Rapid rock mass classification framework using machine learning and semantic segmentation for underground mining

Abstract

Conventional rock mass classification still heavily relies on manual field investigations and laboratory testing. Despite advancements in Artificial Intelligence (AI)-enhanced methods, integrating machine learning for reliable classification remains challenging. This study proposes an optimized AI-powered rock mass classification framework. First, three types of machine learning models – tree-based models, a stacking ensemble model, and a Graph Neural Network (GNN) – were employed to predict uniaxial compressive strength (UCS) based on field data. Next, three semantic segmentation architectures – U-shaped Network (U-Net), Fully Convolutional Network-8s (FCN-8s), and Segmentation Network (SegNet) – were comparatively employed to detect joint fractures from field-captured rock surface images. Then, a multi-stage clustering-fitting algorithm was developed to extract key joint parameters from the segmented images. Finally, The framework was applied to a real mining case for ore rock mass classification. Results show that the GNN exhibits superior capability in modeling nonlinear parameter interactions for uniaxial compression strength prediction, while U-Net delivers the highest segmentation accuracy for joint detection under complex geological conditions. The AI-powered method achieves higher classification accuracy in ore body zones compared to surrounding rock and aligns well with the Rock Mass Rating (RMR) system. By automating parameter extraction and integrating rapid field testing, the framework reduces reliance on manual work and expensive equipment, advancing automation and reliability in geological applications.