Angular change in direction between successive steps	Tortuosity and direction persistence
	Straightness Index	Ratio of net displacement to total path length	Movement efficiency [8]
Space Use	Net Squared Displacement	Squared distance from trajectory start point	Range expansion and migration
	First Passage Time	Time required to exit circle of radius r from point	Area-restricted search behavior [8]
Recursion	Residence Time	Time spent within a defined area	Resource importance or preference
	Revisitation Rate	Frequency of returns to a specific area	Site fidelity or cache recovery
	Return Time	Time interval between consecutive visits	Temporal patterns in resource use

The application of information theory to movement analysis represents a promising frontier for handling high-resolution data. Recent research proposes "a fine-scale approach that rests heavily on concepts from Shannon's Information Theory" to analyze second-by-second movement data [47]. This approach enables researchers to "provide entropy measures for movement paths, compute the coding efficiencies of derived StaMEs and CAMs, and to assess error rates in the allocation of strings of m StaMEs to canonical activity modes (CAMs)" [47].

Spatial Analysis and Clustering

GPS data enables not only individual movement analysis but also the characterization of group-level spatial patterns and habitat use. Unsupervised clustering algorithms like k-medoids have been successfully applied to GPS data to "track location and spatial scatter of herds" [10]. This approach helps identify core activity areas, seasonal ranges, and patterns of pasture utilization that might indicate unbalanced resource use.

The integration of spatial clustering with behavioral classification creates powerful insights into animal-environment interactions. For example, linking grazing behavior identified from accelerometry with spatial positions from GPS can reveal preferential foraging areas and landscape features that influence behavior. This combined approach facilitates the "detection of anomalous situations on farms," such as predator threats or disease transmission, through identifying behavioral and spatial patterns that deviate from established norms [10].

The Scientist's Toolkit: Research Reagent Solutions

Table 3: Essential Research Tools for Bio-logging Studies

Tool Category	Specific Tool/Technique	Primary Function	Application Notes
Hardware Solutions	MEMS Tri-axial Accelerometer	Measures 3D acceleration at high frequency (>10 Hz)	Low-power consumption, suitable for long-term deployment [10] [13]
	GPS Loggers	Records animal positions at configurable intervals	5-minute intervals balance battery life with ecological relevance [10]
	Animal-borne Video Systems	Provides ground truth for behavior validation	Enables labeled dataset creation for supervised learning [10]
Computational Frameworks	Random Forest Classifier	Supervised behavior classification from accelerometer features	Demonstrates high accuracy (>0.93) for distinct behaviors [10]
	K-medoids Clustering	Unsupervised spatial analysis of GPS locations	Identifies herd aggregation patterns and core areas [10]
	Information Theory Measures	Quantifies entropy and complexity in movement paths	Analyzes fine-scale (second-by-second) behavioral sequences [47]
Data Processing Libraries	Movement Metrics Toolkits	Computes step lengths, turning angles, net displacement	Standardizes trajectory analysis across studies [8]
	Signal Processing Libraries	Extracts time and frequency domain features from acceleration	Enables computation of 100+ behavioral features [10]
	Deep Learning Frameworks	Implements neural networks for complex pattern recognition	Suitable for multimodal data integration (e.g., CNN, RNN, GAN) [81]

The optimization of data processing for large, complex bio-logging datasets requires an integrated approach that combines appropriate hardware configurations, robust preprocessing methodologies, and sophisticated analytical techniques. By implementing the protocols outlined in this document, researchers can transform raw sensor data into biologically meaningful insights about animal behavior, spatial ecology, and energy expenditure. The rapid advancement of tracking technologies continues to push the boundaries of what can be observed in free-ranging animals, and corresponding developments in analytical approaches are essential for maximizing the scientific value of these remarkable data streams. As the field progresses, approaches based on information theory and advanced machine learning offer promising avenues for extracting deeper insights from the rich tapestry of animal movement data [47] [80].

Benchmarking Truth: Validating Models and Comparing Analytical Approaches

Ground-Truthing with Simultaneous Behavioral Observations

Within the broader context of GPS tracking and accelerometer data analysis in animal movement research, ground-truthing is a critical process that links raw sensor data to observable animal behaviors. Simultaneous behavioral observations, typically via video recording, provide the foundational dataset for training and validating automated classification models. This process transforms accelerometer and GPS signals into meaningful, behaviorally annotated data, enabling researchers to answer fundamental questions about animal energetics, ecology, and conservation. Without rigorous ground-truthing, the vast quantities of data generated by modern biologging devices remain difficult to interpret accurately. This protocol outlines detailed methodologies for establishing this essential link between sensor data and animal behavior.

The Critical Role of Ground-Truthing in Movement Ecology

Remote recognition of behavior using accelerometers requires ground-truth data based on human observation or knowledge [82]. Accelerometers are sensitive to movement and orientation but cannot deduce behavior independently; they must be calibrated against a trusted source of behavioral information [82] [13]. The primary goal is to create a labeled dataset where specific patterns in the sensor data are matched to defined behaviors from an ethogram. This dataset then serves as a training foundation for machine learning algorithms, allowing them to later identify these behaviors from sensor data alone in new, unlabeled datasets [82]. This process is particularly vital for cryptic behavioral events that are difficult to observe directly in the wild, such as transient foraging actions or responses to subtle environmental cues [13]. Furthermore, ground-truthing mitigates the inherent limitations of direct human observation, including observer bias, the physical limitations of researchers, and the potential for the observer's presence to alter natural animal behavior [13].

Experimental Protocols for Simultaneous Data Collection

Core Equipment and Synchronization Setup

The following equipment is required for the simultaneous collection of behavioral observations and sensor data.

Table 1: Essential Research Reagents and Equipment

Item Name	Function/Description	Key Specifications
Tri-axial Accelerometer	Measures surge, sway, and heave acceleration (change in velocity) of the animal's body [13].	Sample rate: Typically 10-100 Hz [83] [13]; Resolution: >10 Hz [13].
GPS Logger	Records animal position and movement trajectory.	Sampling interval: Can vary from seconds to minutes [84] [85].
Video Recording System	Captures continuous behavioral observations for ground-truthing.	Resolution: >=1080p recommended for clarity [86].
Synchronization Mechanism	Aligns video footage and sensor data streams in time.	Can be a shared start signal, a visible event captured by both systems, or specialized sync hardware.

The workflow for initial setup and synchronization is critical and can be visualized as follows:

Behavioral Annotation and Ethogram Creation

The first procedural step involves an expert creating a comprehensive ethogram—a formal catalog of the behaviors to be studied [82]. For a study on wild meerkats, an ethogram might include resting, foraging, vigilance, and running [82]. In cattle research, common classes are grazing, ruminating, laying, and steady standing [84]. The video recording is then meticulously annotated according to this ethogram, creating a continuous timeline of observed behaviors [82]. This annotation process links the recorded acceleration signal directly to the stream of observed behaviors that produced it, forming the core of the ground-truthed dataset.

Data Processing and Model Training Workflow

Once simultaneous data is collected, the subsequent processing and analysis follow a structured pipeline to build a robust behavioral classification model.

Signal Processing and Feature Engineering

The raw, high-resolution accelerometer data is processed by segmenting it into finite windows of a pre-set size (e.g., 2-5 seconds) [82]. From the data within each window, quantitative features are engineered to summarize the signal's characteristics. The quality of these features is paramount; good features will have similar values for the same behavior and different values for different behaviors [82]. Typically, 15-20 features are computed in both the time and frequency domains [82] [84]. A biomechanically informed approach might focus on engineering a smaller set of powerful features that specifically quantify posture (via static acceleration), movement intensity (via dynamic acceleration), and movement periodicity [82].

Table 2: Example Performance Metrics of Classification Models

Study Organism	Behavioral Classes	Classification Algorithm	Reported Accuracy	Key Validation Method
Cattle [84]	Grazing, Ruminating, Laying, Standing	Random Forest	0.93 (Best, for Grazing)	Validation split/Hold out
Four Albatross Species [83]	Flapping Flight, Soaring Flight, On-water	Hidden Markov Model (HMM)	0.92 (Overall)	Expert classification of sensor patterns
Wild Meerkats [82]	Resting, Vigilance, Foraging, Running	Hierarchical Tree-like Scheme	>0.95 (for behaviors constituting >95% of time budget)	Leave-One-Individual-Out (LOIO)

Machine Learning and Model Validation

The ground-truthed features are used to train machine learning algorithms, such as Random Forest or Hidden Markov Models (HMMs) [82] [84] [83]. The general workflow for this phase is comprehensive:

A critical best practice is to use appropriate cross-validation methods. Leave-One-Individual-Out (LOIO) cross-validation is highly recommended for characterizing a model's ability to generalize to new, unseen individuals [82]. In this method, the model is trained on data from all individuals but one and tested on the left-out individual. This process is repeated until every individual has been used as the test set. LOIO helps mitigate the effects of non-independence in data extracted from the same individual's time series [82]. When reporting results, it is essential to look beyond simple overall accuracy and report behavior-wise sensitivity and precision, as overall accuracy can be misleading when class durations are naturally imbalanced [82].

Essential Toolkit for the Researcher

The following tools and software packages are instrumental in implementing the protocols described above.

Table 3: Key Software and Analysis Tools

Tool Name	Application in Protocol	Relevant Citations
Animal Tag Tools Wiki (MATLAB)	Pre-processing and calibration of accelerometer and magnetometer data.	[83]
Hidden Markov Models (HMMs)	Classifying behavioral states from sensor data; effective for time-series with serial autocorrelation.	[83] [85]
Random Forest	A machine learning algorithm for supervised classification of behaviors based on engineered features.	[82] [84]
AlphaTracker	A video-based tool for multi-animal, markerless pose estimation and behavioral analysis, useful for annotation.	[86]
ezTrack	A free software for video analysis, including positional tracking and freeze analysis.	[87]

The analysis of animal tracking data, which increasingly combines GPS locations with accelerometer data, relies on sophisticated statistical models to infer habitat selection and movement behaviors. Key among these are Spatial Logistic Regression Models (SLRMs), Spatio-Temporal Point Process Models (ST-PPMs), and Integrated Step Selection Models (iSSMs) [41]. These models differ in their theoretical foundations, their approach to critical issues like autocorrelation in tracking data, and ultimately, their statistical performance. This application note provides a comparative analysis of these methods, offering guidance on their selection and implementation for researchers in movement ecology. The insights are framed within the broader context of a thesis utilizing integrated GPS-accelerometer tracking, where classifying behavior from accelerometers is a key prerequisite for detailed movement analysis [45] [88].

A simulation-based study directly compared SLRMs, ST-PPMs, SSMs, and iSSMs for inferring local resource selection and large-scale attraction/avoidance. The study assessed models based on their Type I error rates (false positive rate) and statistical power (ability to detect a true effect) [41].

Table 1: Statistical Performance Comparison of Habitat Selection Models

Model	Type I Error Rate	Statistical Power	Key Strengths	Key Limitations
SLRM	Frequently and strongly exceeds nominal levels [41]	Not Specified	Conceptual simplicity [41]	Neglects spatio-temporal autocorrelation, leading to inflated Type I errors [41]
ST-PPM	Nominal (acceptable) in all studied cases [41]	Robust, but on average lower than iSSM [41]	Directly models spatio-temporal structure; mathematically rigorous handling of availability [41]	View autocorrelation as a nuisance; longer computation times [41]
iSSM	Nominal (acceptable) in all studied cases [41]	Highest average power [41]	Integrates movement and habitat selection; robust power; short computation times; predictive capacity [41]	More complex model formulation [41]

The core finding is that only iSSMs and ST-PPMs maintained statistically acceptable Type I error rates across all scenarios tested. The iSSM approach demonstrated superior statistical power compared to ST-PPMs, making it the most robust method for accurately identifying factors influencing animal movement and space use [41].

Experimental Protocols for Model Evaluation

The following protocol is adapted from the comparative simulation study that evaluated these models [41].

Protocol: Simulated Habitat and Animal Tracking Data Generation

Objective: To generate controlled, realistic data with known properties for benchmarking model performance.

Workflow:

Materials:

• Computing environment (e.g., R, Python)
• Spatial data simulation package (e.g., Raster in R)

Procedure:

• Habitat Simulation: Create 400 distinct raster landscapes with randomly distributed habitat properties to represent ecological covariates [41].
• Movement Parameterization: For each landscape, randomly generate a set of animal movement properties. These should include:
  • Directional persistence: The tendency to continue moving in a similar direction.
  • Large-scale attraction/avoidance: An attraction center within the landscape.
  • Local habitat attraction: The strength of attraction to specific habitat features.
  • Random walk component: Unexplained movement variance [41].
• Track Generation: Use an individual-based movement model to simulate animal tracks on these landscapes. For each parameter set:
  • Generate five tracks where movement is influenced by local habitat attraction (to test statistical power).
  • Generate five tracks where movement is independent of local habitat (to test Type I error rates) [41].
  • This results in approximately 4,000 animal tracks for subsequent model testing.

Protocol: Model Fitting and Validation

Objective: To fit the SLRM, ST-PPM, and iSSM models to the simulated data and evaluate their performance.

Procedure:

• Data Preparation: For each simulated track, prepare the data according to the requirements of each model.
  • For SLRM and ST-PPM: Combine used animal locations with a set of "available" locations (pseudo-absences or dummy points) generated across the landscape [41].
  • For iSSM: For each observed step (movement between two consecutive locations), generate a set of alternative, available steps the animal could have taken [41].
• Model Fitting: Apply each of the six statistical models (from the three main classes) to the prepared datasets. Ensure that ST-PPMs appropriately account for spatio-temporal autocorrelation and that iSSMs simultaneously estimate movement and resource selection parameters [41].
• Performance Assessment:
  • Type I Error Rate: For tracks simulated without habitat selection, calculate the proportion of times each model incorrectly detected a significant habitat effect (false positive).
  • Statistical Power: For tracks simulated with habitat selection, calculate the proportion of times each model correctly identified the known habitat effect (true positive) [41].

The Scientist's Toolkit: Essential Research Reagents & Materials

Table 2: Key Materials and Tools for Animal Movement Analysis

Item	Function/Description	Relevance to Models
GPS/GPRS Tracking Device	Logs spatio-temporal location of animals. Devices should allow for programmable fix intervals, as shorter intervals (e.g., 1 min) generally provide higher accuracy [72].	Provides the fundamental (x, y, t) location data used as the response variable in all models (SLRM, ST-PPM, iSSM).
Tri-axial Accelerometer	Senses acceleration forces, allowing inference of animal behavior (e.g., flying, resting) and energy expenditure via metrics like ODBA [45] [88].	Critical for defining behavioral states, which can be used to fit separate iSSMs for different behaviors or as covariates within a single model.
Movement Data Repository (e.g., Movebank)	Online platform for managing, storing, sharing, and visualizing animal tracking data [72].	Facilitates data management and collaboration prior to analysis with any of the models.
Computational Environment (R/Python)	Provides the software and statistical computing framework for implementing models. Key R packages include glmm for SLRMs, inlabru for ST-PPMs, and amt or animove for iSSMs.	Essential for executing the statistical procedures and computations required by all models.
Kalman Filter Integration Algorithm	A data fusion technique that optimally integrates GPS and accelerometer data, improving location estimates and robustness in complex environments [89].	Can be used for pre-processing location data to reduce noise before it is used in any of the habitat selection models.

Based on the comparative analysis, Integrated Step Selection Models (iSSMs) are the recommended method for inferring habitat selection from animal tracking data. iSSMs provide the most robust statistical properties, maintaining nominal Type I error rates while offering the highest statistical power to detect true effects [41]. Their key advantage lies in explicitly integrating the animal's movement mechanism with its habitat selection process, thereby correctly handling the inherent autocorrelation in tracking data. Furthermore, benefits like shorter computation times and predictive capacity make iSSMs a versatile and powerful tool for movement ecology [41].

For researchers whose primary focus is on the spatial point pattern of animal locations rather than the movement steps, Spatio-Temporal Point Process Models (ST-PPMs) are a mathematically rigorous and valid alternative, though they may be less powerful [41]. The use of Spatial Logistic Regression Models (SLRMs) is not recommended for standard tracking data analysis due to their high propensity for generating false positives from unaccounted autocorrelation [41]. The integration of on-board processed accelerometer data, which provides continuous behavioral classification, further enhances the biological relevance and precision of all these models by allowing researchers to analyze habitat selection specific to particular behavioral states [45].

Assessing Statistical Power and Type I Error Rates in Different Methods

The analysis of animal tracking data has been revolutionized by advanced statistical methods, each with distinct strengths and weaknesses. Understanding the statistical power and type I error rates of these methods is crucial for researchers in movement ecology to draw reliable inferences from their data. Statistical power, defined as the probability that a test will correctly reject a false null hypothesis, and type I error, the probability of incorrectly rejecting a true null hypothesis, are fundamental metrics for evaluating methodological performance [41]. The selection of an appropriate analytical approach depends on both the research question and the properties of the tracking data itself, with different methods exhibiting varying susceptibility to false positives and capacity to detect true effects [41] [90].

Recent comparative studies have demonstrated that method selection significantly impacts ecological inferences derived from animal movement data [91]. The complexity of animal movement, characterized by inherent autocorrelation and scale-dependent processes, presents unique challenges for statistical analysis that not all methods handle equally well. Furthermore, the temporal scale of data collection—ranging from high-frequency GPS fixes to coarser sampling intervals—interacts with analytical methods to influence behavioral state estimation and habitat selection analyses [72] [91]. This protocol systematically evaluates the statistical performance of predominant methods used in animal movement analysis, providing researchers with evidence-based guidance for method selection.

Comparative Performance of Statistical Methods

Quantitative Comparison of Method Performance

Table 1: Statistical performance of methods for analyzing animal tracking data

Method	Type I Error Rate	Statistical Power	Key Strengths	Key Limitations
Spatial Logistic Regression Models (SLRMs)	Frequently and strongly exceeds nominal levels [41]	Not reported	Simple implementation [41]	Highly inflated false positive rates; sensitive to autocorrelation [41]
Spatio-Temporal Point Process Models (ST-PPMs)	Nominal across all cases [41] [90]	Moderate [41] [90]	Automatically handles dummy points; accounts for spatio-temporal autocorrelation [41]	Lower power than iSSMs; population-level viewpoint [41]
Step Selection Models (SSMs)	Slightly exceeds nominal levels [41]	Moderate [41]	Individual-level perspective; reasonable dummy point location [41]	May slightly exceed type I error rates [41]
Integrated Step Selection Models (iSSMs)	Nominal across all cases [41] [90]	High and robust [41] [90]	High statistical power; handles autocorrelation via stratification; predictive capacity [41]	Requires appropriate stratification [41]
Hidden Markov Models (HMMs)	Performance varies with temporal scale [91]	Identifies 3-5 behavioral states [91]	Estimates discrete behavioral states; handles regular time series [91]	Assumes Markov process; may not capture complex behavioral patterns [91]
Move Persistence Models (MPMs)	Performance varies with temporal scale [91]	Identifies fine-scale patterns at 1h resolution [91]	Estimates continuous behavioral parameter; identifies fine-scale patterns [91]	Less effective at coarser temporal scales [91]
Mixed-Membership Method for Movement (M4)	Performance varies with temporal scale [91]	Similar to HMMs [91]	Fewer assumptions than HMMs; handles missing values [91]	Segment-level approach weights metrics with available data [91]

Impact of Temporal Scale on Method Performance

The temporal scale of tracking data significantly influences method performance and behavioral state estimation. Research comparing movement persistence models (MPMs), hidden Markov models (HMMs), and mixed-membership methods for movement (M4) has demonstrated that sampling movement at coarser time scales smooths estimates of behavioral transitions [91]. At longer time steps (e.g., 8 hours), all three models effectively distinguish area-restricted search behavior from migratory behavior, with HMMs and M4 providing greater nuance. Conversely, MPMs were the only models that successfully identified fine-scale behavioral patterns when analyzing short time steps (1 hour) in green sea turtles, revealing likely periods of resting during long-distance migration that were previously only hypothesized [91].

Table 2: Method performance across temporal scales

Temporal Scale	Optimal Methods	Behavioral Insights Achievable	Method Limitations
Fine-scale (1h)	MPMs [91]	Identifies resting periods during migration; fine-scale behavioral patterns [91]	HMMs and M4 lose fine-scale resolution [91]
Intermediate (4h)	HMMs, M4, MPMs [91]	Balanced detail and generalization	Trade-off between fine and coarse pattern detection [91]
Coarse-scale (8h)	HMMs, M4 [91]	Distinguishes ARS from migration; broader behavioral classification [91]	MPMs lose effectiveness [91]

Experimental Protocols for Method Evaluation

Protocol for Simulating Animal Tracking Data

Purpose: To generate standardized animal tracking data for comparing statistical method performance under controlled conditions [41].

Workflow:

• Habitat Generation: Simulate 400 different habitats with randomly varying habitat properties to represent diverse environmental conditions [41].
• Movement Property Assignment: For each habitat, randomly generate animal movement properties including directional persistence, random walk components, and response to environmental features [41].
• Track Generation: Create 10 animal tracks per habitat (5 with attraction effects, 5 without) totaling approximately 4,000 tracks [41].
• Effect Introduction: Incorporate known strengths of local habitat attraction and large-scale attraction/avoidance processes to evaluate statistical power [41].
• Null Model Tracks: Generate control tracks without attraction effects to assess type I error rates [41].

Validation: Implement published simulation frameworks that incorporate four key movement influences: (1) local habitat attraction, (2) directional persistence, (3) large-scale attraction centers, and (4) random walk components [41].

Protocol for Method Implementation and Testing

Purpose: To systematically evaluate statistical methods using simulated tracking data [41].

Implementation Steps:

• Method Application: Apply each statistical method (SLRM, ST-PPM, SSM, iSSM, HMM, MPM, M4) to the simulated tracks [41] [91].
• Parameter Estimation: For each method, estimate parameters representing habitat selection and movement processes [41].
• Hypothesis Testing: Conduct standardized hypothesis tests for habitat selection and attraction/avoidance effects [41].
• Performance Calculation:
  • Compute type I error rates as the proportion of significant results in null effect tracks [41].
  • Compute statistical power as the proportion of significant results in tracks with known effects [41].
• Comparison: Compare performance metrics across methods and conditions [41].

Specialized Considerations:

• For HMMs: Pre-specify the number of behavioral states based on biological knowledge and validate state interpretation [91].
• For iSSMs: Implement appropriate data stratification to account for autocorrelation while estimating resource selection [41].
• For M4: Account for the mixed-membership approach that segments tracks into homogenous periods before clustering [91].

The Researcher's Toolkit: Essential Materials and Reagents

Tracking Technologies and Data Collection Tools

Table 3: Essential research tools for animal movement studies

Tool Category	Specific Examples	Key Functions	Performance Considerations
GPS Tracking Devices	Movetech Telemetry Flyways-50 [72]	Records animal positions; solar-powered; remote data transmission	Accuracy: 3.4-6.5m horizontal, 4.9-9.7m vertical [72]
Inertial Measurement Units (IMUs)	Daily Diaries [92]	Measures acceleration, magnetic field intensity, pressure; dead-reckoning	Enables fine-scale movement reconstruction [92]
Integrated Sensor Suites	Wildlife Computers SPLASH10-F-385A [91]	Combines Argos and Fastloc GPS; multiple sensors in single package	Varying location errors depending on habitat and technology [91]
High-Precision GNSS	Leica RTK GNSS [93]	High-accuracy positioning; dual-frequency; 20Hz sampling	Reference standard for speed and position validation [93]
Consumer-Grade GNSS	Garmin Forerunner 305 [93]	Cost-effective positioning; suitable for large-scale deployments	Lower sampling rates (1Hz); higher latency [93]

Analytical Frameworks and Software Solutions

Dead-Reckoning Implementation: The Gundogs.Tracks() function in R provides comprehensive dead-reckoning capabilities, incorporating speed filtering, track scaling, and drift correction algorithms [92]. This approach significantly enhances path resolution between verified positions while optimizing battery life of primary tracking systems.

Movement Analysis Platforms: Movebank serves as a centralized data repository for animal tracking data, facilitating data management, visualization, and analysis across research groups [72]. This platform supports various data formats from different tracking technologies and enables standardized methodological comparisons.

Statistical Programming Environments: R provides extensive capabilities for implementing movement analyses including HMMs, SSMs, iSSMs, and specialized packages for trajectory analysis, resource selection, and behavioral state estimation [91] [41]. The flexibility of programming-based analysis facilitates method customization and simulation studies.

Advanced Methodological Considerations

Handling GPS Error and Data Quality

The accuracy of GPS tracking devices varies based on fix acquisition intervals and environmental conditions. Research demonstrates that average horizontal accuracy ranges between 3.4 to 6.5 meters, while vertical accuracy varies between 4.9 to 9.7 meters across high-frequency (1-minute) and low-frequency (60-minute) GPS fix intervals [72]. The GPS-Error metric provided by some devices can effectively identify inaccurate positions (>10 meters) in high-frequency intervals (eliminating over 99% of inaccurate positions by removing the 3% of data with highest GPS-Error), though this metric proves less effective for low-frequency intervals [72].

Dead-Reckoning Correction Protocols

Dead-reckoning using inertial measurement units (IMUs) significantly enhances the resolution of animal movement paths between verified positions, but requires careful correction for accumulating drift. The optimal frequency for verified position (VP) correction depends on the movement medium and species [92]. Research demonstrates that dead-reckoning error is greatest for animals travelling within air and water compared to terrestrial environments, requiring more frequent correction for aerial and aquatic species [92].

Protocol for VP-Corrected Dead-Reckoning:

• Data Collection: Deploy integrated IMU and GPS tags recording tri-axial acceleration, magnetic-field intensity, and pressure [92].
• Path Reconstruction: Implement dead-reckoning algorithms to sequentially integrate travel vectors (heading and speed estimates) [92].
• Drift Assessment: Quantify position error as the distance between temporally aligned dead-reckoned and VP positions [92].
• Correction Optimization: Determine optimal VP correction rate based on movement type:
  • Terrestrial species (e.g., lions): Lower correction frequency sufficient
  • Aquatic species (e.g., penguins): Moderate correction frequency required
  • Aerial species (e.g., tropicbirds): Higher correction frequency necessary [92]
• Environmental Flow Incorporation: Account for external air or tidal flow vectors for animals moving in fluid media [92].

Based on comprehensive performance evaluations, integrated step selection models (iSSMs) generally provide the optimal balance of nominal type I error rates and high statistical power for inferring habitat selection or large-scale attraction/avoidance from animal tracking data [41] [90]. Additional advantages include relatively short computation times, predictive capacity, and the ability to derive mechanistic movement models [41]. However, method selection should be guided by specific research questions, with hidden Markov models (HMMs) and mixed-membership methods (M4) better suited for discrete behavioral state estimation, particularly at coarser temporal scales [91].

Researchers should carefully consider the temporal scale of their data collection relative to their ecological questions, as this significantly influences methodological performance [91]. For fine-scale movement analysis (e.g., 1-hour intervals), move persistence models (MPMs) outperform other approaches, while HMMs and M4 provide superior behavioral classification at coarser scales (e.g., 8-hour intervals) [91]. Method selection should also account for species-specific movement characteristics and environmental contexts, with particular attention to handling autocorrelation structures inherent to animal tracking data [41].

Future methodological development should focus on enhancing model performance for species moving in fluid environments (aerial and aquatic), where dead-reckoning error accumulates most rapidly and requires specialized correction approaches [92]. Integration of multiple data streams from complementary technologies, including high-resolution GPS, IMUs, and environmental sensors, will continue to improve the statistical power and reliability of animal movement analyses across diverse ecological contexts.

Uncertainty Quantification in Behavioral Classification

Behavioral classification using animal-borne sensors, particularly accelerometers, has revolutionized movement ecology by enabling remote, continuous monitoring of animal behavior. However, the models that translate complex sensor data into behavioral categories are inherently uncertain. Uncertainty quantification (UQ) provides the critical framework for assessing the reliability of these classifications, distinguishing between well-supported predictions and speculative inferences. Within the broader context of GPS tracking and accelerometer data analysis research, UQ transforms behavioral classification from a black-box prediction into a scientifically rigorous measurement process with defined confidence boundaries. This is particularly vital when these classifications inform conservation policies, ecological interpretations, or physiological studies [47] [13].

The fundamental challenge stems from multiple sources: model limitations in capturing biological complexity, data quality issues from sensor noise, and behavioral ambiguity where distinct behaviors produce similar sensor signatures. Furthermore, as research scales from individual animals to populations, understanding how uncertainty propagates through analytical pipelines becomes essential for robust ecological inference [94]. This application note establishes comprehensive protocols for quantifying, managing, and reporting these uncertainties throughout the behavioral classification workflow.

Theoretical Foundations of Uncertainty in Behavioral Classification

In behavioral classification systems, uncertainty arises from a cascade of sources throughout the data lifecycle. The GBADs programme framework categorizes uncertainty into epistemic (from limited knowledge), ontological (from defining system boundaries), and ambiguous (from unclear terminology) types, each operating at substantive, strategic, and institutional levels [94] [95].

Data acquisition uncertainty originates at the sensor level, including measurement errors from accelerometer calibration drift, GPS positioning inaccuracies, and temporal sampling limitations. MEMS accelerometers, while cost-effective for large deployments, introduce uncertainty through sensitivity nonlinearity, cross-axis effects, and environmental influences on performance [96]. Model structure uncertainty emerges from selecting algorithms, defining behavioral categories, and choosing input features. For instance, combining behaviors with similar kinematic signatures increases classification ambiguity [97].

Parameter uncertainty relates to the estimated coefficients within models, while projection uncertainty concerns the applicability of models trained in captive environments to wild contexts [95]. Each uncertainty type propagates through the analytical chain, ultimately affecting the confidence in final behavioral assignments and subsequent ecological conclusions.

Information Theory in Movement Analysis

Information theory provides a mathematical foundation for quantifying uncertainty in animal movement tracks. By treating movement paths as information streams, researchers can apply Shannon's Information Theory to measure the predictability and information content of behavioral sequences [47].

This approach involves decomposing movement into smallest viable statistical elements (StaMEs) and clustering them into canonical activity modes (CAMs). The Jensen-Shannon divergence measure then assesses differentiation between behavioral clusters, while entropy measures quantify the predictability of behavioral sequences [47]. This theoretical framework enables researchers to compute coding efficiencies of derived movement elements and establish error rates in behavioral assignments, providing a rigorous quantitative foundation for uncertainty assessment in path segmentation analysis.

Table 1: Classification Accuracy Under Different Experimental Conditions

Model Type	Number of Behaviors	Epoch Length (samples)	Reported Accuracy	Key Uncertainty Factors
Super Learner [97]	4	7 (0.28s)	Highest	Model selection, epoch definition
Super Learner [97]	6	7 (0.28s)	Reduced	Behavioral category distinction
Various ML algorithms [97]	4	75 (3s)	Lower	Temporal resolution loss
Hidden Semi-Markov Models [97]	2	Variable	Higher	Fewer categorical distinctions
Decision Trees [97]	7	Variable	Low (attack/peck unclassifiable)	Behavioral complexity

Table 2: Sensor-Related Uncertainty Contributions

Uncertainty Source	Typical Magnitude	Impact on Classification	Mitigation Strategies
MEMS Accelerometer Sensitivity [96]	1.0-1.5% (standard uncertainty)	Medium	Laboratory calibration with uncertainty budget
GPS Position Error [98]	3-10 meters	Context-dependent	Higher fix rates, filtering algorithms
Device Attachment [97]	Unquantified	High	Standardized attachment methods
Sampling Frequency [97]	7-75 epochs tested	Medium	Match to behavioral kinetics
Battery Power Limitations [99]	Variable deployment duration	High	Solar augmentation, power management

Methodological Framework for Uncertainty Quantification

Integrated Uncertainty Quantification Workflow

The following workflow integrates multiple UQ approaches throughout the behavioral classification pipeline:

Experimental Protocols for Uncertainty Assessment

Protocol 1: Sensor Calibration and Validation

Purpose: Quantify and minimize measurement uncertainty from accelerometers prior to deployment.

Materials: Custom calibration test bench with precise rotation control (e.g., high-precision turntable), reference accelerometers, environmental chamber for temperature testing, data acquisition system [96].

Procedure:

• Mount accelerometers on calibration bench ensuring precise alignment to minimize cross-axis sensitivity effects.
• Apply known acceleration profiles (typically ±1g to ±8g range) across operational frequency range (0-100Hz for most behavioral studies).
• Record sensor outputs simultaneously with reference measurements across multiple axes.
• Calculate sensitivity, linearity, cross-axis sensitivity, and bias parameters for each sensor.
• Develop uncertainty budget accounting for reference instrument uncertainty, alignment errors, environmental conditions, and repeatability.
• Validate calibration with independent acceleration profiles not used in calibration.

Uncertainty Quantification: Express results as expanded uncertainty with coverage factor k=2 (approximately 95% confidence level). Report combined standard uncertainty incorporating Type A (statistical) and Type B (systematic) components [96].

Protocol 2: Super Learner Algorithm Implementation

Purpose: Implement ensemble machine learning to reduce model uncertainty in behavioral classification.

Materials: Annotated accelerometer dataset with corresponding video validation, computing infrastructure capable of parallel processing, software environment (R or Python with appropriate machine learning libraries) [97].

Procedure:

• Data Preparation: Segment accelerometer data into epochs (recommended: 7-13 samples at 25Hz for otariid pinnipeds). Extract 147 summary statistics including mean, median, standard deviation, skewness, kurtosis, percentiles, axis correlations, ODBA, and VeDBA [97].
• Base Learner Selection: Implement diverse set of candidate algorithms including Random Forests, Support Vector Machines, Neural Networks, and Gradient Boosting Machines.
• Training Configuration: Employ V-fold cross-validation (typically V=10) to prevent overfitting and provide robust performance estimation.
• Super Learner Implementation:
  • Train all base learners on cross-validated folds
  • Create optimal weighted combination of base learners minimizing cross-validated risk
  • Validate ensemble model on held-out test dataset
• Uncertainty Metrics: Calculate classification probabilities, confusion matrices, and out-of-bag error rates. Perform sensitivity analysis on epoch length and behavioral category definitions.

Uncertainty Quantification: Report cross-validated accuracy with variance estimates. Compare super learner performance against individual base algorithms. Quantify improvement in classification variance reduction [97].

Protocol 3: Behavioral Validation in Captive Settings

Purpose: Establish ground-truth dataset for model training and quantify validation uncertainty.

Materials: Triaxial accelerometers (e.g., CEFAS G6a+), synchronized video recording systems (e.g., GoPro Hero cameras), data synchronization software, captive animal facilities with appropriate ethics approvals [97].

Procedure:

• Device Attachment: Secure accelerometers between shoulder blades using either tape or custom-designed harnesses to minimize movement artifacts.
• Video Recording: Record continuous behavioral observations with time-synchronized cameras covering entire observation periods.
• Ethogram Development: Create detailed ethogram with explicit behavioral definitions. For fur seals and sea lions, this may include 26 unique behaviors aggregated into 4-6 broader categories (feeding, grooming, resting, traveling) [97].
• Data Annotation: Time-match video observations with accelerometer outputs to create labeled dataset.
• Inter-observer Reliability: Multiple trained observers should score identical sequences to quantify annotation consistency using Cohen's Kappa or similar metrics.

Uncertainty Quantification: Report inter-observer reliability statistics. Document any behavioral sequences with ambiguous classification. Quantify temporal alignment precision between video and sensor data [97].

Essential Research Reagent Solutions

Table 3: Key Research Materials and Analytical Solutions

Category	Specific Product/Technique	Function in Uncertainty Quantification	Implementation Considerations
Sensor Systems	Triaxial MEMS Accelerometers (e.g., STMicroelectronics LSM6DSR) [96]	Capture raw movement data with minimal power consumption	Require laboratory calibration; assess cross-axis sensitivity
Biologging Platforms	Gipsy Remote (Technosmart) [98]	Integrated GPS-accelerometer data collection	Solar-battery capability enables long-term deployment
Calibration Equipment	Precision rotating table with angular encoder [96]	Generate known acceleration profiles for sensor characterization	Enables simultaneous multi-axis calibration
Machine Learning Algorithms	Super Learner ensemble method [97]	Optimally combines base learners to reduce model variance	Computationally intensive but provides superior accuracy
Validation Tools	Synchronized video-accelerometer systems [97]	Establish ground truth for behavioral classification	Requires standardized ethograms and inter-observer reliability assessment
Data Processing	Dynamic Body Acceleration (ODBA, VeDBA) metrics [97]	Summarize complex acceleration signals into ecologically relevant features	Standardized calculation enables cross-study comparisons

Discussion and Implementation Guidelines

The protocols outlined establish a comprehensive framework for quantifying uncertainty throughout the behavioral classification pipeline. Implementation requires careful consideration of trade-offs between analytical complexity and practical utility.

Model Selection Trade-offs: While super learning demonstrates superior accuracy and reduced variance compared to individual machine learning algorithms, it demands substantial computational resources and expertise [97]. For resource-limited projects, well-implemented Random Forests or Support Vector Machines may provide acceptable performance with lower complexity.

Behavioral Categorization Impact: The granularity of behavioral classification directly impacts uncertainty. Studies requiring detailed ethograms (6+ behaviors) should anticipate higher classification uncertainty compared to broader categorical systems (4 behaviors) [97]. Research objectives should drive this balance between resolution and reliability.

Temporal Scaling Considerations: Epoch length significantly influences classification accuracy. Shorter epochs (0.28-0.52 seconds) generally outperform longer windows (3 seconds) for capturing discrete behavioral elements [97]. However, longer windows may better characterize sustained behavioral states. Matching epoch length to behavioral kinetics is essential.

Uncertainty Communication: Following the GBADs framework, clearly document all uncertainty sources, modeling assumptions, data quality rankings, and validation results [94]. This transparency enables proper interpretation of behavioral classifications and supports meta-analytical approaches across studies.

As accelerometer technologies advance and machine learning methods become more sophisticated, the framework presented here provides a foundation for increasingly rigorous uncertainty quantification in behavioral classification. This rigor transforms animal-borne sensors from mere data collection devices into properly calibrated scientific instruments for behavioral measurement.

The analysis of animal movement and behavior through biologging technologies, such as GPS and accelerometers, represents a cornerstone of modern movement ecology. Within the broader context of GPS tracking and accelerometer data research, this application note addresses a critical task: the automated classification of specific, fine-scale behaviors like grazing, ruminating, resting, and walking. Accurate behavioral classification is fundamental to studies in animal ecology, welfare assessment, and conservation biology, as it transforms raw sensor data into biologically meaningful information [10] [13]. This case study synthesizes recent research to compare the performance of classification methods across different behaviors and species, provides detailed protocols for implementing these methods, and visualizes the underlying frameworks.

Research demonstrates that classification accuracy is highly behavior-dependent. Certain behaviors have distinct movement signatures, leading to high classification accuracy, while others are more challenging to distinguish. The following tables summarize findings from key studies.

Table 1: Classification Accuracy for Cattle Behaviors Using a Random Forest Model (Data from [10])

Behavior	Classification Accuracy
Grazing	0.93
Ruminating	0.90
Laying	0.89
Steady Standing	0.88

Table 2: Comparison of Classification Method Accuracy for Seabird Behaviors (Data from [88])

Classification Method	Thick-billed Murres (Accuracy)	Black-legged Kittiwakes (Accuracy)
Overall Average Accuracy	>98%	89% (Incubation) to 93% (Chick Rearing)
Movement Thresholds	>98%	Information not specified in source
k-Means Clustering	>98%	Information not specified in source
Random Forest	>98%	Information not specified in source
Hidden Markov Models	>98%	Information not specified in source

Table 3: Behavior Classification Based on Movement Parameters (Data from [52])

Behavioral State	Mean Speed	Mean Turning Angle
Resting	Low	Low
Walking	High	Low
Foraging	Low	High

Experimental Protocols for Behavior Classification

The process of classifying behavior from sensor data involves a sequence of critical steps, from data collection to model validation. The workflow below outlines this general process, which is followed by detailed protocols for two common analytical approaches.

Protocol 1: Supervised Classification with Random Forests

This protocol leverages a supervised machine learning approach, which requires a labeled dataset for training.

• Step 1: Sensor Deployment and Data Collection

  • Deploy tri-axial accelerometer sensors securely on animals, preferably on the neck for cattle to capture head-movement-related behaviors like grazing and ruminating [10].
  • Sample accelerometer data at a minimum of 10 Hz to capture fine-scale movements [10] [59].
  • Collect synchronized video recordings of the instrumented animals for a sufficient duration to capture all behaviors of interest. This video serves as the "ground truth" for labeling the sensor data.

• Step 2: Data Labeling and Preprocessing

  • Manually label the accelerometer data stream based on the synchronized video, assigning each time segment a behavioral class (e.g., Grazing, Ruminating, Laying) [10].
  • Split the continuous sensor data into analysis windows (e.g., 3-10 second epochs).
  • For GPS data, calculate movement parameters such as step length (straight-line distance between locations), speed, and turning angle (change in direction) [8] [52].

• Step 3: Feature Extraction

  • From each accelerometer data window, extract features in both the time and frequency domains. One study extracted 108 such features [10].
  • Key features include:
    • Static Acceleration: The constant gravitational component, used to calculate body posture (pitch and roll) [13] [59].
    • Dynamic Acceleration: The high-frequency variation due to movement [13].
    • Overall Dynamic Body Acceleration (ODBA) / Vectoral Dynamic Body Acceleration (VeDBA): The sum of the dynamic components of acceleration across all axes, a proxy for energy expenditure [8] [59].
    • Spectral Features: The dominant frequency and amplitude from a Fast Fourier Transform (FFT) of the signal, which can help identify repetitive movements like walking or chewing [88].

• Step 4: Model Training and Validation

  • Use a Random Forest (RF) classifier, which is robust and often provides high accuracy [10] [88] [59].
  • Split the labeled feature dataset into a training set (e.g., 60-80%) and a testing set (e.g., 20-40%).
  • Train the RF model on the training set. The model will learn the complex relationships between the extracted features and the behavioral labels.
  • Predict behaviors on the held-out test set and compare them to the ground truth labels to calculate accuracy, precision, recall, and F-measure.

• Step 5: Model Optimization

  • Refine Variables: Test different combinations of features to identify the most parsimonious set that maintains high accuracy [88] [59].
  • Adjust Data Frequency: Experiment with different data resolutions. Higher frequencies (>10 Hz) may better capture fast locomotion, while lower resolutions (e.g., 1 Hz mean) can improve identification of slower, aperiodic behaviors like grooming or feeding [59].
  • Balance Training Data: Ensure the training dataset contains a roughly equal duration of each behavior to prevent the model from being biased toward the most common behaviors [59].

Protocol 2: Segmentation and Clustering of Movement Paths

This protocol is well-suited for GPS data and aims to identify behavioral states directly from movement geometry without pre-defined labels.

• Step 1: Data Preparation and Parameter Calculation

  • Obtain a high-frequency GPS trajectory (a time-series of locations).
  • From the trajectory, calculate primary movement parameters for each step between locations:
    • Persistence Velocity: The speed of movement in the direction of heading [8] [52].
    • Turning Angle/Veocity: The change in direction from one step to the next [8] [52].

• Step 2: Trajectory Segmentation

  • Apply Behavioral Change Point Analysis (BCPA) to the time series of movement parameters [52].
  • BCPA uses a moving window to identify significant points where the mean and variance of the movement parameters (e.g., persistence velocity) change substantially. This partitions the trajectory into segments that are internally homogeneous in terms of movement.

• Step 3: Segment Clustering

  • Characterize each segment by the probability distribution of its movement parameters (e.g., speed, turning angle).
  • Compute a distance matrix between all segments using a statistical distance metric like the Kolmogorov-Smirnov distance, which compares distributions.
  • Perform agglomerative hierarchical clustering on this distance matrix to group segments with similar movement characteristics [52].

• Step 4: Behavioral State Inference

  • Interpret the resulting clusters as distinct behavioral states. For example:
    • Cluster 1 (Low speed, Low turning): Interpret as Resting.
    • Cluster 2 (High speed, Low turning): Interpret as Walking/Commuting.
    • Cluster 3 (Low speed, High turning): Interpret as Foraging/Grazing [52].
  • The accuracy of this method can be validated against field observations, with one study reporting an average agreement of 80.75% [52].

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 4: Key Materials and Software for Behavioral Classification Experiments

Item Name	Function/Application	Specification Notes
Tri-axial Accelerometer	Measures acceleration in three orthogonal planes (surge, sway, heave), capturing posture and fine-scale movement [10] [13].	Sample rate ≥10 Hz; Dynamic range typically ±2g to ±8g; MEMS-based for low power and size.
GPS Logger	Records animal position over time, enabling analysis of movement paths and speeds [10] [8].	Configurable fix intervals (e.g., 5 min to 1 sec); Error <5m; Low power consumption to extend battery life.
Animal Collar/Harness	Secures sensors to the study animal with minimal impact on natural behavior [10].	Weatherproof casing; Species-appropriate attachment; Secure but non-restrictive.
Video Recording System	Provides ground truth data for labeling accelerometer signals and validating models [10] [59].	Synchronized timekeeping with sensors; Sufficient resolution and frame rate to identify behaviors.
R or Python Software	Data processing, feature extraction, machine learning, and statistical analysis [88] [59].	Key packages: accelerometry, moveHMM, scikit-learn, caret, adehabitatLT.
Random Forest Classifier	A supervised machine learning algorithm that achieves high accuracy for behavioral classification tasks [10] [88] [59].	Robust to overfitting; Handles large numbers of features well.

Technology and Analysis Workflow

The core technology of accelerometry is based on measuring the components of movement. The diagram below illustrates how raw sensor data is decomposed and translated into meaningful behavioral and ecological metrics.

Conclusion

The integration of GPS and accelerometer data, when supported by rigorous calibration, appropriate analytical models, and thorough validation, provides an unparalleled window into animal behavior and movement ecology. The field is moving toward more accessible, reproducible, and powerful analytical platforms that can handle the increasing volume and complexity of bio-logging data. Future directions will likely involve greater integration of AI and machine learning for automated behavioral classification, the development of more sophisticated multi-sensor fusion techniques, and the creation of standardized protocols that ensure data comparability across studies and species. For biomedical and clinical research, these methodologies offer a framework for quantitatively assessing animal models of disease, monitoring the efficacy of therapeutic interventions, and understanding behavioral phenotypes in unprecedented detail, ultimately strengthening the translational pathway from basic research to clinical application.

References

• 1. (Wearbles) - GPS , Validity and Reliability | Science for Sport Technology [https://www.scienceforsport.com/gps-wearables-validity-and-reliability/]
• 2. Inside the GPS Unit | Global Perf Insights [https://www.globalperformanceinsights.com/post/inside-the-gps-unit]
• 3. Frontiers | Editorial: Advances in modelling and analysis of animal ... [https://www.frontiersin.org/journals/ecology-and-evolution/articles/10.3389/fevo.2025.1655346/full]
• 4. DeepBhvTracking: A Novel Behavior Tracking Method for ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC8581673/]
• 5. ECODATA and MoveApps: new tools for animating wildlife ... [https://conservationcorridor.org/digests/2023/05/ecodata-and-moveapps-new-tools-for-animating-wildlife-movements-across-corridors/]
• 6. Collection: Towards a mechanistic understanding of animal ... movement [https://link.springer.com/collections/bbffhffgdd]
• 7. Animations of wildlife tracking data help explore animal ... [https://news.osu.edu/animations-of-wildlife-tracking-data-help-explore-animal-movements/]
• 8. Ecological metrics and methods for GPS movement data - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC6322554/]
• 9. Flexible characterization of animal movement pattern using ... [https://movementecologyjournal.biomedcentral.com/articles/10.1186/s40462-016-0080-y]
• 10. Analysis of Accelerometer and GPS Data for Cattle Behaviour Identification and Anomalous Events Detection - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC8947510/]
• 11. Ecological inference using data from accelerometers needs careful protocols - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC9303593/]
• 12. Plunge-diving into dynamic body acceleration and energy ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11607695/]
• 13. Observing the unwatchable through acceleration logging of animal ... [https://animalbiotelemetry.biomedcentral.com/articles/10.1186/2050-3385-1-20]
• 14. Doubly labelled water assessment of energy expenditure [https://pmc.ncbi.nlm.nih.gov/articles/PMC5486561/]
• 15. Validation study of energy expenditure and intake during ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC2662402/]
• 16. Validation of ActiReg® to measure physical activity and ... [https://www.cambridge.org/core/journals/british-journal-of-nutrition/article/validation-of-actireg-to-measure-physical-activity-and-energy-expenditure-against-doubly-labelled-water-in-obese-persons/B834336BEF412F35BACBB6CC426689A0]
• 17. The statistical building blocks of animal movement simulations [https://movementecologyjournal.biomedcentral.com/articles/10.1186/s40462-024-00507-4]
• 18. Integrating animal tracking and trait data to facilitate global ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11883293/]
• 19. A compilation pipeline for wildlife tracking datasets ... [https://www.sciencedirect.com/science/article/pii/S1574954125002298]
• 20. SmartWilds: Multimodal Wildlife Monitoring Dataset [https://arxiv.org/html/2509.18894v2]
• 21. Analysis of Accelerometer and GPS Data for Cattle ... [https://www.mdpi.com/1099-4300/24/3/336]
• 22. Using GPS and accelerometer data to remotely detect ... [https://animalbiotelemetry.biomedcentral.com/articles/10.1186/s40317-024-00385-y]
• 23. How Animals Trackers Works — In One Simple Flow (2025) [https://www.linkedin.com/pulse/how-animals-trackers-works-one-simple-flow-2025-h5otf/]
• 24. Improving the spatial accuracy of wildlife tracking data with ... [https://animalbiotelemetry.biomedcentral.com/articles/10.1186/s40317-025-00431-3]
• 25. A pipeline with pre-processing options to detect behaviour ... [https://peercommunityjournal.org/articles/10.24072/pcjournal.545/]
• 26. Practical guidelines for validation of supervised machine ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12214441/]
• 27. Combined use of tri-axial accelerometers and GPS reveals the flexible... [https://pmc.ncbi.nlm.nih.gov/articles/PMC5462363/]
• 28. Environmental correlates of activity and energetics in a wide-ranging... [https://animalbiotelemetry.biomedcentral.com/articles/10.1186/s40317-021-00272-w]
• 29. A protocol for assessing bias and robustness of social network ... [https://movementecologyjournal.biomedcentral.com/articles/10.1186/s40462-024-00494-6]
• 30. An artificial neural network for automated behavioral state ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC8435206/]
• 31. Introducing machine learning approaches for classification of ... [https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0323893]
• 32. Multi-Sensor Integration and Machine Learning for... | Preprints.org [https://www.preprints.org/manuscript/202503.0614/v1]
• 33. Neural models for detection and classification of brain ... [https://www.nature.com/articles/s42003-025-07991-3]
• 34. Applications of step-selection functions in ecology and ... [https://movementecologyjournal.biomedcentral.com/articles/10.1186/2051-3933-2-4]
• 35. [1512.01614] Integrated : bridging the gap... step selection analysis [https://arxiv.org/abs/1512.01614]
• 36. A 'How to' guide for interpreting parameters in habitat‐ ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC8251592/]
• 37. Methods for implementing integrated - step functions with... selection [https://movementecologyjournal.biomedcentral.com/articles/10.1186/s40462-024-00476-8]
• 38. Rayleigh step-selection functions and connections to ... [https://pubs.usgs.gov/publication/70251443]
• 39. Using step selection functions to analyse human mobility ... [https://elifesciences.org/reviewed-preprints/107153]
• 40. Modelling individual variability in habitat selection and movement ... [https://repository.library.noaa.gov/view/noaa/66703]
• 41. Analysis of local habitat selection and large-scale attraction/avoidance based on animal tracking data: is there a single best method? | Movement Ecology | Full Text [https://movementecologyjournal.biomedcentral.com/articles/10.1186/s40462-021-00260-y]
• 42. Analysis of local habitat selection and large-scale attraction/avoidance based on animal tracking data: is there a single best method? - PMC [https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8063450/]
• 43. A guide to pre‐processing high‐throughput animal tracking data ... [https://ncbi.nlm.nih.gov/pmc/articles/PMC9299236]
• 44. [1911.00151] Estimating animal utilization distributions from multiple... [https://arxiv.org/abs/1911.00151?]
• 45. Increasingly detailed insights in animal behaviours using continuous on-board processing of accelerometer data | Movement Ecology | Full Text [https://movementecologyjournal.biomedcentral.com/articles/10.1186/s40462-022-00341-6]
• 46. AlphaTracker: a multi-animal tracking and behavioral analysis tool - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC10266280/]
• 47. Gaining Insight Through Animal Tracking - Simple Science Movement [https://scisimple.com/en/articles/2025-08-22-gaining-insight-through-animal-movement-tracking--a3wgnn2]
• 48. of carnivore kill sites is improved by verified... Identification [https://animalbiotelemetry.biomedcentral.com/articles/10.1186/s40317-020-00206-y]
• 49. Frontiers | The (misleading) role of animal models in drug development [https://www.frontiersin.org/journals/drug-discovery/articles/10.3389/fddsv.2024.1355044/full]
• 50. Frontiers | Open-source software for automated rodent behavioral analysis [https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2023.1149027/full]
• 51. The concept of animals' trajectories from a data analysis perspective - ScienceDirect [https://www.sciencedirect.com/science/article/abs/pii/S1574954108000654]
• 52. Deriving Animal Movement Behaviors Using Movement Parameters Extracted from Location Data [https://www.mdpi.com/2220-9964/7/2/78]
• 53. Path segmentation for beginners: an overview of current methods for detecting changes in animal movement patterns | Movement Ecology | Full Text [https://movementecologyjournal.biomedcentral.com/articles/10.1186/s40462-016-0086-5]
• 54. A hidden Markov movement model for rapidly identifying behavioral states from animal tracks - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC5383489/]
• 55. : a serverless no-code MoveApps platform for analysis ... animal [https://movementecologyjournal.biomedcentral.com/articles/10.1186/s40462-022-00327-4]
• 56. Optimal Methods of RTK-GPS/Accelerometer Integration to Monitor the Displacement of Structures - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC3279252/]
• 57. Using Accelerometer and GPS Data for Real-Life Physical Activity Type Detection - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC7038120/]
• 58. -Environment Interactions Animal [https://www.ab.mpg.de/safi]
• 59. Identifying animal behaviours from accelerometers: Improving predictive accuracy of machine learning by refining the variables selected, data frequency, and sample duration - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC11097004/]
• 60. GitHub - michaelwro/ accelerometer - calibration : A general method ... [https://github.com/michaelwro/accelerometer-calibration]
• 61. and Programming | Adafruit Analog Calibration ... Accelerometer [https://learn.adafruit.com/adafruit-analog-accelerometer-breakouts/calibration-and-programming]
• 62. Identification of behaviours from accelerometer data in a wild social primate | Animal Biotelemetry | Full Text [https://animalbiotelemetry.biomedcentral.com/articles/10.1186/s40317-017-0121-3]
• 63. sampling requirements for Accelerometer behaviour... animal [https://animalbiotelemetry.biomedcentral.com/articles/10.1186/s40317-023-00339-w]
• 64. in the Real World: 5 Uses You'll Actually See ( Animals ) Trackers 2025 [https://www.linkedin.com/pulse/animals-trackers-real-world-5-uses-youll-actually-see-lqhoc]
• 65. GPS animal tracking - Wikipedia [https://en.wikipedia.org/wiki/GPS_animal_tracking]
• 66. Neckband or backpack? Differences in tag design and their effects on... [https://animalbiotelemetry.biomedcentral.com/articles/10.1186/s40317-016-0104-9]
• 67. Measurement Toolkit - Accelerometers [https://www.measurement-toolkit.org/physical-activity/objective-methods/accelerometers]
• 68. Frontiers | Analysis of dog movement using a single accelerometer in... [https://www.frontiersin.org/journals/veterinary-science/articles/10.3389/fvets.2025.1551341/full]
• 69. Optimal Sensor Placement for Measuring Physical Activity with a 3D Accelerometer - PMC [https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3958275/]
• 70. Wildlife Animal Tracking System using GPS and GSM – IJERT [https://www.ijert.org/wildlife-animal-tracking-system-using-gps-and-gsm]
• 71. Accelerometer - Wikipedia [https://en.wikipedia.org/wiki/Accelerometer]
• 72. Performance of GPS /GPRS tracking devices improves... | PLOS One [https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0265541]
• 73. a marine predator: the effects of precision, resolution... GPS tracking [https://link.springer.com/article/10.1007/s00227-004-1328-4]
• 74. How Do I Optimise Usage in Location Apps? GPS Battery [https://thisisglance.com/learning-centre/how-do-i-optimise-gps-battery-usage-in-location-apps]
• 75. cambridge.org/core/journals/journal-of-navigation/article/abs/weak- gps ... [https://www.cambridge.org/core/journals/journal-of-navigation/article/abs/weak-gps-signal-detection-in-animal-tracking/B36F353530A97D615E1EECE6FACC0EAE]
• 76. Animal Movement Data: GPS Telemetry, Autocorrelation and the Need for Path-Level Analysis | SpringerLink [https://link.springer.com/chapter/10.1007/978-4-431-87771-4_7]
• 77. 6.3 Analysing Spatio- Temporal Data – Measuring change over time [https://generic.wordpress.soton.ac.uk/gem/unit-6/6-3-analysing-spatio-temporal-data-measuring-change-over-time/]
• 78. Disentangling the effects of forage, social rank, and risk on movement ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC2614723/]
• 79. Testing cellular phone-enhanced GPS tracking technology for urban carnivores | Animal Biotelemetry | Full Text [https://animalbiotelemetry.biomedcentral.com/articles/10.1186/s40317-019-0180-8]
• 80. Cell 长文综述：机器学习如何助力网络生物学 | 集智俱乐部 [https://swarma.org/?p=13908]
• 81. 基于深度学习的生物医学数据分析进展 - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC9927609/]
• 82. Biomechanically-Aware Behaviour Recognition using Accelerometers [https://methodsblog.com/2020/06/10/biomechanically-aware-behaviour-recognition-using-accelerometers/]
• 83. Hidden Markov models identify major movement modes in... [https://movementecologyjournal.biomedcentral.com/articles/10.1186/s40462-021-00243-z]
• 84. Analysis of Accelerometer and GPS Data for Cattle Behaviour Identification and Anomalous Events Detection - PubMed [https://pubmed.ncbi.nlm.nih.gov/35327847/]
• 85. Ecological metrics and methods for GPS movement data - PMC [https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6322554/]
• 86. Frontiers | AlphaTracker: a multi- animal and tracking ... behavioral [https://www.frontiersin.org/journals/behavioral-neuroscience/articles/10.3389/fnbeh.2023.1111908/full]
• 87. ezTrack-A Step-by-Step Guide to Behavior Tracking [https://pubmed.ncbi.nlm.nih.gov/34610215/]
• 88. A comparison of techniques for classifying behavior from accelerometers for two species of seabird - PMC [https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6434605/]
• 89. GNSS/accelerometer integrated deformation monitoring algorithm based on sensors adaptive noise modeling - ScienceDirect [https://www.sciencedirect.com/science/article/abs/pii/S0263224123007431]
• 90. of local habitat selection and large-scale attraction/avoidance... Analysis [https://pmc.ncbi.nlm.nih.gov/articles/PMC8063450/]
• 91. Method selection and temporal scale greatly influence ecological inferences on estimated animal behavioral states | Animal Biotelemetry | Full Text [https://animalbiotelemetry.biomedcentral.com/articles/10.1186/s40317-025-00434-0]
• 92. How often should dead-reckoned animal movement paths be corrected for drift? | Animal Biotelemetry | Full Text [https://animalbiotelemetry.biomedcentral.com/articles/10.1186/s40317-021-00265-9]
• 93. Comparison of Global Navigation Satellite System Devices on Speed Tracking in Road (Tran)SPORT Applications - PMC [https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4299074/]
• 94. Frontiers | A framework for handling uncertainty in a large-scale programme estimating the Global Burden of Animal Diseases [https://www.frontiersin.org/journals/veterinary-science/articles/10.3389/fvets.2025.1459209/full]
• 95. A framework for handling uncertainty in a large-scale programme estimating the Global Burden of Animal Diseases - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC11927218/]
• 96. Dynamic Multi-Axis Calibration of MEMS Accelerometers for Sensitivity and Linearity Assessment [https://www.mdpi.com/1424-8220/25/7/2120]
• 97. Super machine learning: improving accuracy and reducing variance of... [https://animalbiotelemetry.biomedcentral.com/articles/10.1186/s40317-017-0123-1]
• 98. Case report: Long-term GPS throughout the breeding and... tracking [https://animalbiotelemetry.biomedcentral.com/articles/10.1186/s40317-024-00397-8]
• 99. How Animals Works — In One Simple Flow ( Trackers ) 2025 [https://www.linkedin.com/pulse/how-animals-trackers-works-one-simple-flow-2025-h5otf]