Classification of Urban Land Subsidence Types in Fuzhou from Time-Series InSAR Using FFT-Based Filtering and Ensemble Learning

Accurate identification of land subsidence types is essential for effective urban risk management, yet remains challenging due to the superposition of deformation signals at different spatial scales and the complexity of urban environments. In this study, land subsidence types in Fuzhou were identified through an integrated framework combining multi-scale deformation analysis and ensemble learning. Ground deformation time-series measurements were derived from 66 Sentinel-1A synthetic aperture radar (SAR) observations acquired between January 2018 and June 2023 using the time-series interferometric synthetic aperture radar (TS-InSAR). Deformation values in decorrelated areas were subsequently reconstructed using regression models driven by multi-source geological, hydrological, land-use, and urban features, resulting in a spatially continuous deformation field. A Fast Fourier Transform (FFT)-based Butterworth filtering approach was then applied to separate regional-scale and local-scale subsidence signals. Based on the extracted local deformation patterns and discriminative auxiliary features, land subsidence was classified into five categories: farmland-related subsidence, linear infrastructure-related subsidence, low-lying stratum-related subsidence, land-use transition-related subsidence, and older building area-related subsidence. Three ensemble learning models, XGBoost, CatBoost, and LightGBM, were implemented for subsidence type classification. All models achieved satisfactory performance, among which LightGBM exhibited the best overall performance. The classification results reveal pronounced differences in spatial distribution and deformation intensity among subsidence types. Farmland-related subsidence occupies the largest proportion of the affected area but is characterized by relatively moderate deformation rates, whereas older building area-related subsidence, despite its limited spatial extent, exhibits the highest deformation intensity. This study demonstrates the potential of ensemble learning for land subsidence type classification.