Cow Study Application

Classifying cattle behavior from low-frequency GPS and accelerometer collar data collected during Montana winters, and modeling the environmental and temporal factors that drive behavioral change.

Project Overview

This project studies low-frequency GPS and triaxial accelerometer sensors attached to cattle collars over Montana winters. At these sampling rates, individual movement signatures are subtle — requiring careful feature engineering and probabilistic modeling to reliably distinguish behaviors such as grazing, resting, and walking.

Beyond classification, the project investigates which environmental and temporal covariates — temperature, daylight duration, snow depth, and others — are associated with shifts in behavioral composition across the herd. The application on this site is a dedicated front end for managing datasets, configuring analyses, and viewing results. The underlying library is open source and available on GitHub.

Data & Sensors

GPS

Collar GPS fixes recorded at low frequency. Positional data informs displacement, range use, and coarse activity level, but is insufficient on its own at these intervals to distinguish fine-grained behaviors.

Low frequencyPositionalRange use

Accelerometer

Triaxial accelerometer bursts capture short windows of raw motion within each sampling epoch. Summary statistics computed from these bursts — means, variances, spectral features — form the primary inputs to the classification models.

TriaxialBurst samplingFeature engineering

Environmental Covariates

Weather station and NOAA records provide daily temperature, precipitation, snow depth, and photoperiod. These covariates are used to model how behavior allocation changes in response to winter conditions.

TemperatureSnow depthPhotoperiod

Classification Models

Hidden Markov Model

HMMs model behavior as a sequence of latent states with Markovian transitions. The temporal structure of collar time series maps naturally onto this framework, simultaneously estimating emission distributions per state and the transition probabilities between them — yielding interpretable outputs comparable across individuals and conditions.

ProbabilisticTemporal structureInterpretable
Algorithm details →

LSTM

Long Short-Term Memory networks learn temporal dependencies directly from the sequence without explicit state parameterization. Evaluated against HMM outputs to characterize the trade-offs between model expressiveness, interpretability, and labeled-data requirements.

Deep learningSequence modelingComparative

Documentation

Data Audit

Coverage, missing epochs, outlier detection, and collar-level summaries of the raw data.

Read more →

Feature Building

Derivation of accelerometer summary statistics and GPS-based features used as model inputs.

Read more →

Unsupervised Clustering

Exploratory clustering of the feature space prior to supervised labeling, used to assess natural groupings and inform state-space design.

Read more →

HMM Algorithm

A detailed walkthrough of the HMM fitting procedure, state selection, and validation against manually labeled segments.

Read more →

HMM Results

Per-collar classification outputs, state composition summaries, and radar charts showing behavioral allocation across the study period.

Read more →