Automatic Extraction of Open-Pit Mining Areas Based on Multilevel Feature Fusion and Vegetation Feature Enhancement

Accurately extracting Open-Pit Mining Areas (OMAs) from high-resolution remote sensing imagery is of great significance for ecological restoration and sustainable resource management. Existing studies on OMAs extraction are limited by insufficient attention to overall mining area characteristics, inadequate spectral feature utilization, and feature learning bias from foreground–background class imbalance. To address this issue, this study proposes OMENet, an OMAs extraction network that combines multi-level feature fusion with vegetation feature enhancement. OMENet adopts a dual-branch encoder-decoder architecture. The main encoder extracts geometric and texture features of OMAs, while the auxiliary encoder captures vegetation disturbance features. The Feature Cross-fusion Module (FCM) was designed to enable complementary fusion of feature information from the two branches. The Dynamic Weighted Hybrid loss function (DWH) and Mining Areas Attention Segmentation Head (MASH) were constructed to address feature learning bias caused by foreground–background class imbalance. In addition, this study constructs the Open-Pit Mining Areas Dataset (OPMSet) based on GF-2 multispectral imagery, which is characterized by high spatial resolution, prominent multispectral features, and diversity in both scenes and target objects. Experimental results demonstrate that OPMSet is suitable for training various semantic segmentation models, achieving accuracies primarily above 90%. And OMENet outperforms ten other comparison methods across seven accuracy evaluation metrics. Notably, for key semantic segmentation metrics such as IoU and F1-score, OMENet achieves improvements of 0.72% to 4.29% and 0.38% to 2.34%, compared with other models. This study can provide technical support for ecological restoration and sustainable resource management in OMAs.