Learning Spatial Ecological Networks from Geolocated Molecular and Earth Observation Data

Ecological populations exchange genes through dispersal, forming spatial ecological networks whose edge weights measure the effective ecological distance separating them. Inverse landscape genetics (ILG) seeks to recover this network by learning an associated conductance grid graph whose edges encode local landscape permeability to movement, so that the effective resistance it induces between sampled populations reproduces their genetic dissimilarity. Existing approaches learn these conductances through low-capacity encoders informed by coarse-grained, feature-engineered covariates, limiting their ability to capture the effect of fine-grained landscape features on ecological connectivity. We introduce a differentiable ILG pipeline that allows training a deep encoder end-to-end through a GPU-accelerated effective-resistance layer. On an empirical dataset, we show that training a deep patch encoder with the pipeline improves predictive performance over \texttt{ResistanceGA}, the classical ILG framework used by practitioners, and a fully data-driven metric-learning approach. The pipeline offers a scalable route to extract rich information from high-resolution Earth observation data to infer spatial ecological networks, with applications to protected-area network design, invasive-species mitigation, and biodiversity-positive renewable-energy planning.