Causal Attribution of the Interannual Variability in Flood Peaks Through Bayesian Networks

Classical regression models, due to the limited computational expense and good performance, can be used for the attribution of interannual variability in flood peaks. However, these models capture the relation between predictand (i.e., flood peaks) and predictors (i.e., climate variables), suffering from the disconnect between correlation and causation. Here, we utilize a causal Bayesian Network model to establish causal relationships between flood peaks and basin- and season-averaged precipitation and temperature, which were found to be useful predictors in previous regression-based attribution studies. We develop these models for seasonal flood peaks for 3,884 gauges across the conterminous Unites States, achieving a median Spearman's rank correlation above 0.7. By performing do-calculus intervention on the predictors, we found a strong causal relationship between seasonal maximum daily discharge and both concurrent and lagged season-precipitation and temperature, consistent with underlying physical processes across different basins. The Bayesian Network model effectively predicts the interannual variability in seasonal and annual peak discharges and establishes a causal link between them. The model identifies key drivers across different seasons and regions in CONUS and highlights that antecedent catchment wetness is particularly relevant for high magnitude flows, while precipitation is the dominant driver of medium flows. This study significantly expands our current knowledge on causal flood drivers and presents a novel approach to flood prediction and attribution.