Accurate segmentation of shale-reservoir microstructures is essential for enhancing shale-gas recovery, yet the scarcity of public datasets, the sub-pixel size of pores and their multi-scale complexity severely limit the application of deep learning in this domain. To address these challenges, we first establish a systematic workflow for constructing a shale SEM-image dataset: forty \(2048 \times 1768\) images are meticulously annotated and augmented with shale-specific strategies, yielding 10 800 augmented \(256 \times 256\) patches for training and validation, alongside 80 strictly unaugmented patches for testing, covering organic matter, organic pores and inorganic pores. A selective regulated augmentation scheme is further designed to improve model generalisation on the validation set. We then propose RDAMU-Net, an enhanced U-Net architecture that couples residual dilated modules with an adaptive multi-scale fusion network to enlarge receptive fields and dynamically aggregate cross-scale features. A multi-dimensional attention mechanism embedded in the skip connections captures long- and short-range dependencies in both spatial and channel dimensions, mitigating the loss of micro-pore information during downsampling. Extensive ablation studies and comparative evaluations demonstrate that RDAMU-Net achieves an IoU of \(\sim \) 63% and a Recall of \(\sim \) 78% while maintaining fewer parameters, outperforming several representative state-of-the-art CNN, Transformer and Mamba baselines, especially for extremely small inorganic pores. This work provides a new data foundation and an efficient segmentation framework for intelligent shale-microstructure analysis, advancing digital-rock technologies in unconventional hydrocarbon development. Our code is available at https://github.com/Runner-xc/RDAMU-Net.
