From weather radars to bird migration fluxes Process-guided machine learning for spatio-temporal forecasting and inference

Every year, billions of birds migrate across the globe. Weather radar networks have become a powerful data source to monitor these migratory movements at large scales, but several technical challenges — such as sparse coverage, partial velocity information, and measurement noise — remain. Moreover, incomplete mechanistic understanding of migration behaviors restricts the application of traditional modeling approaches. This thesis explores how partial process knowledge can be combined with machine learning to improve the analysis and prediction of bird migration from weather radar data. Part I focuses on continental-scale forecasting. By integrating physical principles from fluid dynamics—such as mass conservation—into deep learning models, we develop hybrid approaches that enhance both the accuracy and interpretability of migration forecasts. Part II addresses the challenge of reconstructing local-scale movement from raw radar signals. We propose probabilistic frameworks that fuse the flexibility of deep learning with graphical models to encode known dependency structures. These methods provide a foundation for reconstructing high-resolution movement patterns with uncertainty quantification and computational efficiency. Together, these contributions highlight the potential of process-guided machine learning in tackling complex ecological systems.