Optimization of land subsidence prediction features based on machine learning and SHAP value with Sentinel-1 InSAR Data

Abstract

Abstract Land subsidence has always been a concern of geoscience, and exploring the factors affecting land subsidence to predict future land subsidence is essential research. However, current research rarely has a scientific and unified feature screening process for land subsidence features. This study applies neural networks and SHAP values to land subsidence prediction. We used SHAP values instead of the traditional random forest (RF) to quantify land subsidence features and neural networks to predict the areas where land subsidence is likely to occur in the cities of Chongqing and Chengdu, encompassing the majority of the possible land subsidence scenarios in the future. The results show that the prediction of land subsidence using neural networks improves the model accuracy by 16% compared to the traditional method. After input features optimization, the performance improves by nearly 22%. We found that the feature optimization method based on SHAP values proposed in this study is more helpful for land subsidence prediction, and the factors affecting land subsidence derived from data analysis with complex terrain are also consistent with the results of previous studies. This feature optimization method can contribute to the input variable selection process for the land subsidence prediction model, improve accuracy, and provide solid theoretical support for preventing urban land subsidence.