SHAP-enhanced interpretive MGTWR-CNN-BILSTM-AM framework for predicting surface subsidence: a case study of Shanghai municipality.

Urban expansion and subsurface resource exploitation have intensified ground subsidence, posing significant geological risks. Conventional prediction models often overlook multi-scale spatiotemporal effects that critically influence accuracy. This study proposes an integrated MGTWR-CNN-BiLSTM-AM (MGCBA) model to address this gap. Utilizing SBAS-InSAR-derived deformation data from Shanghai's primary subsidence zones, validated through GNSS and PS-InSAR observations, we developed a Multi-scale Geographically and Temporally Weighted Regression (MGTWR) framework. This model quantifies nonlinear spatiotemporal relationships between subsidence and driving factors, including monthly-scale variables (groundwater extraction, precipitation) and annual-scale parameters (land use, soil type), generating dynamic weight matrices. The integrated CNN-BiLSTM-AM (CBA) deep learning network extracts critical time-series features to optimize spatiotemporal weights adaptively. Experimental results demonstrate a prediction accuracy of 0.99347 (RMSE: 1.8643 mm), outperforming the standalone CBA model (0.98494) by 0.85%. SHAP value analysis identifies monthly groundwater levels, soil moisture, and annual-scale soil type/DEM as dominant contributors to Shanghai's urban core subsidence. The proposed multi-scale spatiotemporal modeling framework advances surface deformation prediction by enhancing the interpretability of key drivers under spatiotemporally variable conditions.