Classify animal behaviour from accelerometer data

Why this is hard

Modern tags record acceleration on three axes in short bursts, and a deployment can produce a great many of them. Each burst tells you how the animal moved in those seconds, but looking at every burst by hand quickly becomes time-consuming, and for large datasets near-impossible.

How Firetail does it: FireSOM

FireSOM is an assistive, unsupervised method. It sorts bursts into groups of similar movement patterns; you decide what each group means.

1. Choose features. You pick feature groups instead of single features: per-axis statistics (minimum, maximum, median, quartiles), the overall dynamic body acceleration (ODBA), fold-changes between axes, the mean-corrected norm, and a continuity measure.
2. Train the map. Firetail normalises the features and trains a self-organizing map, a small grid in which every node stands for a potential category.
3. Assign every burst. Each burst lands in the category it resembles most. The categories appear as annotation layers directly above the raw acceleration data.
4. Give them meaning. Look at each category in context, alongside the GPS track, the map, other sensors or video, then rename, merge or discard categories.

Start simple, then refine

Start with a 3×3 map, nine categories at most, and basic statistics: this will likely separate active from inactive patterns. Categories that still mix several behaviours can be refined: turn their regions into a selection and train a local model on just those bursts; select at least a few hundred of them. More categories need a larger map, and a larger map needs more training data.

Where you already know what the animal did, you can also annotate regions by hand: select them in the bursts or on the map and assign a category.

Reuse a model on other animals

A trained model can be saved, including your renamed and regrouped categories, and applied to other tags, individuals or deployments. Firetail can't detect biases between tags on its own, so transferred predictions should be checked against an independent reference.

Check the result against ground truth

Firetail has no built-in way to judge whether a model is right, so the check comes from outside: manual annotations, video observations synchronized with the acceleration data, or annotations from tools like BORIS. Firetail's evaluation module compares such a reference with the predicted categories and reports precision, recall and F-measure from the overlapping segments.