Assessing Global Hand Datasets as Priors for SAR-Based Bayesian Flood Mapping

Abstract

Floods continue to affect millions of the global population annually. SAR-based methods are one of the most reliable tools for mapping floods expeditiously. Among these are Bayesian flood mapping methods that rely on conditional and prior probability formulations to make labeling decisions. Recent work demonstrated a globally applicable Height Above the Nearest Drainage (HAND)-based prior probability function to improve Bayesian flood mapping. However, limitations were identified due to the input DEM. In this contribution, we assess the performance of three (near-)globally available HAND datasets as input to this function. Compared to the HAND dataset used in the previous study, the MERIT and Deltares HAND datasets were derived from improved SRTM DEMs and finer detailed drainage networks. We hypothesize that these finer-resolution HAND datasets can potentially improve probabilistic flood mapping further. Thus, we compare the flood mapping performance using the baseline SRTM-derived, MERIT, and Deltares HAND data as priors on (the original) six study sites for both flooded and non-flooded scenarios. Our results show similar performance in the flooded scenarios using the MERIT and Deltares HAND datasets. The MERIT dataset shows slightly better performance among the three. However, an increase in False Positive Rates was apparent in non-flooded scenarios attributed to smaller drainages in the new datasets tested. These results suggest caution in applying the HAND prior method with HAND datasets derived from drainage networks with small upstream contributing areas.