Planar-Constrained 3D Instance Segmentation of Rock Discontinuities
Year: 2026
Type: conference-paper
Abstract: The accurate identification and characterization of rock discontinuities are fundamental to geotechnical engineering safety assessment. While deep learning-based instance segmentation models have shown promise in point cloud analysis, generic models often fail to produce geologically plausible results for rock masses due to a lack of domain-specific knowledge. This paper introduces GeoMask3D, an innovative instance segmentation model that explicitly incorporates geological priors into its learning process. The core of GeoMask3D is a novel planarity-constrained loss function, which leverages the inherent planar geometry of rock discontinuities as a constraint to guide model optimization. This constraint enforces the flatness of segmented surfaces, aligning the model’s output with geological reality. Experiments conducted on a custom-built dataset of a large-scale open-pit mine demonstrate the superiority of our approach. GeoMask3D achieves a mean Average Precision (mAP) of 57.5%, a significant improvement of 4.7% over the baseline model. Qualitatively, it produces remarkably regular and smooth discontinuity surfaces. Furthermore, we establish an automated pipeline from raw point clouds to key geotechnical parameters (e.g., Jv, Kv), with results showing high consistency with manual field measurements (average orientation error < 5°). This work not only provides a high-performance segmentation tool but also validates the powerful “data-driven learning guided by physical knowledge” paradigm, offering robust technical support for applications like digital twin rock masses and geological hazard early warning.
Authors: Guanghao Wang, Zhenhao Gu, Lijun Fu, Duan Jian
Topic: Rock Mechanics and Modeling
Subfield: Mechanics of Materials
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7164670873