A two-stage cascaded prediction framework for urban inundation depth using a dual-branch CNN-ConvLSTM network

Abstract

Study region Shenzhen, China, is a densely urbanized subtropical megacity characterized by extensive impervious surfaces, rapid runoff responses, and frequent rainfall-induced inundation. Study focus To address the underestimation of urban inundation depth driven by severe zero-inflation, this study proposes a two-stage Cascaded Multi-scale CNN-ConvLSTM Network (CMC-Net). A Random Forest classifier first identifies hazardous inundation samples using a 0.15-m threshold. A dual-branch CNN-ConvLSTM then estimates water depth by integrating the local spatial context represented by static geographical features with the spatiotemporal evolution of multiscale cumulative rainfall. Structural ablations evaluate the contributions of multiscale convolution, channel attention, and recurrent spatial interactions. New hydrological insights for the region Evaluation on chronologically held-out positive samples shows that the conditional CNN-ConvLSTM is better than other benchmark regression models, achieving an NSE of 0.6150 and an RMSE of 0.2778 m. More importantly, under realistic full-sample scenarios (comprising both positive and negative samples), the cascade framework achieved an NSE of 0.4969, Pearson’ s r of 0.7996, RMSE of 0.3176 m, and MAE of 0.2523 m, outperforming the three direct-regression alternatives.The attribution and ablation results indicate that multiscale rainfall features provide useful information for inundation-depth estimation, while combining static spatial context with dynamic rainfall evolution improves estimation performance. Overall, the cascaded strategy reduces depth underestimation associated with severe class imbalance.