Multi-source Image Feature Extraction and Segmentation Techniques for Karst Collapse Monitoring

Linwen Bo^*

• Gansu Industrial Vocational and Technical College, Tianshui, China

Karst collapse monitoring is challenging due to data sparsity, complex underground dynamics, and the need for real-time adaptation. Traditional methods often struggle to provide accurate and timely risk assessments. To address these issues, we propose the Integrated Karst Collapse Prediction Network (IKCPNet), a framework that integrates multi-source imaging, geophysical modeling, and machine learning techniques. IKCPNet processes seismic imaging, hydrological patterns, and environmental factors to predict collapse risks with high precision.The model leverages an advanced data encoding mechanism to fuse diverse inputs and a physics-informed module to capture subsurface changes. Additionally, a dynamic risk assessment strategy enhances decision-making by incorporating real-time feedback and probabilistic mapping.Experimental results show that IKCPNet significantly improves prediction accuracy and risk mitigation. On the OpenSARShip dataset, it achieves an accuracy of 94.34±0.02 and an IoU of 90.23±0.02, outperforming the previous best model by 1.22 and 0.89 points, respectively. These findings highlight the potential of our approach for enhancing geological hazard monitoring through multi-source data integration.