Machine Learning for Classification of Satellite Geolocation Data: Addressing the Path Segmentation Challenge in Animal Movement Ecology

Knowledge of global animal movement offers insight into our changing planet, and direct observation of patterns from space is an ideal vantage point. Due to hardware miniaturization of animal trackers and satellites (CubeSats), increasing numbers of geolocation records are becoming available. Ecologists, biologists, and conservationists apply this data in their research and initiatives, but robust methods for automated classification of the data are lacking. In order to quantify behavioral changes at scale for the study and stewardship of nature, a system is needed that can automatically segment and label movement states. Such a system can benefit science by reducing the setup time for research, thereby improving resource allocation of people, time, and funding. This manuscript explores the viability of machine learning models to address the challenge of segmenting active migration from summer or winter range residency. Recurrent neural network (RNN) and long short-term memory (LSTM) architectures are both evaluated and compared. Results show encouraging accuracy with F1-scores exceeding 80%, and work is scoped for future optimizations and feature inclusion.