Formation mechanism analysis and the prediction for compound flood arising from rainstorm and tide using explainable artificial intelligence

• Published in: Journal of environmental management Vol. 388; p. 125858
• Main Authors: Lai, Chenguang, Liao, Yanghao, Yu, Haijun, Wang, Zhaoli, Liao, Yaoxing, Yang, Bing, Niu, Qiang, Jiang, Zezhou, Li, Xuefang, Xu, Chong-Yu
• Format: Journal Article
• Language: English
• Published: England Elsevier Ltd 01.07.2025
• Subjects: Artificial Intelligence; Compound flooding; Deep learning; Explainable artificial intelligence; Floods; Multi-head attention; Rain; Deep learning; Explainable artificial intelligence; Compound flooding; Multi-head attention
• ISSN: 0301-4797; 1095-8630; 1095-8630
• DOI: 10.1016/j.jenvman.2025.125858

The compounded effect of heavy rainfall and high tide backwater significantly exacerbate the load on urban drainage systems in coastal cities, leading to an escalating risk of compound flood disasters. The formation mechanism of compound floods is of great complexity, and the research concerning it constitutes a highly challenging subject. While deep learning (DL) techniques have been increasingly applied in flood forecasting, their "black-box" nature often obscures the internal decision-making logic, limiting insights into the mechanisms driving compound flooding. To address this, our study proposes an explainable artificial intelligence (XAI) framework, utilizing a Long Short-Term Memory (LSTM) network integrated with a Multi-Head Attention (MHA) mechanism as a surrogate model for urban flood simulation. The SHapley Additive exPlanations (SHAP) method is employed to elucidate the model's decision-making process, uncovering critical driving factors and their interactions in compound flooding scenarios. Results demonstrate that the MHA mechanism enhances the model's ability to capture rainfall-tide interactions, with the LSTM-MHA model outperforming data-driven baseline models and achieving performance slightly below physics-based models, as evidenced by an R2 of 0.971, MAE of 0.040 m, and RMSE of 0.065 m. Furthermore, the LSTM-MHA model significantly improves computational efficiency, completing simulations 216 times faster than traditional physics-based models in the study case. SHAP analysis reveals consistent trends across typical scenarios, highlighting the dominant roles of rainfall and tidal factors across spatiotemporal scales and validating the surrogate model's decision-making rationality. By integrating XAI with SHAP, this study enhances both the accuracy and transparency of flood simulations, quantifying the relative contributions and interaction mechanisms of compound variable, and offering new perspectives for analyzing the underlying causes of compound flooding. This approach holds significant potential for developing more robust disaster mitigation systems and strengthening the resilience of coastal cities.