As the demand for regional geological disaster risk assessments in large cities continues to rise, our study selected Hangzhou, one of China’s megacities, as a model to evaluate the susceptibility to two major geological hazards in the region: ground collapse and ground subsidence. Given that susceptibility assessments for such disasters mainly rely on knowledge-driven models, and data-driven models have significant potential for application, we proposed a high-accuracy Random Forest—Back Propagation Neural Network Coupling Model. By using nine evaluation factors selected based on field surveys and expert recommendations, along with disaster data, the model's predictive results indicate a 3–40% improvement in model performance metrics such as AUC, accuracy, precision, recall, and F1-score, compared to single models and traditional SVM and logistic regression models. Ultimately, using the predictive results of this model, we created susceptibility maps for individual disasters and developed a muti-hazards susceptibility map by employing the expert weight discrimination method and the overlay evaluation method. Furthermore, we discussed the feature importance in the prediction process. Our study validated the feasibility of using advanced machine learning models for urban geological disaster assessment, providing a replicable template for other cities.