Construction of AI-based Dataset for Rock Type Classification and Joint Detection Using Bedrock Drill Core Images

This study aims to address frequent ground subsidence issues in urban areas by constructing an artificial intelligence (AI)-based dataset for rock type classification and joint detection using bedrock drill core images. In Korea, current ground assessments heavily rely on subjective and qualitative interpretations, lacking accuracy and objectivity. To overcome this limitation, high-resolution images of bedrock drill core samples were collected and meticulously annotated, resulting in a comprehensive dataset comprising 661,425 images. The dataset includes 550,080 images for rock type classification and 111,345 images for joint detection, with labeling data provided in JSON format. The rock type classification dataset is categorized into igneous, metamorphic, and sedimentary rocks, while the joint detection dataset employs a polygon segmentation method. The efficacy of the dataset was validated by applying ResNet152 V2 for rock classification and Deeplab V3+ for joint detection models, achieving a Top-1 Accuracy of 90.25% and an intersection over union (IoU) of 84.04%, respectively. The constructed dataset is publicly available via AI-Hub and is expected to significantly contribute to the development of quantitative and rapid bedrock evaluation technologies in research and industry.