Land subsidence prediction model based on its influencing factors and machine learning methods

Abstract

Abstract Land subsidence has caused huge economic losses in the Beijing plains (BP) since 1980s. Building land subsidence prediction models that can predict the development of land subsidence is of great significance for improving the safety of cities and reducing economic losses in Eastern Beijing plains. The pattern of evolution of land subsidence is affected by many factors including groundwater level in different aquifers, thicknesses of compressible layers, and static and dynamic loads caused by urban construction. First, we used the small baseline subset Interferometric Synthetic Aperture Radar (SBAS-InSAR) technology on 47 ENVISAT ASAR images and 48 RADARSAT‐2 images and used Persistent Scatterers Interferometric Aperture Radar (PS-InSAR) technology on 27 Sentinel-1 images to obtain the land subsidence monitoring results from June 2003 to September 2018. Second, the accuracy of the InSAR monitoring results were validated by using leveling benchmark land subsidence monitoring results. Finally, we built land subsidence rate prediction models and land subsidence gradient prediction models by combining land subsidence influencing factors and four machine learning methods including support vector machine (SVM), Gradient Boosting Decision Tree (GBDT), Random forest (RF) and Extremely Randomized Trees (ERT). The findings show: (1) The InSAR monitoring results revealed that the maximum land subsidence rate reached − 110.7 mm/year, -144.4 mm/year and − 136.8 mm/year during the 2003–2010, 2011–2015 and 2016–2018 periods, respectively. (2): The InSAR monitoring results agreed well with the leveling benchmark monitoring results with the Pearson correlation coefficients of two monitoring results were 0.97, 0.96 and 0.95 during the 2003–2010, 2011–2015 and 2016–2018 periods, respectively. (3): We found that the land subsidence prediction based on ERT method is the optimal model among four land subsidence prediction models and that the prediction performance of land subsidence prediction model based on ERT method will be greatly improved when apply this prediction model in sub study areas where the land subsidence mechanism is similar owning to the similar hydrogeological parameters.