DUTransUNet: dual-decoder transformer U-net for differential interferogram deformation regions segmentation

Differential Interferometric Synthetic Aperture Radar (DInSAR) is widely used for surface deformation monitoring. However, automatic segmentation of deformation regions in differential interferograms remains challenging because deformation-related fringes in wrapped differential interferograms are often superimposed on non-deformation phase contributions, including atmospheric phase screens, orbital errors, and decorrelation noise. Accordingly, we propose an enhanced Transformer U-Net with a dual-decoder design, DUTransUNet. Built on TransUNet, our model adds an auxiliary supervision decoder to strengthen intermediate feature learning and multi-scale representation, while also incorporating a Convolutional Block Attention Module (CBAM) and refinement convolutions to enhance target responses and boundary precision. Experiments on the Hephaestus dataset show that DUTransUNet achieves an mIoU of 83.07%, outperforming TransUNet and Swin-UNet by 3.67% and 4.33%, respectively. It also improves precision, recall, and F1-score, with notable gains in deformation segmentation and boundary delineation under challenging backgrounds, confirming its effectiveness in complex interferometric data.