AutoRoadSAM: Multimodal Remote Sensing Road Extraction With Structure-Semantic Awareness via Auto-Prompting Vision Foundation Models

The integration of multimodal data holds great promise for advancing road extraction in remote sensing. However, existing approaches are limited by the lack of unified end-to-end frameworks for diverse modality combinations, suboptimal multimodal feature fusion, and challenges in capturing the slender, winding, and complex topological structures of roads. In this paper, we propose AutoRoadSAM, a novel end-to-end framework for multimodal road extraction that fully exploits the powerful visual representation capabilities of the Segment Anything Model (SAM) and, for the first time, introduces an Auto-Prompting Mechanism via a Dynamic Snake Convolution-based Decoder. This decoder adaptively generates task-specific prompts by capturing fine-grained local geometric features from auxiliary modality branches, enabling precise alignment with complex road structures. To further enhance multimodal feature fusion and topological perception, we design the Cross-Modal Information Interaction (CMII) module, which facilitates global context modeling and cross-modal interaction, while strengthening the representation of intricate road topology through multidirectional snake scanning. Moreover, we incorporate a Mask Decoder with Cross Polarity-aware Linear Attention to boost decoding efficiency and effectively address pixel imbalance. Together, these innovations enable AutoRoadSAM to achieve superior structure- and semantic-aware road extraction across diverse modality combinations. Extensive experiments on six public datasets and four modality combinations demonstrate that AutoRoadSAM consistently outperforms state-of-the-art methods, validating the effectiveness and generalization capability of each proposed component. The code is available at https: //github.com/NWPUFranklee/AutoRoadSAM.git.