Evaluation factors information.

To address the prevailing scenario where comprehensive susceptibility assessments of ground deformation disasters primarily rely on knowledge-driven models, with weight judgments largely founded on expert subjective assessments, this study initially explores the feasibility of integrating data-driven models into the evaluation of urban ground collapse and subsidence. Hangzhou city, characterized by filled soil and silty sand, was selected as the representative study area. Nine pertinent evaluation factors were identified, and the RF-BP neural network coupling model was employed to assess the susceptibility of ground collapse and subsidence in the study area, the results indicate that the stacked model achieved a 7% increase in AUC value compared to the single model. Subsequently, this study utilized the advanced large language model (LLM), ChatGPT-4, to supplant expert judgment in the weight determination of ground deformation disasters. The advantages of ChatGPT-4, such as its ability to process vast amounts of data and provide consistent, unbiased judgments, were highlighted. ChatGPT-4’s assessments were validated by geological experts in the study area through the analytic hierarchy process. The results show that, by analyzing the same textual materials, the weights determined by experts differed by only 3% from those judged by ChatGPT, demonstrating the reliability and human-expert-like logic of ChatGPT-4’s judgments. Finally, a comprehensive susceptibility assessment of ground deformation disasters was conducted utilizing ChatGPT-4’s judgment results, yielding favorable outcomes.