A Deep Learning-Based Detection of Ground Deformation Using a Single Interferogram

Accurate classification and localization of geophysical deformation, such as volcanic and coseismic events, are essential in geohazard monitoring and risk mitigation. Interferometric Synthetic Aperture Radar (InSAR) is a valuable tool for detecting ground deformation. However, analyzing interferograms, especially when using single interferograms, remains challenging due to the complex nature of the mixed signals. In this study, we investigate the application of deep learning (DL) models—namely, the residual convolutional neural network (ResNet), Vision Transformer (ViT), and Vision Mamba (ViM)—for the automatic classification and localization of deformation patterns from single interferograms. Synthetic and real interferogram datasets were used for model training and evaluation. By incorporating transfer learning (TL), we significantly enhanced the performance of the model, particularly on real-world data where labeled samples are scarce. Our results show that the ViT model outperforms ResNet and ViM, achieving higher classification accuracy and more precise localization of deformation regions. The global attention mechanism of ViT allows it to capture both local and long-range dependencies in the data, making it particularly well-suited for detecting more diffuse deformation patterns, as is often the case in volcanic activity. In contrast, ResNet and ViM exhibited limitations in handling these more complex spatial patterns. TL was instrumental in bridging the performance gap between synthetic and real data, enabling better generalization across different interferogram scenarios. This work demonstrates the potential of transformer-based models in geophysical applications and highlights the importance of TL when dealing with limited real-world datasets. Our findings contribute to advancing the integration of DL and InSAR technologies, offering promising avenues for developing automated, real-time geohazard detection systems capable of accurately identifying and localizing deformation events across diverse environments.