A robust deep learning approach for rock discontinuity identification from large scale 3D point clouds

Rock discontinuities are critical geological features that govern rock mass mechanical behavior. Accurate identification and quantitative characterization of these discontinuities underpin vital applications for slope stability analysis, underground excavation design, and rock blasting. Conventional approaches to large-scale point cloud analysis are often limited by poor fine-scale feature representation and high parameter sensitivity. To address these limitations, this study proposes RL-JointNet, an end-to-end deep learning approach specifically developed for robust discontinuity segmentation. The key novelty of RL-JointNet lies in its enhanced local feature extraction module. This module integrates two main innovations: (1) an explicit relative position encoding to capture fine-scale spatial relationships, and (2) a multi-path feature fusion strategy to better represent complex neighborhood geometries. The model's effectiveness was validated on high-resolution point cloud datasets from two rock slopes. Results demonstrate superior performance, achieving a Global Accuracy (GA) up to 98.7% and a mean Intersection over Union (mIoU) of 98.1%, with recognition accuracy for individual discontinuity classes consistently exceeding 95%. Crucially, a comprehensive hyperparameter sensitivity analysis revealed that RL-JointNet exhibits significantly enhanced robustness compared to conventional point cloud deep learning models, ensuring consistent performance across diverse engineering scenarios. The proposed RL-JointNet offers a reliable approach for automated discontinuity analysis, thereby enhancing the capability for detailed characterization of large-scale and complex rock masses.