Satellite Imagery-Based Traversability Analysis Using Self-supervised Learning Policy for Global Path Planning

This paper proposes a self-supervised traversability analysis method based on satellite imagery to enhance long-range navigation capability for Unmanned Ground Vehicles (UGV) in unstructured environments. The method integrates high-resolution satellite images and historical trajectory data to construct a vehicle-centric terrain similarity map, enabling continuous modeling of traversability over large-scale environments. The system computes structural similarity between current observations and historical terrain patterns to generate a continuous traversability score map. Compared to segmentation-based approaches such as Segment Anything (SAM), the proposed method requires no manual annotations or pixel-wise labels during training or deployment, offering lower application cost and stronger adaptability across diverse scenarios. Experimental results show that the method effectively identifies traversable regions across various complex terrains and provides stable support for path planning without requiring prior map construction.