A spatial explainable deep learning framework for prediction classification of hydrological drought in ungauged basin

Abstract

Study Region

The Upper Yellow River Basin (UYRB), China

Study focus

The increasing frequency, spatial extent, and intensity of hydrological droughts pose devastating impacts on water security, ecosystem stability, and sustainable development. While deep learning models have demonstrated significant promise in drought forecasting, particularly in data-scarce basins, their inherent opacity hinders the understanding of drought mechanisms. Therefore, we proposed a Spatial Explainable Deep Learning (SEDL) framework suitable for ungauged basins that can be integrated with various deep learning models. This framework aims to quantitatively analyze the spatial driving mechanisms that govern hydrological drought occurrence and enhances the accuracy of categorical drought prediction.

New hydrological insights for the region

This study quantitatively demonstrated that natural runoff (with the mean contribution of 42.45 %) was the primary driving factor of hydrological droughts, surpassing the effects of precipitation (26.90 %) and average temperature (31.05 %). Crucially, regional hydrological drought occurrence was spatially influenced by temperature-runoff interactions in upstream headwater catchments and local precipitation variations. The improved model based on the SEDL framework achieved significant improvements in performance, with maximum increments of 8.7 % in accuracy, 62.7 % in Kappa coefficient (K), and 10.3 % in F₁-score. By integrating deep learning with explainable artificial intelligence, the SEDL framework revealed the spatial physical driving factors of hydrological droughts while achieving 85.2 % accuracy on the test set, thus establishing a new research paradigm for explainable drought prediction.