Improving River Flood Mapping with Adaptive Sampling and Artificial Intelligence Techniques for Enhanced Flood Risk Assessment

There is a growing concern that existing flood frequency maps require re-assessment by the adoption of artificial intelligence (AI) techniques, but a persistent limitation of AI-based flood predictions is the selection of pseudo-absence samples due to lack of ground truth data and the risk of true negative bias. This study develops AI-assisted adaptive sampling techniques driven by precipitation and topographic factors with 17 flood driver forcings to forecast flood recurrence extents. The aim is to evaluate the applicability of such sampling techniques for enhancing existing flood design maps. Point-based and object-based flood events were sampled along existing delineated 1% annual exceedance probability (AEP) extent of the Miho River main channel, and equal non-flooded pseudo-absences were obtained based on multi-factor sensitivity analysis of the sampling methods. The elevation-adapted Random Forest (RFElev) relatively outperformed other models across flood and non-flood classes, achieving AUC = 0.98, Recall = 0.99, 48.52% true flood positive, 50.93% true non-flood negative, and 0.567% total error on test data. Simulated floodwater depths affected more agricultural, industrial, and residential asset clusters which are concentrated downstream at high depth zones of 3.94 m and 4.82 m for 200-yr and 500-yr respectively. Elevation-adapted pseudo-absence samples produced the strongest flood-no-flood feature pipeline, improving the delineation of transitional flood zones critical for infrastructure protection and emergency planning. The increased areal extent of predicted inundation extents offers a relatively better assessment for capturing unpriced asset vulnerability outside the old flood insurance zones, and help to update existing flood tax system, and facilitate disaster preparedness.