Sai Kiran Kuntla , Manabendra Saharia, Pierre Kirstetter

The increasing risk of floods across the globe needs focused attention because of the extensive damage to human lives and economy. A comprehensive understanding of its causative factors is of vital importance. Yet catchment characterization studies are generally limited to case studies or regional domains. A comprehensive global characterization is currently unavailable, which requires collecting and collating a large number of datasets over vast areas. This study embraces large-sample data-driven science as a new paradigm to characterize streamflow extremes by utilizing global datasets of physiographic explanatory variables that could explain various facets of extreme streamflows. Along with the spatial and temporal variations of high streamflow extremes, their correlation with various catchment characteristics such as geomorphology, meteorology, climatology, landcover, lithology, etc. were examined. The multidimensional relationships between the streamflow extremes and catchment characteristics were modeled using a Random Forest approach and combined with an interpretable machine learning framework to identify the most dominant factors in varying climate classes. Interpretation with SHAP (SHapley Additive exPlanations) reveals that meteorological variables are the most influential variables across the climatic classes. However, the variables and their influences change among different climatic classes. Moreover, different geomorphological variables come into dominance across climatic classes (such as basin relief in warm temperate and drainage texture in arid climates). Overall, the insights from the study could play a crucial role in predicting the unit peak discharge at ungauged stations from the known catchment characteristics. Moreover, these findings can also play a crucial role in formulating risk management strategy.