Ground surface displacement measurement from SAR imagery using deep learning

Abstract

Offset tracking using synthetic aperture radar (SAR) amplitude imagery is a valuable technique for detecting large ground displacements. However, the traditional offset tracking methods with the SAR datasets are computationally intensive and require significant time for processing. We have developed a novel cross-connection Siamese ResNet (CC-ResSiamNet). The model leverages multi-kernel offset tracking for preprocessing, followed by deep learning architectures that incorporate U-Net, cross-connections, and residual and attention blocks to predict pixel offsets between two SAR amplitude images. It is trained and tested on 200 K pairs of reference and secondary SAR amplitude images, alongside corresponding target offset data from Alaska's glaciers. The comparative analysis with multiple deep learning models confirmed that our designed model is highly generalizable, achieving rapid convergence, minimal overfitting, and high prediction accuracy. Through multi-scenario inference with glacier movements, earthquakes, and volcanic eruptions worldwide, the model demonstrates strong performance, closely matching the accuracy of traditional methods while offering significantly faster processing times through parallel computing. The model's rapid displacement mapping capability shows particular promise for improving disaster response and near real-time surface monitoring. While the approach encounters challenges in accurately capturing small-scale displacements, it opens new possibilities for SAR-based surface displacement prediction using machine learning. This research highlights the advantages of combining deep learning with SAR imagery for advancing geophysical analysis, with future applications anticipated as more commercial and scientific SAR missions launch globally.