Abstract:
Coarse-grained soil is widely used in embankments, earth-rock dams, and other fill engineering. However, the traditional sieving method is time-consuming and inefficient, failing to meet the rapid quality testing requirements for gradation. To address these issues, this paper constructs an "image-gradation" relational database for loess and quartz sand coarse-grained soil, comprising 22,080 photos. In response to the mismatch between two-dimensional images and three-dimensional gradation, a Searcher-Analyzer Network (SaNet) is developed to handle any number of image inputs. The model's accuracy steadily improves with an increase in the number of images, with average errors of 1.63% and 1.21% for the identification of loess and quartz sand gradations, and goodness of fit values of 0.995 and 0.992, respectively. The results demonstrate that the machine learning model built on the SaNet architecture exhibits high accuracy in gradation identification, meeting the real-time non-destructive gradation detection requirements in fill engineering.