The paper introduces GeoGaussian, a novel approach to enhance the geometry preservation in 3D Gaussian Splatting for scene rendering. The method addresses the issue of geometry degradation in non-textured regions during the optimization process, which significantly affects the quality of novel views. GeoGaussian introduces a pipeline that initializes thin Gaussians aligned with smooth surfaces, using normal vectors extracted from point clouds. This is followed by a densification strategy that ensures new Gaussians lie on the same plane as the original, maintaining geometric consistency. The optimization module includes a geometrically consistent constraint to encourage thin Gaussians to align with smooth surfaces, improving the overall rendering quality. Experimental results on public datasets demonstrate that GeoGaussian outperforms state-of-the-art methods in terms of geometry accuracy and photo-realistic novel view rendering. The method is particularly effective in structured environments and sparse view scenarios, showing robustness and generalization capabilities.The paper introduces GeoGaussian, a novel approach to enhance the geometry preservation in 3D Gaussian Splatting for scene rendering. The method addresses the issue of geometry degradation in non-textured regions during the optimization process, which significantly affects the quality of novel views. GeoGaussian introduces a pipeline that initializes thin Gaussians aligned with smooth surfaces, using normal vectors extracted from point clouds. This is followed by a densification strategy that ensures new Gaussians lie on the same plane as the original, maintaining geometric consistency. The optimization module includes a geometrically consistent constraint to encourage thin Gaussians to align with smooth surfaces, improving the overall rendering quality. Experimental results on public datasets demonstrate that GeoGaussian outperforms state-of-the-art methods in terms of geometry accuracy and photo-realistic novel view rendering. The method is particularly effective in structured environments and sparse view scenarios, showing robustness and generalization capabilities.