Abstract: To tackle the challenge of accurately determining the volumetric gangue content rate under actual stacking conditions of coal-gangue in fully mechanized caving faces, a prediction method based on the DeepLab v3+ model was proposed. A dataset consisting of images depicting coal-gangue accumulation was constructed, and a semi-automatic data labeling method, along with Contrast Limited Adaptive Histogram Equalization (CLAHE), was employed for image preprocessing. The DeepLab v3+ model was utilized for the semantic segmentation of coal-gangue images, which facilitated the calculation of the projected area gangue content rate. A numerical model was established using the PFC^3D numerical simulation software based on the reconstructed three-dimensional coal-gangue block, simulating the top coal drop and the coal transport process via scraper conveyor. The volume of each gangue or coal particle was extracted using the FISH programming language, enabling the calculation of the volumetric gangue content rate of the coal-gangue accumulation. By analyzing the quantitative relationship between the projected area gangue content rate and the volumetric gangue content rate under varying top coal thickness conditions, a predictive model for the volumetric gangue content rate of coal flow was developed. Experimental results indicated that the accuracy, mean pixel accuracy, and mean intersection-over-union (IoU) of the DeepLab v3+ model were 97.68%, 97.72%, and 95.33%, respectively, all surpassing those of classical semantic segmentation models such as FCN8s and PSPNet. This enabled precise and rapid identification of the projected area gangue content rate of coal-gangue accumulations. The coefficient of determination (R²) for the volumetric gangue content rate prediction model was 0.9828, demonstrating robust predictive performance.