Identification of maceral groups in Chinese bituminous coals based on semantic segmentation models

• Published in: Fuel (Guildford) 2022-01, Vol.308, p.121844, Article 121844
• Main Authors: Wang, Yue, Bai, Xiangfei, Wu, Linlin, Zhang, Yuhong, Qu, Sijian
• Format: Article
• Language: English
• Subjects: Accuracy; Algorithms; Bituminous coal; Chinese bituminous coal; Clay minerals; Coal; Coal maceral; Computer applications; Convolutional neural networks; Deep learning; DeepLab V3+ model; Image analysis; Image processing; Image quality; Image segmentation; Macerals; Machine learning; Mathematical models; Minerals; Pixels; Pyrite; Semantic segmentation; Semantics; Training
• ISSN: 0016-2361; 1873-7153
• DOI: 10.1016/j.fuel.2021.121844

•It is the first time that semantic segmentation models based on deep learning are introduced to make a pixel-level identification of coal macerals.•The semantic segmentation method based on deep learning meet the accuracy requirements of quantitative analysis of coal macerals as the average pixel accuracy of macerals determined by DeepLab V3+ model is 92%. Automatic identification of coal macerals has been a long-term pursuit for coal petrologists. As the rapid developments of computational processing capacities and algorithms in recent years, deep learning has been experiencing a considerable progress and successfully applied to image analysis. Semantic segmentation algorithms based on deep learning were introduced to make a pixel-level classification of maceral groups in Chinese bituminous coals. A high quality dataset with 739 representative petrographic images was established based on the characteristics of coal macerals in Chinese bituminous coal. The petrographic images were annotated into six classes, i.e., vitrinite, inertinite, liptinite, dark minerals (clay minerals), bright minerals (pyrite), and epoxy (background). The unbalanced proportions of annotated classes were calibrated by mathematical methods. Three classic semantic segmentation models, U-Net, SegNet, and DeepLab V3+, were used to make a pixel-level identification of the coal macerals. Pixel Accuracy (PA), Intersection over Union (IOU) and BFScore were introduced to evaluate the performances of the segmentation. The results have shown that the loss curves and accuracy curves of the three models in the training process rapidly converged after iterations. The training time of U-net, SegNet, and DeepLab V3 + was 172 min, 637 min, and 135 min respectively. The average pixel accuracy of the three models was 73%, 73%, and 92% respectively. The IoU and BFScore of DeepLab V3 + were also higher than those of U-Net and SegNet. The DeepLab V3 + was the preferred method for maceral group identification considering the training efficiency and segmentation accuracy. The maceral group compositions determined by DeepLab V3 + model achieved similar results as the manual point-counting method. The semantic segmentation method based on deep learning meet the accuracy requirements of quantitative analysis of coal maceral groups.