Explaining Geographic AI Embeddings for Geospatial Data

Geographic implicit neural representations (INRs) are increasingly used to map Earth coordinates to location embeddings, effectively encoding geospatial data into neural network weights. While these embeddings serve as general-purpose geospatial representations, there has been a lack of principled tools to understand precisely what geographic or semantic information they capture. This research addresses that gap by providing an explainability analysis. The study proposes a method to decompose these complex location embeddings into three human-interpretable feature types: sparse latent concepts, natural language concepts, and visual features. Sparse autoencoders are used to learn the latent concepts, while Sparse Linear Concept Embeddings (SpLiCE) are applied with a predefined geospatial dictionary to recover natural language concepts. Visual features are extracted using saliency maps derived from CLIP Surgery. The findings demonstrate that location embeddings can be effectively broken down into these interpretable representations while maintaining high reconstruction capability. This decomposition reveals clear geographic structures such as forests, deserts, urban areas, roads, and landmarks, as well as broader biome and climate signals. This work represents a significant step towards creating more transparent and auditable geospatial AI representations.