<p>Thin section identification is a fundamental task in petroleum geology research for understanding reservoir characteristics and obtaining parameters related to oil and gas reservoirs, playing a significant role in the exploration and development of hydrocarbon resources. However, the multi-source and multi-modal characteristics of mineralization information, visual observation errors, and the subjective experience of experts can introduce certain biases into thin section identification results. This study employs the PyCharm platform as the deep learning framework and selects suitable microscopic images with distinct features from three publicly available thin section image datasets on the China Scientific Data website. Using data augmentation techniques such as random flipping, zooming, shifting, cropping, and scaling, a dataset of 4,400 thin section images was constructed, including dolomite, siliceous rock, gneiss, schist, and sandstone. A classification model named SENet-ConvNeXt-FL was developed by integrating the SENet attention mechanism into the ConvNeXt network and introducing the Focal Loss function. This approach enhances the model's feature representation capability and effectively addresses class imbalance issues. The results demonstrate that the SENet-ConvNeXt-FL model achieves an overall accuracy of 92.5% in thin section identification, exhibiting strong robustness and generalization ability. Therefore, further refinement and application of deep learning-based thin section image classification models hold great promise for providing critical support in the accurate evaluation of hydrocarbon reservoirs.</p>