C-LSTM for MT-InSAR Ground Deformation Prediction

Currently, Multi-Temporal InSAR (MT-InSAR) is extensively employed to predict the trend of ground deformation. The deformation data acquired through MT-InSAR has been utilized as a single parameter in the model for predicting land deformation. Nevertheless, these models still necessitate enhancement in terms of their ability to generalize and accurately predict outcomes. In this paper, a combined Long Short Term Memory (C-LSTM) is proposed to combine groundwater level, rainfall, and surface deformation information from MT-InSAR. We assess the predictive accuracy of single-factor and multi-factor models. The results show that after feature combination optimization, the R
²
of the C-LSTM subsidence prediction model with multi-feature training improves the prediction results by 9.7%, 0.48%, and 21.82%, respectively, over the prediction results of the single-feature-factor model. By improving the C-LSTM’s feature factors, this method improves the accuracy of the forecast of ground subsidence change areas.