Abstract: To address the limitation of ore visual analysis confined to 2D images, which results in the loss of three-dimensional color information and textural features, this paper proposes a high-precision three-dimensional color reconstruction method for ores based on 3D Gaussian Splatting （3DGS）. An ore image sequence is acquired using a multi-view image capture system. Sparse point clouds, generated by COLMAP, are used to initialize the 3D Gaussian model. A hierarchical optimization strategy is designed: firstly, the 3DGS algorithm is employed to achieve multi-scale geometric modeling and real-time rendering of the ore. During the 3D Gaussian modeling process, Gaussian parameters are jointly optimized within a differentiable rendering framework using a 2D image loss function. This is combined with an adaptive density control strategy to dynamically adjust the Gaussian distribution density. The iteration count is fixed at 30 000 to obtain the final high-quality 2D rendered images. Subsequently, the renderer is extended to generate depth maps. Multi-view depth information is fused based on Truncated Signed Distance Function （TSDF） to enhance surface continuity. The Open3D library is then utilized to extract a high-precision mesh model through voxelization and the Marching Cubes algorithm. To validate the effectiveness, 3-6 cm phosphate ores are used as the research subject. A systematic analysis is conducted on the impact of image quantity on reconstruction quality, and comparative experiments are performed against Multi-View Stereo （MVS） and Poisson reconstruction methods. The study demonstrates that reconstruction quality stabilizes when the input image count reaches 35, yielding high-quality results. Compared to using 15 input images, the Peak Signal-to-Noise Ratio （PSNR） increases by 46.5%, the Structural Similarity Index （SSIM） increases by 31.4%, and the Learned Perceptual Image Patch Similarity （LPIPS） decreases by 34.8%. Furthermore, the proposed method exhibits superior performance compared to MVS and Poisson reconstruction: PSNR increases by 22.6% and 9.2%, SSIM increases by 20.4% and 9.8%, and LPIPS decreases by 27.6% and 16.8%, respectively. The 3D ore models generated by this method effectively characterize mineral morphological features, providing a low-cost, high-quality 3D reconstruction solution for mineral resource research.