A Decoupled Segmentation-Classification Strategy Based on Semantic-SAM for Precise Semantic Segmentation in Coal Mine Areas

To address complex semantic segmentation in coal mine areas, this study proposes the SAM-SEF (SAM-based Semantic Enhancement Framework). It integrates Semantic-SAM’s zero-shot segmentation capability with specialized deep learning models through a “segmentation-classification-update” strategy. Semantic-SAM initially provides category-agnostic multi-granularity segmentations; specialized deep learning models then perform semantic understanding of these regions, and a category mapping mechanism generates precise results. Evaluating 18 models, SAM-SEF-ENet-B0 achieved the highest IoUs for Open-pit Coal Mining Sites (OCMS) (93.68%) and Composite Coal-related Sites (CCS) (81.52%), with an mIoU of 87.60%, significantly outperforming traditional models. A Semantic-SAM-based semi-automatic annotation tool reduced manual annotation time by approximately 60%. A multi-scale evaluation system (based on 33% and 66% target area percentiles) revealed a positive correlation between target scale and segmentation performance, with small-scale target segmentation being the primary challenge (IoUs of 0.79 for OCMS and 0.67 for CCS in small scales). Systematic experiments validated Semantic-SAM’s superiority over SAM (mIoU 54.03%) and SAM2 (mIoU 45.33%) in semantic understanding, architectural adaptability, and computational efficiency. The study also found that classification networks within the SAM-SEF framework outperformed specially designed segmentation networks. This demonstrates the effectiveness of a “classification-driven segmentation” paradigm, where lightweight classifiers operating on purified targets from a frozen foundation model can surpass traditional end-to-end segmentation networks in specific, complex domains, offering new technical solutions for remote sensing monitoring.