<p>Rainfall-induced karst collapse poses serious threats to infrastructure and human safety, particularly in regions with complex underground geological structures. Traditional statistical or shallow learning methods struggle to effectively capture the intricate spatiotemporal dependencies that govern collapse events. To address this challenge, we propose a novel predictive framework—ESTG-KC (Enhanced Spatiotemporal Graph Neural Network for Karst Collapse prediction). This framework integrates multi-source geospatial and environmental data into a dynamic graph structure. Unlike existing approaches, ESTG-KC introduces a rainfall-driven dynamic graph convolution mechanism that adaptively updates node connectivity based on real-time precipitation patterns by computing the pairwise rainfall differences and applying an exponential decay function to determine edge weights. This allows the graph topology to evolve in response to changing rainfall conditions, effectively reflecting the dynamic connectivity of geological structures. Additionally, the framework incorporates a multi-scale temporal attention mechanism that captures both abrupt short-term rainfall events and gradual long-term accumulation effects. To verify the model’s performance, extensive experiments were carried out on several publicly accessible environmental datasets. The results indicate that ESTG-KC consistently surpasses leading existing methods in terms of prediction accuracy and early warning effectiveness. This work provides a reliable and interpretable tool for forecasting karst collapses triggered by rainfall, supporting efforts in geological hazard prevention and risk management.</p>