Hybrid Swin Transformer-CNN Model for Pore–Crack Structure Identification

Accurate classification and characterization of pore–crack structures are substantial to carbonate reservoirs in terms of reservoir exploration and development. Although experience-dominated manually classifying pore–crack structures achieves a milestone, these methods usually encounter significant uncertainties and heavily rely on the interpreter’s experience. Nevertheless, as a classification problem, using the 2-D image input dataset, instead of 1-D logging data, could achieve a higher accuracy. Consequently, we developed a Swin Transformer-convolutional neural network (SWT-CNN) hybrid network to capture multilevel features of the pore–crack structure simultaneously using 2-D resistivity imaging logging (RIL) images as an input, thereby eliminating the uncertainty of manual interpretation and enabling automatic feature extraction. Furthermore, to fully utilize rare and valuable dataset, the proposed SWT-CNN model incorporates the data augmentation strategy which has been modified to fit the dataset. In addition, the idea of transfer learning is introduced to improve the accuracy of pore–crack types classification in carbonate rock and accelerate convergence. Finally, the field validation data test shows that the proposed SWT-CNN can achieve an accuracy rate of 95.92%. Moreover, the visualization of the feature map indicates the proposed SWT-CNN is more accurate in recognizing the position of pore–crack structures.