Abstract

One of the pivotal components of the material handling system in mining sites is the haul road, which serves as the lifeline of material movement within a mining operation. These extensive road networks are specifically engineered to withstand the immense weight and frequent use of heavy mining vehicles, such as haul trucks. An efficient haul road design and maintenance facilitates seamless and expedited materials transportation. It ensures that trucks can travel at optimal speeds without unnecessary delays due to road conditions or maintenance issues.

However, frequent changes in road trajectory and complex shapes make it challenging to update these networks regularly. Unlike urban roads, mining roads lack clear edges and share spectral similarities with their surroundings, making them difficult to distinguish. Furthermore, their unstructured layout contrasts sharply with the straight, well-defined roads of urban settings. Together, these factors underscore the urgent need for efficient methods to accurately generate and frequently update road network data in open-pit mining environments.

Recently, the use of remote sensing technologies for optical detection methods has gained significant popularity. Moreover, recent advances in learning-based algorithms, particularly in deep learning and computer vision, have significantly enhanced the capability to analyze remote sensing data. Consequently, interest in utilizing this data has surged. This research aims to investigate these advancements and outline a novel approach to effectively leverage satellite imagery for road network analysis in mining operations.

Fir this aim, a transformer-based framework proposed which focuses on automated mining haul road extraction using satellite imagery is proposed. To achieve high-level precision, the proposed model combines several key modules tailored to the unique challenges of mining road extraction. The architecture is designed as an encoder-decoder framework, followed by a decision-head part, where the encoder extracts relevant features, and the decoder predicts accurate outputs. SegFormer, a state-of-the-art transformer-based framework for semantic segmentation, serves as the backbone of the encoder, chosen for its capability to capture multi-scale features and context across complex textures. To emphasize significant spatial and channel-wise details, the model incorporates Efficient Channel Attention immediately at the end of the encoder, which filters key channels. At the same time, the Convolutional Block Attention Module is applied before the decision head to refine spatial features and enhance attention to relevant regions. In the decision head, the model utilizes pixel connectivity and road edge detection to improve segmentation accuracy further. An AugMix-based augmentation strategy further improves robustness by introducing greater variability. Satellite data collected from different mine sites in North and South America were used for this study. Evaluated on this comprehensive mining road dataset, our approach achieved %93.8 accuracy, %72 F1-score, %67.8 mIoU, %73.2 recall, and %73.8 precision. In addition, the model’s modular design allowed us to examine each component’s effectiveness individually, demonstrating how edge detection yields sharper boundaries and how the connectivity module improves continuous segmentation. This comprehensive analysis highlights the distinct contributions of each module and provides valuable insights for further optimization.