A framework for automatic discontinuity trace extraction using multi-scale surface variation index and transfer-learning enhanced artificial neural network

Discontinuity traces significantly impact the mechanical properties of rock masses, making their rapid and accurate identification crucial for stability analysis. We propose a framework using the multi-scale surface variation index (MsSVI) and transfer-learning enhanced artificial neural network (ANN) for efficient discontinuity trace extraction from rock mass point clouds. Leveraging the similarity between regular geometric bodies and engineering rock masses, we extract trace feature points without manual threshold selection. Our contributions include: (1) An adaptive radius MsSVI calculation method based on density information; (2) a universal trace feature point classification model trained using MsSVI and ANN via inductive transfer learning; and (3) a random sampling L1-medial skeleton algorithm for precise trace feature point extraction, bypassing point cloud triangulation. Experimental results show that our model achieves a 90.2% F1-score on test sets, demonstrating its accuracy and robustness. Furthermore, our method excels in trace detail extraction on two datasets, surpassing existing models and highlighting its potential for rock mass structural analysis.