Figure 1 RQD measurement using automatic crack detection on an uploaded image. 3. CASE STUDY 3.1 Data preparation The classification task was performed on tuff, mudstone, and other rocks obtained from advanced borehole cores in tunnelling. In this study, the construction of the labelled dataset required for machine learning was performed by the learning module of the developed application. As shown in Figure 2, the user selects the label name of interest and specifies the Region of Interest (ROI) for the corresponding region. This efficient method allows the user to construct training data. In this experiment, three types of rocks (designated Rocks A, B, and C) were designated as classification classes, and a total of 351,006 pixels of hyperspectral data were acquired as training data. The dataset was then divided into three parts: 80% was designated as the training set, 10% as the validation set, and 10% as the test set, for the purpose of evaluating the model. Figure 2 ROI selection interface for labelling hyperspectral training data 3.2 Result of segmentation In this experiment, the three rock types were assigned labels, and machine learning models were trained. The dataset employed for classification was the hyperspectral dataset delineated in Section 3.1. Hyperspectral images obtained from actual borehole core samples were utilized for training, and the classification model was constructed based on the spectral information obtained. For the hyperspectral images that were not utilized in the training process, the previously constructed model was applied to determine the rock type. The outcomes of this analysis are presented in Figure 3(a). To evaluate the accuracy of the application, a comparison was made between the results of the rock types determined by the naked eye and the results of the machine learning model constructed using a Python script (Figure 3(b)). The correct answer data (labels based on ocular appraisal) used for the comparison is shown in Figure 3(c), and it was confirmed that the results
RkJQdWJsaXNoZXIy MTM0Mzk2