Microscopic pore combination type identification of tight sandstone reservoir based on improved swin transformer architecture

Tight sandstone reservoirs differ fundamentally from conventional medium to high permeability reservoirs due to their complex and heterogeneous microscopic pore structures. This complexity poses significant challenges for accurate reservoir characterization and often results in suboptimal development performance. The specific configuration of microporosity combinations plays a decisive role in determining the storage and seepage capacities of tight sandstone reservoirs. Therefore, the precise identification of microporosity combination types is essential for improving both reservoir evaluation accuracy and development effectiveness. However, traditional computer vision models exhibit limitations in capturing fine-grained textures and spatial relationships among microscopic pores with complex morphologies, leading to inadequate generalization capabilities. To address these issues, this study proposes an enhanced Swin Transformer-based neural network architecture, termed SwinLSC (Swin Transformer with Linformer and Self-Adaptive Channel Attention). The model incorporates a global-local attention mechanism and is trained on image datasets of cast thin sections from tight sandstone reservoirs in the Yanchang Oilfield. To evaluate model performance, Top-1 Accuracy , Loss , and Recall metrics were employed, and the SwinLSC model was benchmarked against three mainstream architectures: Swin Transformer, Vision Transformer (ViT), and ResNet. Experimental results demonstrate that SwinLSC achieves a prediction accuracy of 93.3 %, significantly outperforming the comparative models. These findings indicate that the SwinLSC model effectively addresses the generalization deficiencies of conventional approaches in recognizing microstructural features in cast thin section imagery. Consequently, it offers a robust and accurate solution for microporosity type identification, thereby providing reliable technical support for the efficient exploration and development of tight sandstone reservoirs. • SwinLSC model achieves 93.3 % accuracy, surpassing Swin Transformer, VIT and ResNet in microporosity prediction. • Global-local attention mechanism in SwinLSC enhances feature extraction and generalization. • Improved microporosity identification enables efficient exploration and development.