Open-Pit Mine Extraction from Very High-Resolution Remote Sensing Images Using OM-DeepLab

Automatic extraction of open-pit mine (OM) areas from very high-resolution (VHR) remote sensing images is important for mineral resource management and monitoring. However, few good results have been achieved on this topic due to the highly heterogeneous OM environment. Based on DeepLabv3+, this study proposes a semantic segmentation network, named OM-DeepLab, for extracting OM from VHR images. First, Xception is employed as the backbone network to extract deep features from the input image patch. Second, Atrous Spatial Pyramid Pooling (ASPP) is attached to capture semantic features at multi-scales. In order to strengthen the learning of important information, attention mechanism is embedded following the ASPP. The foregoing operations constitute an encoder. Third, in order to retain detail information in the process of feature sampling, a low-level feature multi-scale fusion (LFMF) module is proposed to form the decoder by connecting shallow features and high-level features derived from encoder. For OM extraction from large-scale VHR images, a novel model prediction method is presented. The proposed network is utilized to achieve the initial OM extraction. Finally, post-processing containing an object-based Conditional Random Fields (CRF) is performed to refine the OM extraction results. To validate the proposed method, an OM dataset containing 136 OM areas from government department of mine supervision was applied in the experiments, and pixel-level F1-score of 0.912 was achieved. In addition, to evaluate the effectiveness of the algorithm in practical applications, a large-scale VHR image covering 736 km(2) was utilized in the experiments, and the pixellevel F1-score, object-level FalseAlarm, object-level MissingAlarm were 0.854, 0.132, 0.042, respectively. This study has practical significance in terms of automatically extracting OM from VHR images, which is beneficial to OM management and monitoring.