The Multi-Level Classification Strategy for Rock Surface 3D Point Clouds

Abstract

This study presents an integrated computational framework for accurate rock discontinuity characterization in complex geological environments, addressing the limitations of traditional field measurements and existing automated approaches through the synergistic combination of UAV-based remote sensing and advanced machine learning techniques. The methodology establishes a three-stage analytical pipeline beginning with an optimized Random Forest algorithm for robust initial classification of discontinuity features in 3D point cloud data, followed by Mean Shift clustering to systematically group discontinuities by principal orientations, and concluding with DBSCAN-based refinement for precise boundary delineation. Field validation demonstrates the framework's effectiveness in overcoming environmental noise and surface irregularity challenges, with the Mean Shift clustering component particularly excelling in maintaining geometric fidelity for complex curved or rough discontinuity surfaces. The approach shows consistent performance advantages over conventional methods in both detection accuracy and computational stability, offering practical improvements for geological surveys by enhancing measurement reliability while reducing field workload, with broad applications in slope stability assessment, underground excavation design, and rock mass quality evaluation.