<p>Rock thin section segmentation serves as a critical technical pillar for reservoir characterization in oil and gas exploration. It enables the effective identification of reservoir lithological characteristics, determination of diagenetic types, and quantification of pore structures, thereby providing essential microscopic evidence for reservoir quality assessment, oil-bearing prediction, and exploration potential evaluation. However, existing deep learning methods still exhibit limitations in accurately delineating grain boundaries and handling grain adhesion. These challenges primarily stem from significant variations in sandstone grain scales, low boundary contrast, and complex cementation or adhesion effects.To address these challenges, this paper proposes a hybrid network model designed for rock thin section grain segmentation, named SA-TransU²Net. By integrating a U²-Net-based backbone with Swin-Transformer attention modules, the proposed model enhances global context representation capabilities and effectively resolves the long-range dependency problem. Furthermore, the model incorporates a parameter-free spatial attention mechanism to intensify feature responsiveness towards grain boundaries and salient structures, achieving robust segmentation in complex, adhered regions.Experimental results demonstrate that SA-TransU²Net exhibits superior performance on a sandstone thin section dataset from the Ordos Basin, achieving a Mean Intersection over Union (mIoU) of 87.68%, a Precision of 91.73%, and a Recall of 92.66%. Compared with existing mainstream models, the proposed method effectively improves both segmentation accuracy and analysis efficiency, offering an efficient and reliable solution for intelligent rock thin section analysis. Experimental results demonstrate that SA-TransU²Net achieves competitive and robust performance on the Ordos Basin sandstone thin-section dataset.</p>