Use of machine learning for classification of sand particles

Particle classification is essential for geotechnical engineering practice since particle shapes correlate with the mechanical and hydraulic properties of sand layers. Traditional shape classification is tedious, subjective, and time-consuming because it depends on manual visual comparison with reference particles. This study demonstrates the feasibility of employing machine learning algorithms for sand classification. Machine learning (ML) models are increasingly being introduced for automatic identification and classification of various objects. Nine types of sand were selected, and the analysis was based on 2000 binary images of each sand that were obtained from dynamic image analysis (DIA). Each particle was represented by six engineering size and four shape descriptors. The efficacy of seven ML models for automatically classifying individual sand particles was explored. The study demonstrates that the size and shape descriptors are efficient and robust to identify up to 75% of sand particles, using a neural network classifier. In addition, use of scale-invariant feature transform (SIFT) features was also explored to permit future generalization of sand classification using image datasets containing images with different scales and resolutions. Adding SIFT to size and shape can increase classification accuracy to 83% using a random forest classifier. The analysis also reveals that histograms of orientation gradients of SIFT keypoints in sand appear well correlated with sphericity and convexity of particles. This study suggests that a dataset of 2000 particles per sand is sufficient for optimal classification performance and that image preprocessing of DIA images was not necessary.