BEACON Enhances Geospatial AI with Behavioral and Semantic Data

A novel framework named BEACON has been introduced to augment geospatial foundation models, such as AlphaEarth, which primarily rely on Earth-observation imagery. While these models excel at capturing physical characteristics, they often weakly encode human activity and urban function. BEACON addresses this by employing a tri-modal contrastive learning approach. This framework aligns three distinct views of urban environments: physical representations from existing geospatial embeddings, semantic information derived from point-of-interest (POI) text data, and human behavioral patterns inferred from hourly POI visitation. The goal is to enrich the deployed image-only representation with these additional signals. Evaluated in the Houston Metropolitan Area, BEACON demonstrated substantial improvements across various downstream tasks. It boosted relative R^2 scores for predicting obesity prevalence by up to 43%, poor mental health by 34%, and median household income by 22% compared to AlphaEarth, while maintaining strong performance on physical and environmental variables. These results highlight the significant value of integrating semantic and behavioral data to extend geospatial models beyond physical observation into human-centered urban analytics.