<p>Surface subsidence induced by mineral exploitation poses significant risks to surrounding environments, lives, and property in mining areas. Accurate prediction of this subsidence is therefore critical for effective mitigation efforts, but remains challenging due to its complex spatio-temporal characteristics, which often exhibit both nonlinear dynamics and underlying trends. Existing single-model approaches may struggle to fully capture this complexity, potentially leading to reduced prediction accuracy. To address this, a surface subsidence prediction model called seasonal trend decomposition–extreme gradient boosting tree (STL-XGBoost) was proposed, which is based on small baseline subset interferometric synthetic aperture radar (SBAS-InSAR). The surface subsidence sequence of a mining area from 2020 to 2023 was obtained using SBAS-InSAR technology; then, STL decomposition was applied to separate the subsidence sequence and obtain the trend term. The non-trend term was used to predict the subsidence sequence for the XGBoost model, while the trend term was predicted using STL decomposition, and the nonlinear and trend term prediction values were equally weighted to obtain the final composite model prediction value. Compared with the mean absolute error (MAE) of the prediction results of the single XGBoost model, that of the STL-XGBoost model was reduced by 31%, and the root mean squared error (RMSE) was reduced by 38%. The prediction results had a higher accuracy and a strong correlation of above 0.9 with the original time series, indicating the effectiveness and reliability of the proposed method in providing a strong technical support for surface-subsidence prediction in mining areas.</p>