A TDFC-RNNs framework integrated temporal convolutional attention mechanism for InSAR surface deformation prediction: A case study in Beijing Plain

The precise time series prediction method is the key technology for the monitoring and management of ground deformation. Current prediction methods mostly rely on independent sampling points for prediction, limiting the effective utilization of spatial features by the model, thereby affecting the overall spatial prediction accuracy, and it also restricts the prediction efficiency of the model. In response to the above-mentioned issues in previous research, this study proposes a Time Distributed Fully Connected (TDFC) Recurrent Neural Networks (RNNs) framework that integrates Temporal Convolutional Attention Mechanism (TCAM) for joint prediction of sampling points in time series Interferometric Synthetic Aperture Radar (InSAR) surface deformation data. Firstly, based on Sentinel-1A imagery over the Beijing Plain, the time series surface deformation data from May 2017 to April 2020 are obtained utilizing the Small Baseline Subset InSAR (SBAS-InSAR) technology. After data processing and production into a dataset, based on the TDFC-RNNs framework integrated with TCAM, five different RNN structures were used as prediction modules to construct time series prediction models for InSAR surface deformation. To investigate the effectiveness of the TCAM module on prediction performance, ablation experiments were conducted specifically targeting it. Furthermore, to explore the relative optimality choice of prediction modules under the current dataset and the compatibility of this framework with non-RNN structures, various other sequence models were selected as prediction modules. The predictive performance of the models constructed by this framework was compared in two aspects with benchmark methods, ablation models, and other exploratory models. This included evaluating the predictive results of the test set using various metrics and analyzing the trends in numerical characteristics of the predicted results for the next 60 time steps (720 days). The comprehensive comparison results indicate that the model constructed by this framework outperforms other methods or models in terms of overall performance across various evaluation metrics. At the same time, the future predicted results exhibit more reliable numerical characteristics, aligning well with the developmental trends of surface deformation. This suggests that the above-mentioned models demonstrate favorable predictive capabilities for time series InSAR surface deformation. Such results can be instrumental in intuitively assessing the overall situation of surface deformation in the study area, promptly identifying risks, and swiftly implementing measures to address potential hazards.