High flood-model accuracy can hide how a map will fail in the next flood
WHAT THE EVIDENCE SAYS
A review of 200 peer-reviewed flood-susceptibility modeling studies found that roughly 73% used validation that did not fully account for spatial dependence, likely exaggerating generalization. About 48% did not quantify uncertainty, and only around 6% used a probability-calibration assessment.
WHY IT MATTERS / OUR READ
The counter-signal is that an impressive AUC on a map is not a flood prediction people can safely use. Model evaluation should hold out spatially independent areas, expose residual uncertainty, and test the actual planning decision that a map is intended to support.
CAVEATS / UNCERTAINTY
The counts describe practices in a heterogeneous sample of published studies, not a meta-analytic correction for every model. The estimated validation inflation varies by data and design; the review does not establish accuracy for a specific floodplain.
SOURCE
Frontiers in Environmental Science · AI flood-mapping methodological review
View original evidence ↗
A 200-study review found widespread spatially naive validation and thin uncertainty reporting in AI susceptibility maps.
Read the Wild Lands assessment →