Enhancement of land subsidence prediction capabilities using machine learning and SHAP value analysis with Sentinel-1 InSAR Data

Abstract

Land subsidence has been a significant focus of geoscience studies, and researching the factors contributing to it and predicting future incidents is essential. However, current research requires a systematic and coherent strategy to identify symptoms of land sinking with scientific rigor. This study employs neural networks and SHAP values to forecast land subsidence. We utilized SHAP values as a substitute for the usual random forest (RF) technique to evaluate the attributes of land subsidence. In addition, we employed neural networks to predict the prospective areas where land subsidence may occur in Chongqing and Chengdu in the future, considering a diverse set of conceivable situations. The results indicate that using neural networks for land subsidence prediction improves the model's accuracy by 16% compared to the traditional approach. The performance is enhanced by approximately 22% after optimizing the input characteristics. The feature optimization strategy proposed in this study, which relies on SHAP values, is particularly advantageous for predicting land subsidence. Moreover, the determinants of land sinking, ascertained by data analysis considering intricate topography, correspond with the conclusions of previous research. Utilizing this feature optimization method can improve the selection of input variables for the land subsidence prediction model, increasing accuracy and providing a solid theoretical foundation for preventing urban land subsidence.