S²AM-Net: Structure-semantic SAM-guided Network for Few-Shot Segmentation in Mining Areas

quan cui , longzhou hu, yan zhou, Gaodian zhou, jianxun Li

Precise mining land-cover classification is essential for monitoring environmental degradation and ecological restoration. Few-Shot Segmentation (FSS) offers a promising solution under limited annotations, but it still faces common challenges such as ambiguous foreground features and overfitting to the small support set. Moreover, mining landscapes are characterized by highly detailed structures, which are often lost during standard downsampling.
To address these issues, we propose a Structure-semantic Adaptive Matching Network (S²AM-Net). It employs an adaptive prior enhancement module to suppress background noise, selective edge attention with multi-scale learnable Sobel convolutions to preserve fine-grained structural details, and mask semantic alignment using SAM-generated masks for texture-invariant alignment.
We also construct a dedicated benchmark, DLRSD-mining-6i, for complex mining categories. Extensive experiments on this dataset and iSAID-5i demonstrate that our method consistently outperforms existing approaches. The source code is available at: https://github.com/cowqer/S2AMNet.