NASA's Prithvi-EO-2.0 reveals flood mapping blind spots across 19 global events

A team led by Venkatesh Kolluru from NASA's IMPACT group has rigorously benchmarked the geospatial foundation model Prithvi-EO-2.0 for satellite-based flood mapping across 19 diverse flood events spanning 2017–2025. The study covered six continents, eight climate zones, and six distinct flood mechanisms, using two independent reference products for validation. The results are sobering: while cropland inundation detection reached 52% IoU and riverine floods achieved F1=0.69, tree-covered and built-up areas saw near-zero detection at just 4% IoU regardless of flood type. This suggests existing AI models systematically fail in urban and forested environments where flood impact is often most severe.

Importantly, the researchers found that apparent model errors were partly due to definitional inconsistency between the two reference flood maps—a critical insight for operational users. By iterating through their pipeline, they identified 23 distinct failure modes, with pipeline engineering (e.g., preprocessing, post-processing) contributing more to initial error than the model's capacity itself. These findings establish clear environment-dependent detection boundaries for satellite flood mapping, guiding where Prithvi-EO-2.0 can be trusted and where further refinement is needed. The work underscores that deploying AI for disaster response requires not just better models but also better understanding of data and pipeline limitations.