SS-MambaUnet for fault segmentation in multi-source remote sensing images

Complex landforms and harsh environments restrict remote sensing-based fault interpretation, and manual visual inspection is inefficient. Existing deep learning methods have evident drawbacks: convolutional neural networks are weak in long-range feature modelling, while Transformers incur high computational costs, easily causing incomplete fault boundaries and scale confusion in complex geological scenes. As emerging state space models, Mamba architectures support efficient long-range dependency capture with linear computational complexity, but they fail to fully exploit multi-band spectral information and may break spatial continuity during feature scanning, impairing fault edge representation. To address these challenges, this paper proposes SS-MambaUnet, a novel multi-source remote sensing model for fault segmentation. To obtain high-quality multi-source input, this paper adopts an enhanced HSVfusion method to integrate Sentinel-1A SAR and Sentinel-2 optical data, combining fine micro-topographic details and abundant spectral characteristics. Taking U-Net as the backbone, the proposed network embeds a Band Attention Module (BAM) for multi-source input to realize adaptive feature selection. Meanwhile, it innovatively introduces Spectral Mamba (Spe-Mamba) and Spatial Mamba (Spa-Mamba) modules to collaboratively capture long-range dependencies in both spectral and spatial dimensions. The study area covers parts of the Qinghai – Tibet Plateau and the southwestern margin of the Junggar Basin in Xinjiang. Experimental results demonstrate that SS-MambaUnet achieves a mean Intersection over Union (mIoU) of 88.23%, outperforming U-Net, DeepLabV3+, and Swin-Unet by 8.71, 7.54, and 3.65% points, respectively, as well as recent Mamba-based models by an average of 1.42% points.