Enriching historical biologging datasets on seabirds using deep neural networks: A transformer‐based approach to infer energy expenditure proxy from <scp>GPS</scp> and environmental data

Abstract Recent advances in biologging have led to the widespread use of accelerometers, which generate high‐resolution movement data essential for understanding animal behaviour. Derived from tri‐axial accelerometry, Overall Dynamic Body Acceleration (ODBA) serves as a proxy for energy expenditure that is less invasive and more cost‐effective than alternative methods, contributing to a better understanding of individual survival and population dynamics. However, many long‐term datasets lack accelerometry due to limitations in instrumentation or older technologies, thereby constraining their utility in multi‐year ecological studies. Marine top predators, particularly seabirds, are valuable ecosystem sentinels due to their foraging strategies, energy constraints and sensitivity to environmental conditions. Understanding energy use during at‐sea travel is essential for forecasting colony viability under climate change, thus supporting long‐term ecological research and conservation efforts. High‐resolution biologging has facilitated data‐driven approaches to monitor seabird behaviour, but gaps in accelerometry data remain a challenge. New developments in deep neural networks (DNNs) offer opportunities to reconstruct missing biologging‐derived metrics by capturing the complex spatiotemporal relationships between movement and environmental data. One way to model such dependencies is to use a transformer architecture, which processes all inputs jointly and produces representations that explicitly encode relationships between them. In this study, we present ODBAFormer, a transformer‐based model designed to infer ODBA from GPS trajectories and environmental covariates. We applied ODBAFormer to seabird tracking data to address gaps in proxy data for energy expenditure. We evaluated the model on unseen data and compared it with state‐of‐the‐art machine‐learning regression method XGBoost as a baseline. We also explored its sensitivity to data heterogeneity and different environmental covariates combinations. Overall, ODBAFormer accurately reconstructs ODBA time series and proxies for both energy budgets and spatial energy expenditure distribution, providing a scalable solution to enrich legacy datasets or for situations where the use of multisensor loggers is limited by cost or weight.