With the continuous intensification of underground space development and the increasing frequency of extreme weather events in China’s megacities, ground collapse incidents have surged nationwide in recent years, posing severe threats to public safety and socio-economic stability. The precise quantification of nonlinear interactions among complex hazard factors and the effective coupling of static risks with dynamic triggers is crucial for the accurate prevention and control of ground collapse. In this study, a machine learning-based framework for ground collapse risk assessment was developed using Shenzhen as the case study area. A comprehensive evaluation index system encompassing ten hazard factors was established through the integration of 1 036 ground collapse incident records in Shenzhen from 2019 to 2022, high-precision geoenvironmental data, rainfall data, and spatial data on urban infrastructure distribution. Three machine learning algorithms—Random Forest, Support Vector Machine, and Logistic Regression—were employed for comparative susceptibility assessment, and Bayesian optimization was adopted for hyperparameter tuning. It is demonstrated that the Random Forest algorithm effectively captures complex nonlinear relationships among hazard factors in multi-source data fusion scenarios, and the best classification and prediction performance is achieved (AUC=0.9326). On this basis, a power-function threshold model based on Early Effective Rainfall and Rainfall Duration (EE-D) was proposed for the Shenzhen area, enabling accurate quantitative expression of rainfall-triggering thresholds. Subsequently, the EE-D model was deeply integrated with static spatial probability predictions derived from machine learning, and thereby a complete technical framework for dynamic risk assessment characterized by “data fusion-intelligent evaluation-dynamic early warning” was established. It is shown by back-validation that 99.6% of rainfall-induced collapse events occurred within moderate-to-high risk zones (AUC≥0.78), indicating high prediction accuracy and early-warning efficacy. A robust scientific basis for the precise prevention and control of ground collapse in megacities is provided by this framework.