Supervised classification-based framework for rock mass discontinuity identification using point cloud data

Mapping and evaluating rock mass discontinuities using point clouds is a critical task in mining, civil, and geological engineering. Rock discontinuities can significantly impact the integrity, strength, and stability of rock masses. The orientation of these discontinuities is also a key characteristic of the rock mass. Accurate orientation estimation from point clouds enables more precise predictions of rock mass behavior, leading to improved safety, more efficient excavation processes, reduced operational costs, and significant time savings. In this context, a supervised classification-based framework is proposed for calculating orientation parameters from point cloud. Supervised classification plays a crucial role in tasks where a model learns complex patterns from labeled data to accurately predict previously unseen instances. The proposed method consists of eight-steps, including: data collection, pre-processing (data filtering), adaptive neighborhood size selection (omnivariance-based), feature extraction (geometric features), feature selection (Minimum Redundancy Maximum Relevance method), classification (Support Vector Machine), clustering (connected component labeling), and plane fitting to calculate orientation parameters (dip angle and dip direction). The framework was applied to two real-world datasets and one synthetic dataset, which was tested in two different subsampled forms (random and uniform subsampling). The results statistically demonstrated that the technique was effective in detecting and characterizing rock mass discontinuities with high Accuracy, Recall, Precision, and F-Score values ranging from 94.64% to 99.57%. The deviations of the method in the measurements of the dip angle and the dip direction, compared to the manual measurements, range from 1% to 4%, indicating strong agreement with the manual measurements and the existing studies.