Quantifying the contribution of multiple factors to land subsidence in the Beijing Plain, China with machine learning technology

Land subsidence is the ground surface response to underground space development, utilization and evolution. Presently, land subsidence has developed into a global, comprehensive and interdisciplinary complex systems problem. More than half a century has passed since the discovery of subsidence in the Beijing Plain in the 1960s. In this study, we investigate the land subsidence in the Beijing Plain over the period of 2003–2015 using ENVISAT ASAR and RADARSAT-2 interferometric datasets and the small baseline subset interferometric synthetic aperture radar (SBAS-InSAR) technique. Furthermore, we introduced the data field model and index-based built-up index (IBI) to obtain the dynamic and static load information of the Beijing Plain. Then, based on a machine learning method, we selected the gradient lifting decision tree (GBDT) model to quantitatively analyze the contributions of groundwater level change, compressible deposit thickness and dynamic and static loads to land subsidence. The results showed that the maximum land subsidence rate was 122 and 141 mm/year in 2003–2010 and 2010–2015, respectively. Comparisons between the SBAS-InSAR results and leveling measurements showed that the minimum absolute error achieved was only 0.2 mm/year. We suggest that the groundwater exploitation in the third confined aquifer has greater impacts on land subsidence in the Beijing Plain than the other factors. The land subsidence likely occurred in compressible deposit thicknesses exceeding 90 m. Moreover, we found that the compressible thickness and groundwater level contributions to land subsidence exceeded 60%. Our results provide a scientific basis for the regulation and control of regional land subsidence.