In order to improve the prediction accuracy of lithology affected by imbalanced geological data, an ECA-MSCB ResNet model was proposed. The model integrates ECA (efficient channel attention) and MSCB (multi-scale convolutional block) into the traditional ResNet architecture to achieve efficient extraction and representation of lithological data features. For the issue of imbalanced lithology categories, prior probability-balanced logit bias was introduced during model training, and the focal loss function was modified to enhance the recognition of minority lithology classes. Experimental results show that the model based on ECA-MSCB ResNet performs well on the imbalanced geological lithology dataset, achieving an average prediction accuracy improvement of approximately 7.45% compared to the original ResNet model and 27.33% compared to the random forest method. Notably, the recognition of minority lithology classes improves by an average of 17.9%. Furthermore, the model demonstrates strong lithology classification ability on public datasets, achieving an F₁-score of 75.77%. In addition, the recognition accuracy of the proposed model outperformed both traditional and mainstream methods. The ECA-MSCB ResNet method holds significant application value in the field of imbalanced geological lithology recognition.