An improved U-net model for the identification of coal macerals

Deep learning has shown strong potential in automating coal maceral segmentation, and an accurate identification of complex lathy (elongated) textures, such as cutinite, however, remains challenging. To address this, this study proposes U-maceralnet, an enhanced U-Net-based semantic segmentation model that integrates external attention and dynamic snake convolution modules to better capture fine, elongated maceral textures. A high-quality dataset of 561 coal maceral images with pixel-level annotations was established, with annotations performed by experienced coal petrographers. U-maceralnet was trained with a tailored loss function and evaluated on the test set. Experimental results show that the model achieves superior segmentation accuracy, especially for macerals with lathy textures, with average performance metrics of 84.0% for pixel accuracy, 75.2% for intersection over union (IoU), 87.5% for precision, and 89.7% for recall. Compared to popular methods such as DeepLabv3+, FCN, PSPNet, SegNet, and U-Net, U-maceralnet consistently outperforms in handling structural complexity. Additionally, it provides more accurate estimations of liptinite content than the conventional 400-point counting method, offering improved efficiency and reliability in coal petrographic analysis.