Landfill Pulse: A Coarse-to-Fine GeoFM Framework for Spatiotemporal Landfill Mapping

Existing global landfill inventories lack systematic updates to confirm ongoing site presence, leaving the temporal validity of individual records unknown and severely constraining longitudinal environmental and health research. Here we show that geospatial foundation model (GeoFM) embeddings offer a scalable route to closing this gap. We present Landfill Pulse, which harmonises four global inventories into 8,957 distinct landfill locations and applies a coarse-to-fine GeoFM framework for spatiotemporal landfill mapping. We trained a coarse classifier on 215-dimensional Clay embeddings at MAJOR TOM grid scale to screen all known landfill locations for annual changes. When a potential change signal emerges, a fine classifier—trained on 1024-dimensional embeddings tailored to landfill size—evaluates whether the shift holds consistent across all Sentinel-2 scenes in that year. The coarse classifier achieves an AUROC of 0.912, while the fine classifier achieves 0.947. Across the harmonised dataset, 47.6% of candidate sites survive fine-scale verification, with 11.6% showing confirmed temporal status changes. This work yields a globally harmonised landfill inventory equipped with explicit temporal annotations, providing critical baseline infrastructure for planetary monitoring.