Optimization of land subsidence prediction features using machine learning and SHAP analysis with Sentinel-1 InSAR data in Chengdu and Chongqing, China

Land subsidence constitutes a major geological hazard, especially in swiftly urbanizing areas. This study examines the principal cities of southwestern China (Chengdu and Chongqing), where comprehending the processes contributing to subsidence and forecasting future dangers is crucial. Nevertheless, contemporary research seldom highlights a systematic feature screening procedure. This study utilized SHAP values in place of the conventional random forest method to measure land subsidence characteristics and employed a neural network to forecast regions at risk of future susceptibility. The findings indicate that the BP neural network model (RMSE = 0.468, MAE = 0.219) surpassed the RF model (RMSE = 0.557, MAE = 0.310). Following feature improvement, the model's performance was enhanced, resulting in a more precise prediction, with the obtained factors aligning with prior research. This feature optimization technique can aid in the selection of input variables and offer a robust theoretical framework for addressing and alleviating urban land subsidence.