Land cover and flood type govern the detection limits of satellite-based flood mapping across diverse global flood events

In three linesPrithvi-EO-2.0, a geospatial foundation model, tested on 19 flood events (2017-2025) across 6 continents. Accuracy varies by land cover: cropland 52% IoU, tree cover 4%. Riverine detection strong (F1=0.69). 23 failure modes identified; pipeline engineering dominates initial errors over model capacity.

Summary generated by Claude — human-verified