Machine Learning for Automatic Identification of Pores in Microscopic Images of Weathered Sandstone from Yungang Grottoes

Pores, as one of the channels for water, air, and microorganisms to enter the rock, accelerate the rock weathering process and change the physical and mechanical properties of the rock. Therefore, the study of the microscopic pore structure is of great significance for stone cultural heritage. This paper proposes a multi-scale pore structure characterization method based on backscattered scanning electron microscope (BSE) images, integrating feature engineering and machine learning techniques. First, for each pixel, 16 features are generated to construct a feature engineering. Then, these features are input into four commonly used machine learning models (Random Forest, Support Vector Machine, Multi-Layer Perceptron, Gradient Boosting Decision Tree) to distinguish pore and mineral matrix. To verify the effectiveness of the multi-channel feature input method, we compared the confusion matrix parameters (accuracy, precision, recall, and F1-score) of the four models before and after adding the feature engineering, and found that the indicators of all models increased by 1%–15%. In addition, the study also found that the Random Forest model performed the best. It can effectively segment the pores in different new images, thus could be directly applied to weathered sandstone images, providing data support for the scientific protection of stone cultural heritages.