Fig 1.
Flowchart of the overall procedure of the proposed method.
Fig 2.
Demonstrate the automated processing of core images.
Drill core images and lithology logs obtained from the geological drill holes were first uploaded to the geological database. The core images were gained from the database, and the cropping program made a judgment on the image and cropped it by center cropping. In total, 10 lithology categories were obtained, namely: (a) Diabase; (b) Diorite; (c) Gneiss; (d) Granite; (e) Limestone; (f) Marble; (g) Monzonite; (h) Mudstone; (i) Shale; and (j) Siltstone.
Fig 3.
Example of non-core sections images.
The parts were easy to cause misclassification by the CNN model, such as red marks and crushing structures.
Fig 4.
Ten typical samples in each label.
The non-core itself was discarded and the core images were annotated according to the geologist’s lithology logs.
Fig 5.
The structure of ResNeSt block.
c,h,w are the number of channels, height, and width of the input featuremap, respectively. Cardinal k is the k-th cardinal group and Split r is the r-th split.
Fig 6.
Diagram of learning rate decay.
The maximum learning rate is 4e−5, the minimum learning rate is 1e−7, and the total number of epochs is 100.
Fig 7.
Ffc represents full connection. The weight parameters of the output layer were initialized by Xavier random initialization.
Fig 8.
The effect of whether to use data augmentation on model training.
Data augmentation can effectively alleviate overfitting and improve the generalization of the model.
Table 1.
Compare the experimental results of the model at three batch sizes.
Fig 9.
Total loss of four different CNN architectures during training (a) and validation (b).
Fig 10.
Accuracy of four different CNN architectures during training (a) and validation (b).
Fig 11.
Performance of various lithology categories evaluated by (a) accuracy, (b) precision, (c) recall, and (d) F1−score, respectively. The subplots of each figure are the numerical distribution of each evaluation metric on each label, and the number next to the boxes represent the mean values.
Fig 12.
Confusion matrix of ResNeSt-50 on testing group.
Fig 13.
Images of mudstone misclassified as siltstone.
Fig 14.
Comparison of Params, FLOPs of CNN models.
Table 2.
Some other evaluation metrics were compared in the training process, testing process, and the model itself.
Fig 15.
Grad-CAM visualization of DenseNet-161 making correct predictions on testing images.