An efficient classification system for excavated soils using soil image deep learning and TDR cone penetration test

Abstract

Soil classification plays a significant role in reutilization of excavated soils, which is produced greatly every year in China. In this work, an efficient system for identifying excavated soil type was developed at the start, transfer or end points of transportation. Firstly, soil image color patterns, cone index (CI), dielectric constant (DC), and electrical conductivity (EC) were identified as indexes for prompt characterization. Accordingly, an excavated soil information collecting system (ESICS) based on time domain reflectometry (TDR) cone penetrometer and digital camera was established at Xiecun Wharf, the largest wharf for transferring excavated soils in East China. After collection of soil information for 2 months, a multi-source soil database with 25,152 groups of soil image, CI, DC, and EC was generated. Then, based on ResNet18 convolutional neural networks, a novel classification framework with four screens (soil images, CI, DC, and EC) was proposed. Through deep learning of the database, all the excavated soils were finely classified into 12 types, which was calibrated by laboratory tests in Unified Soil Classification System and soil mineralogy. The system can realize classification with 88.7 %-accuracy within 50 s (even 97 % for the soils with simple color patterns), which leads to cost-effective management of excavated soils.