Hyperbolic regularization-PointNet++: Automated three-dimensional soil particle roundness classification

Abstract

Particle roundness significantly affects the macro-mechanics of granular soils. Nevertheless, current three-dimensional(3D) computational geometry (CG) -based methods for soil particle roundness classification exhibit inefficiency and struggle with defective particles. To overcome these limitations, this paper proposed a deep learning-based approach for intelligent soil particle classification to automatically assess roundness. We labeled 2400 soil particle point clouds into six classes: well rounded, rounded, subrounded, subangular, angular, and very angular. Using the labeled dataset, we trained the PointNet++ model with a novel hyperbolic regularizer, extracted hierarchical roundness features, and built a roundness classifier. The achieved automatic classification accuracy is 92.19%. Moreover, the classification results are similar to the 3D CG-Manual Charting method in terms of quantity, volume, surface area, and convex hull volume. Notably, our method overcomes the limitation of 3D CG's inability to classify soil particles with void defects. Moreover, it shows a classification speed of approximately 77 times faster than 3D CG.