Study on the automated characterization of particle size and shape of stacked gravelly soils via deep learning

Abstract

The mechanical behavior of gravelly soils is influenced by their particle size and shape. Traditionally, particle size and shape can be determined through sieve analysis, the Krumbein & Sloss chart, or mathematical calculations based on discrete particle images. However, these methods are challenging to apply to the scenarios where particles are difficult to collect, such as when particles are oversized or located in remote areas. This study explores the feasibility of quantifying the size and shape of on-site stacked gravelly soils using a deep learning model. The deep learning model SOLOv2, based on convolutional neural network, is utilized as the foundational network framework for identifying particles in stacked particle images. The model’s performance is enhanced using modulated deformable convolution networks. To explore the impact of particle overlap in stacked particle images on quantification results, two recognition modes are considered: one that recognizes only nonoverlapping particles and another that recognizes all particles. The recognized particles are used to quantify corresponding size and shape parameters through mathematical calculations, such as Feret diameter, equivalent area diameter, roundness and sphericity. In addition, in order to verify the reliability of the quantitative results, this study conducts a systematic comparison with traditional methods. The comparative results indicate that, for particle size quantification, the trained model yields reliable measurements regardless of the recognition mode employed. However, for particle shape quantification, only the mode that recognizes nonoverlapping particles produces accurate shape measurement results. Consequently, the proposed trained model can accurately quantify the size and shape parameters of gravelly soils based solely on on-site stacked particle images, without the need for particle collection and separation.