Recognizing gradations of coarse soils based on big artificial samples and deep learning

Abstract

In earth-rockfill dams, roadbeds, airports, and other embankment projects, gradation information serves as the basis for evaluating the quality and suitability of fill materials. Addressing the limitations of existing image-based contour recognition methods and machine learning approaches in recognizing small particle size ranges, this study establishes the first publicly available coarse-grained soil database including Yellow River Silt and Quartz Sand datasets, with particle sizes ranging from 0.075 to 20 mm, comprising a total of 22,380 images. Subsequently, a novel Convolutional Neural Network (CNN) architecture, the Searcher-Analyzer Network (SaNet), based on the Deep Residual Network (ResNet), was proposed to enhance the accuracy of gradation recognition by taking multiple images under a single gradation as input. Finally, the interpretability of the model was discussed through feature map visualization. The results demonstrate that SaNet achieves $\overline{\mathrm{MAE}}$ of 1.63 × 10⁻² and $\overline{R^2}$ of 0.995 for Yellow River Silt, and $\overline{\mathrm{MAE}}$ of 1.21 × 10⁻² and $\overline{R^2}$ of 0.992 for Quartz Sand. Concurrently, the additional computational time and storage requirements are only 3.5 % and 0.3 % more than those of ResNet, allowing the recognition of a single image to be completed within 10 ms. The findings of this study indicate that the proposed SaNet model can instantly achieve high accuracy in gradation recognition, meeting the demands for real-time, non-destructive gradation testing in related tasks.