Analyzing animal movement using deep learning
Preprint
Abstract
Understanding how animals move through heterogeneous landscapes is central to ecology and conservation. In this context, step selection functions (SSFs) have emerged as the main statistical framework to analyze how biotic and abiotic predictors influence movement paths observed by radio tracking, GPS tags, or similar sensors. A traditional SSF consists of a generalized linear model (GLM) that infers the animal's habitat preferences (selection coefficients) by comparing each observed movement step to random steps. Such GLM-SSFs, however, cannot flexibly consider non-linear or interacting effects, unless those have been specified a priori. To address this problem, generalized additive models have been integrated in the SSF framework, but those GAM-SSFs are still limited in their ability to represent complex habitat preferences and inter-individual variability. Here we explore the utility of deep neural networks (DNNs) to overcome these limitations. We find that DNN-SSFs, coupled with explainable AI to extract selection coefficients, offer many advantages for analyzing movement data. In the case of linear effects, they effectively retrieve the same effect sizes and p-values as conventional GLMs. At the same time, however, they can automatically detect complex interaction effects, nonlinear responses, and inter-individual variability if those are present in the data. We conclude that DNN-SSFs are a promising extension of traditional SSF. Our analysis extends previous research on DNN-SSF by exploring differences and similarities of GLM, GAM and DNN-based SSF models in more depth, in particular regarding the validity of statistical indicators that are derived from the DNN. We also propose new DNN structures to capture inter-individual effects that can be viewed as a nonlinear random effect. All methods used in this paper are available via the 'citoMove' R package.
What the paper shows and why it matters
(AI-generated)
Step selection functions relate recorded animal movement paths to the environment. Each observed step (the move between consecutive locations) is compared with random alternative steps, usually in a conditional logistic regression (a generalised linear model, GLM), requiring the shape of every response and interaction to be specified in advance. Klappstein et al. (2024) relaxed this with smooth curves (generalised additive models, GAMs) and random effects, a reasonable step for nonlinear single-predictor responses and individual differences. Interactions, also between habitat and movement, still have to be specified. Neural networks can learn such interactions, but earlier deep-learning step selection models (Cífka et al. 2023; Forrest et al. 2026) concentrated on complex predictors and prediction. It remained open whether a network can still give the effect sizes, p-values and confidence intervals ecologists need, on the small datasets typical of animal tracking.
The paper shows that a simple neural network can replace the pre-specified functional form of a step selection model for animal movement and, read out with explainable-AI methods (techniques that show how output depends on input), still deliver effect sizes, p-values and confidence intervals with acceptable statistical properties in simulations. The network recovered linear effects with near-nominal false-positive rates and coverage, and nonlinear responses of a single predictor as accurately as a GAM. With nine predictors and 36 interactions, both the network and a GLM identified the three true ones. Averaged over all effects, the network had the lower error (0.08 against 0.14), because it shrank absent interactions towards zero while GLM estimates for them varied more; for the true interactions it underestimated the effects, with larger error (0.27 against 0.15). A further layer places individuals in a low-dimensional space where similar individuals lie close together, acting as a nonlinear random effect that can be read like an ordination biplot. It separated four simulated groups of individuals and three types of neighbouring animal (neutral, attracting, repelling). In eight wild boar, rural animals clustered apart from more variable urban ones.
The results suggest that such networks extend GLM- and GAM-based step selection rather than forming a separate black-box model class. Nonlinear responses, interactions and individual differences can be explored without specifying them beforehand, though the paper notes that flexibility can reduce statistical power. The paper implements the methods in the R package citoMove.
References
- Klappstein et al. (2024). Step selection functions with non‐linear and random effects. Methods in Ecology and Evolution 15, 1332–1346. doi:10.1111/2041-210x.14367
- Cífka et al. (2023). MoveFormer: a Transformer-based model for step-selection animal movement modelling. bioRxiv. doi:10.1101/2023.03.05.531080
- Forrest et al. (2026). Predicting animal movement with deepSSF : A deep learning step selection framework. Methods in Ecology and Evolution 17, 371–391. doi:10.1111/2041-210x.70136