Classifying roundness and sphericity of sand particles using CNN regression models to alleviate data imbalance

Abstract

Determining the shape parameters of sand particles helps to understand the geotechnical properties of sand. This study aims to determine the roundness and sphericity of Jumunjin sand utilizing artificial intelligence (AI). A dataset comprising 1000 sand particle images from Jumunjin sand was used for testing. The training set included approximately 28,000 images, created through a combination of synthetic data (5000 images) and additional data augmentation techniques to address data imbalance issues. Unlike traditional methods for determining roundness and sphericity, this research proposes a model that combines a regression model with a convolutional neural network (CNN), using ResNet and DenseNet as the backbone networks. The results, evaluated based on the coefficient of determination (R²) between the predicted values using the DenseNet169 model and the true values, yielded an R² of 0.695 for roundness and 0.979 for sphericity. When classifying based on the Krumbein and Sloss chart using the trained model, the DenseNet169 model demonstrated the highest accuracy (73.6%), precision (77.9%), and recall (77.2%). A comparison between AI predictions and human evaluations revealed considerable variation in human classification, depending on the observers, whereas the AI model consistently exhibited robust performance in determining both roundness and sphericity.