A Semi-automatic Method to Recognize Discontinuity Trace in 3D Point Clouds Based on Stacking Learning

Discontinuity trace is an important parameter to characterize the quality of rock mass. To solve the problems of sample imbalance, feature unnoticeability, and poor recognition performance of a single model in discontinuity trace recognition using machine learning, a new method is proposed in this paper. The new method includes using the SMOTE oversampling technique to increase the sample size of discontinuity traces, selecting normal vectors and three curvatures as input features for machine learning, and using the Stacking method to ensemble three base models. By analyzing the real slope point cloud acquired by the Trimble X7 scanner, it is proved that the proposed method can effectively recognize discontinuity traces. The SMOTE technique in the method can improve the performance of the model by balancing the number of samples of the two classes, and the selected curvature features have a strong correlation with the discontinuity traces, so the selected features are correct. Finally, it is proved that the Stacking model achieves the best recognition performance and generalization by combining the three base models.