Abstract: High−precision road networks in open−pit mining areas constitute a critical prerequisite for the implementation of intelligent dispatch and autonomous driving technologies. However, the unstructured roads in such environments present challenges including blurred edges and severe background interference, rendering accurate acquisition of road network data difficult. To address this issue, this paper proposed a method for constructing unstructured road networks in open−pit mines based on an improved CSHformer model. This method first constructed a CSHformer semantic segmentation network. Features were extracted through a cross−shaped sliding−window self−attention mechanism and locally enhanced positional encoding. Feature fusion was achieved using the lightweight Hamburger model, while a loss function combining binary cross−entropy and Focal Loss was adopted to address class imbalance. Subsequently, the extracted road network was subjected to gap filling, refinement, and super−resolution optimisation to ensure timely and effective data updates. Finally, latitude and longitude information from the original images was assigned to the extracted road network, completing the construction of an unstructured road network model for open−pit mining areas. Experimental results demonstrate that the CSHformer network achieves 86.54% accuracy in road network extraction, enabling efficient and precise extraction and updating of open−pit mine roads. The constructed road network map boasts an accuracy of 0.05 m, meeting the high−precision positioning requirements of transport vehicles in open−pit mines. This method effectively resolves the challenge of acquiring unstructured road network data in open−pit mines, providing reliable support for intelligent mine scheduling and autonomous driving applications.