Mapping a novel metric for Flash Flood Recovery using Interpretable Machine Learning

Abstract

Flash floods are one of the most devastating natural disasters, yet many aspects of their severity and impact are poorly understood. The recession limb is related to post-flood recovery and its impact on communities, yet it remains less documented than the rising limb of the hydrograph to predict the peak discharge and timing of floods. . This work introduces a new metric called the flash flood recovery or recoveriness, which is the potential for recovery of a watershed to pre-flood conditions. Using a comprehensive database of 78 years and supervised machine learning algorithms, flash flood recovery is mapped in the conterminous United States. A suite of geomorphological and climatological variables is used as predictors to provide probabilistic estimates of recoveriness. Slope index, river basin area and river length are found to be the most significant predictors to predict recoveriness. Several new localized hotspots were identified, such as the western slopes of the Appalachians consisting of Kentucky, Tennessee, and West Virginia and the interlinked areas of western Montana and northern Idaho. This new metric can be useful for prioritizing relief and rehabilitation efforts as well as precautionary measures for disaster risk reduction.