SA-TransU²Net: a rock thin section grain segmentation network based on multi-scale RSU and global context enhancement
Yaohua Gong1, Di Shi1, Ling Zhao2
1School of Computer & Information Technology, Northeast Petroleum University, Daqing, 163318, China.
Scientific Reports
|July 14, 2026
Summary
This study introduces SA-TransU²Net, a novel hybrid network for segmenting rock thin sections. It improves accuracy in identifying geological features for oil and gas exploration by better handling grain boundaries and adhesion.
Area of Science:
- Geoscience and Artificial Intelligence
- Petroleum Geology and Image Analysis
Background:
- Rock thin section segmentation is crucial for reservoir characterization in oil and gas exploration.
- Current deep learning methods struggle with accurate grain boundary delineation and adhesion due to scale variations and low contrast.
Purpose of the Study:
- To develop an advanced deep learning model for precise rock thin section grain segmentation.
- To overcome limitations in existing methods for complex sandstone microstructures.
Main Methods:
- Proposes SA-TransU²Net, a hybrid network combining U²-Net with Swin-Transformer attention modules.
- Integrates a parameter-free spatial attention mechanism to enhance boundary and structure feature responsiveness.
Main Results:
- SA-TransU²Net achieved high performance on an Ordos Basin sandstone dataset with mIoU of 87.68%, Precision 91.73%, and Recall 92.66%.
- The model demonstrated superior segmentation accuracy and analysis efficiency compared to existing mainstream models.
Conclusions:
- SA-TransU²Net offers an efficient and reliable solution for intelligent rock thin section analysis.
- The proposed method effectively improves segmentation accuracy in challenging geological samples.
