Predicting the particle size distribution of fine-grained and sandy soils using deep learning for classifying recovered soils separated from tsunami deposits

Disaster wastes, particularly those generated in the tsunami, comprise soil and sediments, and recovered soil is yielded through treatment such as separation from disaster waste and tsunami deposits. The sieving method is used to determine the particle size distribution of recovered soil before reuse. This study uses a convolutional neural network (CNN) model to predict the particle size distribution of fine-grained and sandy soils. The VGG-16 model was modified for use with the CNN model. Soil with a particle size range of < 4.75 mm was size-fractionated and used as training data. The photo image of the size-fractionated soil was divided and merged to prepare the training data as data augmentation. In the model without data augmentation using photo image merging, in some cases, the particle size distribution curves overlapped between fine-grained and sandy soils. In the models with data augmentation, the predicted particle size distribution curves did not overlap in both fine-grained and sandy soil samples. The model with data augmentation manifested root mean square error (RMSE) lower than that of the model without it. This study shows that the particle size distribution of fine-grained and sandy soils could be predicted with the RMSE of < 0.11 using photo images.