Enhanced Land Subsidence Interpolation through a Hybrid Deep Convolutional Neural Network and InSAR Time Series

Abstract. Land subsidence, the gradual or sudden sinking of the land, poses a global threat to infrastructure and the environment. This paper introduced a hybrid method based on deep convolutional neural networks (CNN) and persistent scattered interferometric synthetic aperture radar (PSInSAR) to estimate land subsidence in areas where PSInSAR cannot provide reliable measurements. This approach involves training a deep CNN with subsidence driving forces and PSInSAR data to learn patterns and estimate subsidence values. Our evaluation of the model shows its efficiency in overcoming the discontinuities observed in the PSInSAR results, producing a continuous subsidence surface. The deep CNN was evaluated on training, validation, and testing data, resulting in mean squared errors of 5 mm, 9 mm, and 11 mm, respectively. In contrast, the kriging interpolation method showed a mean square error of 37.19 mm in the experimental data set. subsidence prediction using the deep CNN method showed a 70 % improvement compared to the Kriging interpolation method.