Convolutional neural network prediction of the particle size distribution of soil from close-range images

Abstract

Accurate soil particle size distributions are essential for various geotechnical applications. In this study, we propose a convolutional neural network approach for predicting the particle size distribution using soil image analysis. Our model is trained on a diverse dataset of soil samples ranging from clayey silt to gravel. We employed transfer learning by using MobileNet pre-trained on ImageNet and adding additional layers to fine-tune the model for our specific task. The soil images were captured under standardised lab conditions using a dark chamber with constant lighting to ensure consistency. We implemented the model in Python and explored various neural network architectures, image resolutions and data augmentation techniques to optimise performance. The model predicts the particle size distribution through two parameters derived from the Weibull distribution. Our approach offers instantaneous predictions and demonstrates robustness across a wide range of soil types. We outperform previous studies by incorporating geotechnical classification and predicting the entire particle size distribution curve. Additionally, we applied explainable artificial intelligence techniques to enhance the transparency and interpretability of the model’s predictions. Our findings highlight the effectiveness of the model and provide valuable insights into the relationship between soil image features and particle size characteristics.