A cascaded deep learning framework for improved forecasting of drought-flood abrupt alternation events in the Hanjiang River Basin

Abstract

Study region The Hanjiang River Basin (HJRB), a meteorologically and topographically complex region in China. Study focus Drought-Flood Abrupt Alternation (DFAA) events threaten water security, yet their prediction remains challenging because of highly nonlinear dynamics and pronounced zero inflation in meteorological records. This study employed the Standardized Drought-Wetness Abrupt Alternation Index (SDWAI) to identify DFAA events and used SHAP to reveal their driving mechanisms. A cascaded framework integrating a parallel convolutional neural network-long short-term memory (CNN-LSTM) classifier, wavelet decomposition, and ensemble regression was developed. By separating event identification from conditional intensity estimation, the framework reduced the influence of non-event samples. New hydrological insights for the region The CNN-LSTM classifier achieved a balanced accuracy of 0.67 and module obtained the best performance (R² = 0.91). The complete cascade framework achieved an overall R² of 0.58 and reduced the peak underestimation commonly observed in conventional zero-inflated regression models. Spatiotemporal analysis revealed distinct phase-locking characteristics of DFAA events, with downstream and southeastern HJRB identified as hazard hotspots. Drought to flood transitions occurred mainly during spring and summer, whereas flood-to-drought transitions increased markedly during late autumn and early winter. Local precipitation and potential evapotranspiration dominated DFAA evolution, while the Arctic Oscillation primarily regulated the large-scale atmospheric circulation background. These findings demonstrate the potential of cascaded deep learning frameworks for regional DFAA forecasting and adaptive water resource management.