Abstract: Regional subsidence has received considerable attention worldwide. Land subsidence encompasses both continuous and seasonal subsidence, i.e., nonrecoverable and recoverable subsidence. Recoverable and nonrecoverable subsidence are identified based on the relationship between subsidence and groundwater-level changes. However, the soil material doesn't follow a linear elastic relation between subsidence and groundwater change. This paper aims to investigate recoverable and nonrecoverable subsidence based on a machine learning approach. Specifically, using the Support Vector Machine (SVM) model allows for the identification of recoverable and nonrecoverable subsidence. Using the Gaussian RBF (radial basis function) Kernel as the SVM classifier, the boundary of a nonrecoverable region can also be defined. The model effectively quantifies recoverable and nonrecoverable subsidence using information on subsidence and groundwater change, e.g., the ratio of subsidence to head change. After classifying recoverable and nonrecoverable subsidence, spatial maps of both recoverable and nonrecoverable subsidence can be generated. These maps show that the average recoverable and nonrecoverable subsidence over a four-year period in the study area is approximately 10.97 cm and 5.72 cm, respectively. Finally, the volume of nonrecoverable groundwater storage loss accounts for 34% of the overall storage.