SubSAM-detect: Segment anything model-driven intelligent Hazard annotation and multi-scale defect detection for substation satellite inspection imagery

Automated hazard detection in high-resolution satellite imagery of electrical substations presents significant challenges, including complex background clutter, high intra-class defect variability, and the prohibitive cost of pixel-level annotation. Existing methods are limited by their reliance on extensive manual labeling and their inability to generalize across diverse substation configurations and equipment types. This paper presents SubSAM-Detect, a novel integrated framework that synergizes the Segment Anything Model (SAM) with domain-adapted large vision-language models (LVLMs) to enable annotation-efficient, zero-shot hazard detection in substation satellite imagery. The framework operates as a unified three-stage pipeline: (1) SAM-LoRA fine-tuning with substation-specific prompt engineering for precise infrastructure component segmentation (transformers, bushings, insulators, conductor arrays); (2) CLIP-enhanced hierarchical hazard classification propagating SAM-generated segments through a multi-granularity defect taxonomy covering vegetation encroachment, oil leakage, thermal anomalies, and structural deformations; and (3) confidence-calibrated active learning for iterative pseudo-label refinement. Compared against five state-of-the-art baselines on the SubStation-Sat (12,000 images, 150 substations) and PowerGrid-Global (8500 images, 85 substations) datasets, SubSAM-Detect achieves 89.3% mAP@0.5, a 6.9% improvement over the best competing method (GDPGO-SAM at 82.4% mAP@0.5), while reducing manual annotation requirements by 85% and improving small-scale defect detection by 15.7%.