GeoSeg-SAM: Task-aligned adaptation of SAM for geological remote sensing interpretation

The Segment Anything Model (SAM) has shown strong performance in natural image segmentation. However, its direct transfer to geological environment remote sensing (GERS) is limited by significant domain discrepancies. GERS elements typically exhibit high class diversity, strong inter-class similarity, and imbalanced sample distribution, which challenge SAM’s prompt-driven paradigm in maintaining semantic stability and spatial contextual consistency. To bridge this gap, we propose GeoSeg-SAM, a geological task-aligned adaptation framework built upon SAM. GeoSeg-SAM progressively aligns generic visual features with geology-discriminative semantics through layer-wise semantic adapters, and introduces a dual-scale structure-aware module in the deeper encoder layers to enhance global region-level contextual representation. Meanwhile, the model employs Pyramid Residual Aggregation (PRA) to fuse multi-level geological semantic representations across stages, thereby improving the representational capacity of features. In addition, at the mask prediction stage, decoded features are fused with PRA-based semantic representations using a semantics-constrained fusion decoding strategy, and a single geological semantic token is incorporated to enhance semantic consistency in GERS scenarios. Experimental results on three representative GERS study areas show that GeoSeg-SAM improves mIoU by approximately 3 percentage points or more over the best competing method in each area, confirming its effectiveness in GERS image interpretation tasks. Code is available at: https://github.com/WanZhan-lucky/GeoSeg-SAM .