Machine learning-constrained projection of bivariate hydrological drought magnitudes and socioeconomic risks

Abstract. Climate change accelerates the water cycle and alters the spatiotemporal distribution of hydrological variables, thus complicating the projection of future streamflow and hydrological droughts. Although machine learning is increasingly employed for hydrological simulations, few studies have used it to project hydrological droughts, not to mention the bivariate risks of drought duration and severity as well as their socioeconomic effects under climate change. We develop a cascade modeling chain to project future bivariate hydrological drought characteristics in 179 catchments over China, using 5 bias-corrected GCM outputs under three shared socioeconomic pathways, five hydrological models and a deep learning model. We quantify the contribution of various meteorological variables to daily streamflow by using a random forest model, then employ terrestrial water storage anomalies and a standardized runoff index to evaluate recent changes in hydrologic drought. Subsequently, we construct a bivariate framework to jointly model drought duration and severity by using Copula functions and the most likely realization method. Finally, we use this framework to project future risks of hydrological droughts as well as associated exposure of gross domestic product and population. Results show that our hybrid hydrological-deep learning model achieves >0.8 Kling-Gupta efficiency in 161 out of 179 catchments. By the late 21^st century, bivariate drought risk is projected to double over 60 % catchments, mainly located in Southwest China. Our hybrid model also projects substantial GDP and population exposures by increasing bivariate drought risks, suggesting an urgent need to design climate mitigation strategies towards a sustainable development pathway.