Deep learning-based turbulent atmospheric delay correction and ground subsidence reconstruction in complex multi-source adjacent subsidence zones

Time-series InSAR is powerful for large-area ground subsidence monitoring, but turbulent atmospheric delay can strongly couple with deformation signals in multi-source, adjacent, and cross-scale subsidence settings, blurring deformation boundaries and masking weak deformation. To address this problem, we propose a lightweight spatiotemporal convolutional autoencoder for turbulent atmospheric delay suppression and complex subsidence reconstruction. The framework integrates 3D convolutions, depthwise separable convolutions, a feature pyramid network, and a convolutional block attention module to enhance spatiotemporal feature extraction, cross-level fusion, and boundary preservation. A composite loss combining L1, SSIM, and semivariogram constraints is further optimized using gradient-balanced dynamic weighting. For training, multi-source adjacent subsidence fields are simulated using elastic half-space dislocation models, while anisotropic semivariogram-based turbulence is introduced to generate directional, multi-scale atmospheric disturbances. Experiments in the Diaobingshan coal-mining area show that the proposed model reduces the maximum standard deviation from 55.2 mm to below 10 mm and decreases the mean from 27.3 mm to 9.2 mm, corresponding to a 66.3% reduction. Tests in the Yellow River Delta further demonstrate its generalization capability. Finally, Coulomb failure stress analysis is used to interpret the coupling between subsidence mechanisms and stress-field distribution.