Image-based prediction of particle size distribution in recycled stabilized soils using convolutional neural networks

This study presents a convolutional neural network (CNN)-based method for estimating the particle size distribution (PSD) of recycled stabilized soils composed of Ao clay, a paper–sludge–ash-based stabilizer, and blast furnace slag cement type B. The proposed method addresses challenges in conventional PSD analysis, including the time-consuming procedures required to assess fine particles within moisture-rich, recycled soils. A comprehensive image database was developed using 1,080 multi-sized fraction images, 4,000 merged sample images, and 10 single-sized fraction images paired with PSDs obtained via sieve analysis. A lightweight CNN model was created to predict PSDs from low-resolution (112 × 112-pixel) soil images. To ensure physical consistency, the model outputs retained the mass percentages for standardized sieve sizes, which were subsequently post-processed into cumulative PSD curves. Unlike traditional image-analysis techniques, CNN-based methods can effectively handle the visual variability caused by moisture, intraparticle porosity, and irregular particle shapes. The trained model achieved high accuracy, with 95% of test images yielding coefficients of determination (R²) above 90% and an average mean absolute error of 0.0158. Accurate predictions were obtained even for fine particles smaller than the pixel resolution, indicating that the CNN learned to interpret key textures and brightness patterns beyond particle outlines. In addition to full PSD curves, the CNN accurately estimated key classification parameters, including fines, sand, and gravel content, with R²* values exceeding 91%. This image-based, nondestructive technique offers a practical approach for rapid quality control during the inspection of construction-sludge recycling facilities.