The traditional evaluation concept for the design of earth-rock dams relies on a limited number of test pit results, leading to issues such as insufficient data and poor representativeness. A rapid gradation detection method for dam-building gravel materials is proposed based on the image recognition technology and a deep threshold convolutional model. The gradation distribution characteristics of the dam materials were statistically analyzed by fitting the distribution patterns of the gradation parameters and conducting hypothesis testing. Using the asphalt concrete face rockfill dam project of the KLYML reservoir in Xinjiang as an example, over 35,000 truckloads of gravel material gradation data were analyzed to statistically examine the distribution patterns of gradation characteristic indices. The analysis revealed that among the five gradation characteristic indices of the dam-building gravel materials, the optimal distribution functions for the coefficient of uniformity C_u, the coefficient of curvature C_c, and the characteristic particle size d₁₀ were all normal distributions, while the optimal distribution functions for the characteristic particle sizes d₃₀ and d₆₀ were Rayleigh distributions. The proposed method in this study can serve as a reference for similar earth-rock dam projects in terms of gradation control during the design phase, quality control of dam materials before construction, and quality evaluation during the construction process.