Application of Dynamic Image Analysis to Sand Particle Classification Using Deep Learning
Nikolaos Machairas, Linzhu Li, Magued Iskander
Abstract
Soil particle size and shape are of great interest to the geotechnical engineering community because they affect soil behavior. Determination of soil type is an important requirement in geotechnical engineering projects such as developing a site-specific soil profile or performing a geotechnical design. Geologic formation of sand usually results in sand particles with distinct visual characteristics such as size, color, and shape. In this study, dynamic image analysis (DIA) is employed to extract particle size and shape descriptors which were then used for the classification of sand. Five different types of siliceous sand with different particle shapes were selected for investigation. A training dataset of grading and shape properties of these sands was compiled from over 50,000 images. This was accomplished using machine learning models based on convolutional neural networks. This method allows for automatic sand/particle classification which may eventually assist engineers on-site to quickly determine geotechnical properties of soil formations that would normally be analyzed in laboratories.