MineFormer: A Lightweight Edge-Aware Transformer for High-resolution Semantic segmentation of Open-pit Mining Area

Accurate identification of open-pit mining areas is crucial for sustainable land use and ecological restoration. Remote sensing (RS) technology provides an efficient and dynamic monitoring solution for large-scale open-pit mining area identification, overcoming the spatiotemporal limitations of traditional manual inspections to enable rapid perception of mining boundaries, mining activities dynamics, and ecological disturbances. However, current methods represented by semantic segmentation still face multiple challenges: spectral ambiguity, terrain-induced edge blurring, and computational burdens of deep models. In this study, a novel multi-scale semantic segmentation framework (MineFormer) is proposed for extracting open-pit mining areas from high-resolution imagery. The MineFormer introduces three key modules: a linear Agent Transformer Encoder (ATE) that models global context efficiently, an Edge Fusion Module (EFM) that integrates Laplacian-based edge priors to enhance spatial boundaries, and an Edge-Aware Transformer Decoder (EATD) for fine-grained detail recovery. We construct a new benchmark dataset covering five mining-intensive cities in Guangdong Province using Google Earth imagery. Extensive experiments demonstrate that MineFormer outperforms eight state-of-the-art (SOTA) baselines—including CNN and Transformer variants—with the highest F1-score of 86.62% on our dataset and mean intersection-over-union (mIoU) of 77.52% on Fine-RSMI dataset, while maintaining computational efficiency. This study provides a scalable and accurate solution for automated open-pit mining monitoring, with direct implications for ecological assessment and sustainable mining management. • Proposes an edge-aware Transformer for precise open-pit mining area extraction. • Designs boundary-guided modules to boost edge integrity and spatial consistency • Presents a new mining dataset (GDMine) with multi-city and foggy scenarios. • Achieves state-of-the-art performance on binary and multi-class mining benchmarks.