Publication Date: 2025
Geoscientific Model Development (1991959X)18(19)pp. 6903-6919
Land subsidence, whether gradual or sudden, poses a significant global threat to infrastructure and the environment. This study introduces a hybrid approach that combines deep convolutional neural networks (CNNs) with persistent scatterer interferometric synthetic aperture radar (PSInSAR) to estimate land subsidence in areas where PSInSAR data are unreliable or sparse. The proposed method trains a deep CNN using subsidence driving forces and PSInSAR data to learn spatial patterns and predict subsidence values. Our evaluation demonstrates that the CNN effectively mitigates discontinuities in PSInSAR results, producing a continuous and reliable subsidence surface. The model's performance was assessed using training, validation, and testing datasets, achieving root mean square errors (RMSEs) of 3.99, 8.47, and 9 mm, respectively. In contrast, traditional interpolation methods such as kriging, inverse distance weighting (IDW), and radial basis function (RBF) interpolation yielded RMSE values of 61.60, 66.21, and 61.76 mm, respectively, on the test dataset. Additionally, the coefficients of determination (R2) for CNN, kriging, IDW, and RBF were 0.98, -0.06, -0.22, and -0.06, respectively. The deep CNN model demonstrated an 85 % improvement in subsidence prediction accuracy compared to conventional interpolation methods, highlighting its potential for accurate and continuous land subsidence estimation. Copyright © 2025 Zahra Azarm et al.
