Abstract: Sand and gravel mixture are widely used fill material for earth and rock dams, with its mechanical properties significantly influenced by particle geometric characteristics such as gradation, shape, and spatial arrangement. Accurate acquisition these geometric characteristics is crucial for studying mechanical properties of gravel, which is of great significance for design and construction of earth and rock dams. This study proposes a novel CT image segmentation method for gravel based on a deep learning model, integrating CT image three-dimensional reconstruction and topology principles to create a comprehensive method for extracting geometric feature parameters of gravel. A corresponding program is developed to provide algorithmic flow and parameter settings. Results show that this method achieves a segmentation accuracy over 95%, allowing precise extraction of geometric parameters such as center of mass coordinates, grain size, aspect ratio, and sphericity. The study reveals that gravel specimens exhibit a spatial distribution where sand grains settle at the bottom and gravel grains are uniformly distributed. Additionally, the aspect ratio and sphericity display a skewed distribution in the predicted probability densities. This study is expected to provide new technical means for investigating mechanical properties of gravel, thereby offering new insights for optimizing design and construction of earth and rock dams.