Machine-learning blends of geomorphic descriptors: value and limitations for flood hazard assessment across large floodplains

Recent literature shows several examples of simplified approaches that perform flood hazard (FH) assessment and mapping across large geographical areas on the basis of fast-computing geomorphic descriptors. These approaches may consider a single index (univariate) or use a set of indices simultaneously (multivariate). What is the potential and accuracy of multivariate approaches relative to univariate ones? Can we effectively use these methods for extrapolation purposes, i.e., FH assessment outside the region used for setting up the model? Our study addresses these open problems by considering two separate issues: (1) mapping flood-prone areas and (2) predicting the expected water depth for a given inundation scenario. We blend seven geomorphic descriptors through decision tree models trained on target FH maps, referring to a large study area (∼ 105 km2). We discuss the potential of multivariate approaches relative to the performance of a selected univariate model and on the basis of multiple extrapolation experiments, where models are tested outside their training region. Our results show that multivariate approaches may (a) significantly enhance flood-prone area delineation (accuracy: 92%) relative to univariate ones (accuracy: 84%), (b) provide accurate predictions of expected inundation depths (determination coefficient ∼0.7), and (c) produce encouraging results in extrapolation.