Lightweight Model <scp>MCCDSNet</scp> : A Novel Approach for <scp>SEM</scp> Image Segmentation of Deep Shale

ABSTRACT Efficient and precise quantitative characterisation of deep shale reservoirs represents one of the key research directions in contemporary studies. To address the issues of low segmentation accuracy, limited segmentation categories, and imprecise boundary segmentation in SEM images of deep shale, an enhanced DeepLabV3+ model named MCCDSNet was proposed for characterising SEM images of deep shale. MCCDSNet employs MobileNetV4HybridLarge as the backbone network, which ensures high processing speed whilst maintaining robust high‐precision feature extraction capabilities. Additionally, the feature representation capability is enhanced by integrating features from adjacent levels. Subsequently, the attention mechanism for channel and spatial information is utilised to improve the accuracy of target segmentation. Lastly, a parallel structure comprising multiple dilated convolution hierarchies and strip pooling is adopted to effectively capture multi‐scale contextual information. In the dataset prepared from the deep shale in the Luzhou Block, MCCDSNet achieved a Mean Intersection over Union (MIoU) value of 0.692 in the segmentation of six categories: inorganic mineral, inorganic pores, organic matter, organic pores, fractures, and pyrite. The results outperform those of other mainstream semantic segmentation networks, thereby verifying the effectiveness of MCCDSNet in the characterisation of scanning electron microscope images of deep shale. When MCCDSNet was applied to Well L6, it was found that organic matter pores are the dominant contributors to the reservoir space of the deep shale in the Luzhou Block. Specifically, in the Longmaxi Formation, the inorganic pores, organic matter, organic pores, pyrite, and fractures are relatively well‐developed within the 3rd sub‐layer and at the bottom of the 4th sub‐layer.