Intelligent recognition of weak discontinuities on outcrops of hard rock masses

Rock mass discontinuities arise from tectonic movements and other geological processes, reflecting the evolution of the Earth’s crust. These discontinuities significantly influence the physical properties, deformation characteristics, and energy release mechanisms of the crust. Therefore, recognizing discontinuities is crucial for understanding the evolution of geological structures, analyzing the physical and mechanical properties of geological bodies, and investigating geological hazards. Traditionally, discontinuity recognition has relied on manual interpretation or automated algorithms based on pixel brightness. However, these methods often struggle to strike a balance between efficiency and robustness. To overcome these limitations, we leveraged deep learning techniques that integrate the strengths of both approaches, enabling the recognition of automated discontinuity with expert-level accuracy. To accomplish this objective, we developed and open-sourced the first large-scale deep learning database for rock mass discontinuities, featuring over 300,000 annotated discontinuities. The YOLOv8x-seg model was extensively trained on this database and evaluated across diverse and complex scenarios. The results demonstrated the model’s capability to accurately recognize discontinuities even under challenging conditions. Furthermore, we expanded the test set to include rock masses from various global locations, as well as underground rock masses, soils, and artificial structures, where the model consistently achieved effective recognition. The model consistently delivered accurate results, highlighting its strong generalization capability. A comparative analysis revealed that its performance closely aligns with expert manual interpretations. Our open-source database enables researchers to train various deep learning models and achieve equally high-performance results.