Interferometric Synthetic Aperture Radar–Guided Multi-Modal Deep Learning for Surface Deformation Target Detection and Boundary Extraction in Complex Mountainous Areas

Abstract

Automatic detection and precise boundary delineation of surface deformation targets in complex mountainous terrain are essential for monitoring geohazards such as landslides and mining-induced subsidence. To address the low efficiency of conventional visual interpretation, the fragmented boundaries derived from single-source interferometric synthetic aperture radar (InSAR) results, and the limited capability of optical imagery to characterize deformation activity, this study proposes an InSAR-guided two-stage multi-modal recognition framework. First, small baseline subset InSAR processing was performed using Sentinel-1 ascending-track data, and candidate deformation regions characterized by negative line-of-sight (LOS) anomalies were automatically generated using robust statistical thresholds, spatial-connectivity constraints, and topographic rules. Subsequently, within the candidate regions, InSAR-derived LOS deformation velocities, Sentinel-2 optical bands and spectral indices, and digital elevation model (DEM)–derived elevation and slope information were integrated to construct a multi-channel input. Deep learning–based semantic segmentation models were then used for target recognition and precise boundary delineation. Experiments were conducted using Fa’er Town in Liupanshui City, Guizhou Province, as the source domain and Panzhou City as the geographically adjacent target domain. The results showed that the convolutional neural network + Transformer model achieved the best overall performance in the within-domain evaluation, with intersection over union (IoU), boundary F1 score, and mean symmetric distance values of 0.9716, 0.9789, and 1.225 m, respectively. The ablation experiments indicated that the fusion of InSAR, optical, and DEM data yielded the best performance, achieving precision, recall, Dice, and IoU values of 0.996, 0.996, 0.996, and 0.991, respectively. The cross-region transfer experiment between geographically adjacent domains showed that directly transferring the source-domain model to the target domain resulted in a marked performance degradation, whereas few-shot fine-tuning substantially restored the model’s recognition capability in the target domain. These findings demonstrate that the proposed method effectively integrates deformation activity, spectral–textural features, and topographic constraints, thereby providing methodological support for the automatic detection and precise boundary delineation of surface deformation targets in complex mountainous terrain.