BEACON: Behavioral and Semantic Enrichment of AlphaEarth Embeddings through Tri-Modal Contrastive Learning
BEACON keeps AlphaEarth deployment image-only, but trains its embeddings against POI text and hourly visitation signals.
The paper frames the gap as AlphaEarth’s strength on physical Earth features but weaker encoding of human activity and urban function. In a Houston metro case study, BEACON is tested on nine downstream tasks against six baselines. With a linear probe, it reports relative R² gains over AlphaEarth of up to 43% for obesity prevalence, 34% for poor mental health, and 22% for median household income. The authors say the approach broadens geospatial foundation models toward human-centered urban analytics while staying competitive on physical and environmental variables. ArXiv · AI/CL/LG's note
The paper frames the gap as AlphaEarth’s strength on physical Earth features but weaker encoding of human activity and urban function. In a Houston metro case study, BEACON is tested on nine downstream tasks against six baselines. With a linear probe, it reports relative R² gains over AlphaEarth of up to 43% for obesity prevalence, 34% for poor mental health, and 22% for median household income. The authors say the approach broadens geospatial foundation models toward human-centered urban analytics while staying competitive on physical and environmental variables. ArXiv · AI/CL/LG's note
score 4