Deep Neural Networks Show Promise for Analyzing Complex Animal Movement Patterns
Preprint demonstrates that DNN-SSFs can match traditional models for linear effects while automatically detecting complex, nonlinear, and individual-specific patterns in animal movement data, offering a flexible new tool for ecologists.
A preprint study proposes using deep neural networks to improve the analysis of how animals navigate heterogeneous landscapes, a key topic in ecology and conservation. Step selection functions (SSFs) are the primary statistical tool for interpreting movement data collected via GPS tags or radio tracking. Conventional SSFs rely on generalized linear models (GLMs) that compare observed movement steps to random alternatives to infer habitat preferences. However, these models struggle with non-linear effects or interactions unless researchers specify them in advance. Generalized additive models (GAMs) offer some improvement but still have constraints in modeling intricate patterns or variation between individuals. The authors tested deep neural network-based SSFs (DNN-SSFs), integrated with explainable AI methods to derive selection coefficients. According to the preprint, DNN-SSFs reproduce the same effect sizes and p-values as GLMs when effects are linear, while also automatically identifying non-linear responses, interaction effects, and inter-individual variability when present in data. This methodological paper compares GLM, GAM, and DNN approaches in depth, including evaluation of statistical indicators extracted from the neural networks, and introduces new DNN structures to model inter-individual effects as nonlinear random effects. No specific animal species, datasets, or sample sizes are detailed in the abstract, as the work focuses on model development and validation rather than a particular empirical case study. As an arXiv preprint (not yet peer-reviewed), the findings require further scrutiny. Limitations include uncertainties around the validity of p-values and other statistical measures derived from DNNs. All code and methods are provided in the 'citoMove' R package. Source URL: https://arxiv.org/abs/2603.24009
HELIX: This means AI could soon help scientists uncover hidden patterns in how animals move and survive, leading to smarter conservation efforts that protect wildlife and ecosystems we all depend on. Ordinary people might see it in better environmental policies or nature documentaries that feel a lot more insightful.
Sources (1)
- [1]Analyzing animal movement using deep learning(https://arxiv.org/abs/2603.24009)