Automated deformation detection and interpretation using InSAR data and a multi-task ViT model

Abstract

Many geological hazards are associated with ground deformations. Prompt and accurate detection and interpretation of ground deformation is therefore vital to geohazard mitigation. Multitemporal Interferometric Synthetic Aperture Radar (MT-InSAR) is an effective geodetic technique for monitoring ground deformation. However, accurate computation and interpretation of deformation using InSAR are often hindered by various errors and a lack of expert knowledge. We present a new advanced deep learning model based on a multi-task vision transformer (MT-ViT) to automatically detect, locate, and interpret deformation using single interferograms. To address the issue of limited training data in InSAR applications, the proposed model utilizes pre-trained weights from optical images and transfers them to a simulated InSAR dataset. Then real interferograms are used to fine-tune the weights in the network. An overall loss function is designed, which considers the classification and localization losses in the model. The effectiveness of the proposed model is demonstrated using both simulated and real InSAR datasets that contain either coseismic or volcanic deformation. The experimental results from the model are also compared with the state-of-the-art convolutional neural network (CNN) based techniques. The results show significant improvement in both the accuracy of the results and the computational efficiency over the CNN-based approaches. The MT-ViT model achieved 99.4 % classification accuracy, 54.1 % mean intersection over union (IOU), and 0.9 km localization accuracy. A comprehensive evaluation of the hyperparameters in training the MT-ViT model was carried out, which will inform future research in this direction. The research results highlight the promising capabilities of MT-ViT in near real-time deformation monitoring and automated deformation interpretation.