Unveiling flood-generating mechanisms using circular statistics-based machine learning approach without the need for discharge data during inference

Abstract

Understanding the drivers of flooding is essential for flood disaster prevention. However, conventional flood prediction methods are hindered by their reliance on local discharge data, which can be constrained by limited spatial resolution. To address this limitation, we present a machine learning model that can categorize floods without requiring discharge data during inference. We first use circular statistics to calculate the relative importance of three candidate flood-generating mechanisms. Global land areas are classified into three primary categories and eight sub-categories based on the proportion of relative importance. A random forest model is then applied to identify the flood types by assuming that the discharge data is unavailable. The findings from circular statistics highlight that globally, soil moisture excess is the most influential driver of floods followed by extreme precipitation and snowmelt, with an average relative importance of 0.535, 0.387, and 0.078, respectively. The RF model performs well in resembling the three primary flood categories with an accuracy of 0.701 and a F1-score of 0.692 in 10-fold cross-validation. The trained gridded-based model provides a swift and efficient approach for analyzing flood mechanisms, even in limited discharge scenarios, allowing for rapid insights.