Abstract
A dead-end road is a roadway whose terminus does not connect with other established networks. In remote mountainous regions distant from human habitation, these endpoints often correlate with specific features like mine shaft entrances, underground engineering access points, or specialized non-residential facilities, making them critical indicative clues in optical remote sensing. However, the effective identification of these dead-end road association zones and the classification of their features remain a key technical challenge. This paper proposes a classification framework for dead-end road association-zone scene classification. Road network information is first exploited to locate dead-end endpoints and construct direction-constrained association zones along road extension directions. Scene classification is then performed by integrating global visual context with object-level semantic relations to enhance discriminability under complex spatial layouts. Using a newly constructed dataset of 4409 image samples spanning seven scene categories, extensive experiments conducted across six representative mountainous regions demonstrate stable and consistent classification performance of the proposed framework, with a combined overall accuracy of 92.16% across all study areas. The results indicate that direction-constrained context modelling effectively reduces background interference and improves scene discriminability in mountainous environments. While the proposed method can improve scene discriminability, we acknowledge that it remains challenging to distinguish visual dead-ends (e.g., tunnel portals) from functional dead-ends (e.g., mine entrances) using only optical imagery. Overall, this work provides an effective solution for analyzing dead-end-road-associated scenes from optical remote sensing imagery, providing a reference for infrastructure monitoring and land-use analysis in complex terrain regions.