Land subsidence simulation considering groundwater and compressible layers based on an improved machine learning method

Abstract

Land subsidence is a significant issue in the Beijing Plain, China, induced by groundwater overexploitation. The regional land subsidence is experiencing a new development trend with the external water source provided by the South-to-North Water Diversion Project (SWDP). The study proposes a novel model to simulate large-scale land subsidence that combines the weight of evidence (WOE) with the light gradient boosting machine (LightGBM) to explore the causes of land subsidence development after SWDP. The model encodes categorical variables to integrate information and evidence, reducing noise in the data, improving their interpretability, and enhancing robustness by transforming input features into more informative representations. The research findings show that SWDP has effectively mitigated subsidence development in the Beijing Plain from 2011 to 2018, reducing the subsidence area from 78 % to 58 % and the maximum rate from 135 mm/y to 110 mm/y. After SWDP, regional land subsidence is mainly attributed to the effects of groundwater and compressible clay layer and is related to engineering activities occurring on other construction land. Despite improved water use structures, water level changes in the second and third confined aquifers continue dominating the subsidence development. Unlike previous machine learning approaches, the proposed method can directly handle discrete data and is more adept at predicting severe subsidence changes. This study can be used to plan remediation strategies for regional land subsidence.