Multiscale Spatiotemporal Prediction Method Of Time-Series InSAR Surface Deformation Using Deep Recurrent Neural Network Model

The Beijing Plain region carries dense buildings,developed transportation,and a large population,making spatiotemporal monitoring and prediction of surface deformation crucial.With the development of deep learning technology,the spatiotemporal prediction of surface deformation using synthetic aperture radar interferometry(InSAR)based on deep learning has become a research hotspot.However,the existing InSAR deep learning spatiotemporal prediction models for surface deformation have low reliability and a short effective prediction time.This study first utilized Small Baseline Subset InSAR(SBAS-InSAR)technology to monitor temporal surface deformation in the Beijing Plain from May 2017 to May 2020;Then,based on time-series InSAR surface deformation data,a deep recurrent neural network multiscale spatiotemporal prediction model for surface deformation is constructed,including a largescale temporal prediction model for surface deformation based on Time Distribution Fully Connected Recurrent Neural Network(TDFC-RNN)and a fine scale spatiotemporal prediction model for surface deformation based on Conv RNN;Finally,based on the proposed model,large-scale and fine-scale spatiotemporal prediction of future surface deformation in the Beijing Plain research area and key areas will be conducted.The specific research content and main conclusions are as follows:(1)The SBAS-InSAR technology was used to obtain the time-series InSAR surface deformation data of Beijing Plain in 2017-2020,and the reliability of InSAR results was verified based on internal precision and leveling data validation.The results show that the surface deformation in areas such as Dongcheng District,Xicheng District,Shijingshan District,and Fengtai District of the Beijing Plain is basically stable;However,there are varying degrees of land subsidence in areas such as the southeast of Changping District,the northwest of Haidian District,the east of Chaoyang District,the northwest of Tongzhou District,the southwest of Shunyi District,and the south of Daxing District.Among them,the northwest of Haidian District,the east of Chaoyang District,and the northwest of Tongzhou District have severe land subsidence,while the east of Chaoyang District has the most severe land subsidence.In view of this,the eastern part of Chaoyang District and the northwestern part of Tongzhou District were selected as key areas for subsequent fine-scale spatiotemporal prediction research.(2)A TDFC-RNN spatiotemporal prediction model for large-scale surface deformation based on RNN was constructed to predict the future trend of surface deformation in the Beijing Plain.Pre-processing of temporal InSAR data such as grid sampling,differentiation,standardization,and dataset partitioning is carried out.A TDFC-RNN surface deformation temporal prediction model is built based on RNN with Encoder-Predictor-Decoder framework pattern as the backbone network,and a newly proposed temporal attention mechanism module is added for the performance comparison of evaluation indicators.All indicators indicate that the proposed spatiotemporal prediction method for surface deformation is superior to traditional machine learning methods.At the same time,the prediction results for subsequent time steps are in line with the future development trend of surface deformation in the research area,verifying the stability and reliability of the proposed TDFC-RNN large-scale spatiotemporal prediction model for surface deformation,and the credibility of the temporal prediction results.(3)A fine scale spatiotemporal prediction model for surface deformation based on Conv RNN neural network was constructed,and the spatiotemporal trend of surface deformation in key areas(Chaoyang-Tongzhou strong subsidence area)was predicted.In order to obtain more accurate prediction results of surface deformation,the model needs to be able to extract spatial neighborhood features more accurately.At the same time,when conducting more refined monitoring and mapping of significant deformation areas within the research area,the required boundary shapes are often more regular.By preprocessing time-series InSAR data with grid sampling,null filling,segmented function fitting numerical compression,and dataset partitioning,a surface deformation spatiotemporal prediction model is built with Conv RNN as the core and Encoder-Predictor-Decoder framework mode as the backbone network.The Conv RNN layer is designed bi-directional,including various Conv RNN variants,and the performance indicators of the model are compared.The results indicate that the constructed fine-scale Conv RNN spatiotemporal prediction model has good and stable performance,and its prediction results are more intuitive and richer in spatial details.This indicates that the proposed Conv RNN fine-scale InSAR spatiotemporal prediction model for surface deformation is reliable and credible,making it convenient for subsequent research on fine-scale surface deformation spatiotemporal prediction.