Land Subsidence Time Series Prediction Method Based on LSTM-AMSGD

Abstract

Accurate prediction of geological subsidence is of great importance for geological hazard risk assessment. Various existing prediction models do not take into account the time correlation between geological subsidence, and the prediction effect lacks practical significance. In this paper, an LSTM-AMSGD-based land subsidence prediction method is proposed. Firstly, the high-precision time series inversion results of large-area land surface deformation are obtained by the small baseline interference technique with multiple principal image coherent targets. Secondly, a recurrent neural network (LSTM-AMSGD) is used as the network architecture. The final cumulative subsidence prediction error is within 0.3 mm, and the single-step prediction of more than 400,000 observation points can be completed in 126s. Therefore, the LSTM-AMSGD model in this paper is effective for the prediction of geological subsidence.