Evaluation of Land Subsidence Susceptibility in Kunming Basin Based on Remote Sensing Interpretation and Convolutional Neural Network

This study aims to utilize the Machine Learning (ML) model to produce highprecision maps of urban ground subsidence susceptibility, providing a scientific basis for disaster prevention and mitigation efforts in the Kunming Basin. In this patent study, remote sensing interpretation of Kunming City was conducted using SBAS-InSAR technology to acquire subsidence data. Based on the frequency ratio method, ten evaluative factors with strong correlations were selected to establish an evaluation index system for the subsidence susceptibility of the Kunming Basin. Five models, including CNN, Back Propagation Neural Network (BPNN), Genetic Algorithm optimized BPNN (GA-BPNN), Particle Swarm Optimization optimized BPNN (PSO-BPNN), and Radial Basis Function Neural Network (RBFNN), were employed. The frequency ratio method and the ROC curve were used to compare the effectiveness and precision of these models. The frequency ratio method indicated that the CNN model had the highest values in the very high and high susceptibility areas, reaching 4.10, which was the highest among all models; in the very low and low susceptibility areas, its value was 0.34, which was the lowest among the models. The ROC curve demonstrated that the CNN model, based on deep learning (AUC = 0.952), was more precise than the machine learning-based models such as BPNN (AUC = 0.896), RBFNN (AUC = 0.917), GA-BPNN (AUC = 0.890), and PSO-BPNN (AUC = 0.906). The CNN model has predicted that 81.06% of the ground subsidence grid cells fall into the very high and high susceptibility categories, demonstrating good predictive performance. According to the established evaluation index system for ground subsidence susceptibility, the fundamental causes of ground subsidence in the Kunming Basin are identified as poor soil mechanical properties and low bearing capacity, while construction activities have exacerbated the development of ground subsidence.