MODEL VALIDATION
A small blind test against observed flooding
Hexaflood compared terrain-based scores with 50 flooded and dry sample points from three European events. The pooled ranking AUC was 0.87, but this small sample is not evidence of universal accuracy.
What was tested
The model received coordinates only. Satellite-derived flood maps were used as ground truth after scoring, with 24 observed flooded points and 26 regional dry controls.
What the result supports
The sample supports using the score as a regional ranking signal for susceptibility screening. Absolute score levels differed by climate, and no single global threshold was perfect.
Important limitations
Three events across two climate settings are not exhaustive. Radar can under-detect water in dense urban areas, samples skew toward observable locations, and the test does not validate property-specific probabilities, depths or forecasts.