Accurate road graph extraction from satellite imagery is essential for large-scale mapping and geospatial analysis. Recent one-shot graph extraction frameworks based on foundation models have achieved promising performance, but their effectiveness decreases in complex environments where road structures are partially obscured by vegetation, buildings, shadows, and other surface conditions. These occlusion-induced disturbances lead to incomplete connectivity and degraded topology reconstruction, particularly under out-of-domain scenarios. This study proposes an occlusion-aware refinement framework to improve the robustness of satellite image road graph extraction while maintaining the original backbone architecture. The proposed framework introduces three complementary strategies: synthetic occlusion augmentation for explicit occlusion-aware representation learning, an occlusion-adaptive extended-line strategy with hard-mining topology optimization for improved connectivity reasoning, and an occlusion-adaptive node-guided resampling mechanism for reliable graph node localization. Experiments conducted on the Global-Scale road graph extraction benchmark demonstrate that the proposed method consistently improves topology reconstruction performance. Compared with the reproduced SAM-Road++ baseline, the proposed framework improves TOPO F1 from 61.81 to 62.56 on the in-domain split and from 46.93 to 51.51 on the out-of-domain split. Furthermore, the ID-OOD performance gap is reduced from 14.88 to 11.05, indicating enhanced robustness under unseen geographic conditions. The results demonstrate that explicitly modeling occlusion as a structured factor can effectively improve the generalization capability of satellite road graph extraction systems.

Remote sensing captions typically interleave domain-specific terminology, object attributes, spatial relations, and scene context within a single sentence, while the corresponding imagery contains targets spanning markedly different spatial extents—from individual vehicles to entire airport layouts. A single global text vector and fixed-scale visual pooling therefore tend to discard the local and hierarchical evidence needed for reliable image–text alignment. Moreover, standard contrastive objectives treat all unmatched pairs as equally unrelated, ignoring the fact that semantically adjacent scenes (e.g., airports and industrial zones) share substantial visual content. To address these issues, we develop a CLIP-style dual encoder with three targeted additions: a text encoder that preserves word-level, phrase-level, and sentence-level representations before fusing them through cross-granularity attention; a visual encoder that associates four hierarchical feature levels with scale-aware receptive-field configurations and incorporates SAM-guided local views during training; and a semantic membership-aware contrastive loss that applies detached, bounded discount weights to semantically close negatives without inverting their penalty sign. On the cleaned RSICD benchmark, the model achieves 87.5% mAP for in-domain prompt-based scene recognition and 34.37% mean recall for scene-constrained retrieval, outperforming the quoted RemoteCLIP and fully fine-tuned CLIP results in the scene-constrained retrieval comparison. Ablation confirms that each of the three components contributes meaningfully to the final performance. The deployable dual encoder contains approximately 59.7 million trainable parameters, and SAM is not required at inference.

Remote sensing captions often combine domain terms, object relations, and scene context, while the corresponding images contain targets that occupy markedly different spatial scales. A single global text vector and fixed-scale visual aggregation can therefore discard information needed for image-text alignment. We develop a CLIP-style dual encoder with three additions: a text encoder that fuses word-level, phrase-level, and sentence-level representations; a visual encoder that associates four feature levels and uses SAM-guided local views during training and validation; and a semantic membership-aware contrastive loss with detached, bounded weights for semantically close negatives. On cleaned RSICD, the model obtains 87.5% mAP and 78.3% Top-1 accuracy for in-domain prompt-based scene recognition. Under scene-constrained retrieval with transductive dual-softmax reranking, it reaches 34.37% mean recall, compared with 30.58% for RemoteCLIP and 30.39% for fully fine-tuned CLIP. Removing the text encoder, scale-aware visual branch, or membership-aware loss lowers mAP by 6.8, 6.3, and 7.8 percentage points, respectively. The deployable dual encoder has approximately 59.7 million trainable parameters. These results apply to daytime optical imagery, a cleaned closed-category benchmark, in-domain prompts, and a fixed-gallery retrieval protocol.

Accurate road graph extraction from satellite imagery is essential for large-scale mapping and geospatial analysis. Recent one-shot graph extraction frameworks based on foundation models have achieved promising performance, but their effectiveness decreases in complex environments where road structures are partially obscured by vegetation, buildings, shadows, and other surface conditions. These occlusion-induced disturbances lead to incomplete connectivity and degraded topology reconstruction, particularly under out-of-domain scenarios. This study proposes an occlusion-aware refinement framework to improve the robustness of satellite image road graph extraction while maintaining the original backbone architecture. The proposed framework introduces three complementary strategies: Synthetic Occlusion Augmentation for explicit occlusion-aware representation learning, an Occlusion-Adaptive Extended-Line strategy with Hard-Mining Topology Optimization for improved connectivity reasoning, and an Occlusion-Adaptive Node-Guided Resampling mechanism for reliable graph node localization. Experiments conducted on the Global-Scale road graph extraction benchmark demonstrate that the proposed method consistently improves topology reconstruction performance. Compared with the reproduced SAM-Road++ baseline, the proposed framework improves TOPO F1 from 61.81 to 62.56 on the in-domain split and from 46.93 to 51.51 on the out-of-domain split. Furthermore, the ID-OOD performance gap is reduced from 14.88 to 11.05, indicating enhanced robustness under unseen geographic conditions. The results demonstrate that explicitly modeling occlusion as a structured factor can effectively improve the generalization capability of satellite road graph extraction systems.