Abstract: Three-dimensional(3D) reconstruction of potted plants remained challenging due to complex morphology and self-occlusion, while existing reconstruction methods often failed to accurately recover fine structures and maintain geometric completeness in such scenarios. These limitations primarily arose from insufficient structural representation and lack of semantic constraints in current approaches. We proposed a Gaussian-based reconstruction framework named Semantic-Structural Synergy-based Self-repairing gaussian reconstruction for potted plants (4SPP). The method comprised two core modules. First, a gaussian semantic dividing stage transferred 2D organ-level segmentation results into the 3D gaussian space, assigning semantic labels to each Gaussian by aggregating multi-view mask responses. Uncertain points were refined through a skeleton-guided proximity check and K-nearest neighbour propagation to ensure local semantic consistency. Second, a progressive structure repair stage detected persistent high-error regions via temporal rendering error analysis. For each detected region, candidate positions for new gaussians were estimated from the local neighbourhood depth, then refined toward the spatial centre of semantically consistent neighbours. The new gaussians inherited averaged geometric and appearance attributes from their neighbours, ensuring smooth integration into the existing structure and reducing the risk of artifacts. This study conducted experiments on nine varieties of potted plants from the real world, each with distinct structural characteristics. Results showed that the proposed method achieved an average improvement of 0.33 dB in peak signal-to-noise ratio(PSNR) compared with standard 3d gaussian splatting(3DGS), along with consistent gains in structural similarity index measure(SSIM) and reductions in learned perceptual image patch similarity(LPIPS). On a public plant dataset, the method achieved an average PSNR of 23.884 dB, outperforming multiple baseline approaches with improvements ranging from 0.50 to 0.75 dB. Visual comparisons demonstrated that the proposed method effectively alleviated reconstruction artifacts such as broken branches, structural gaps, and floating noise, particularly in regions affected by occlusion and weak texture. The reconstructed models exhibited improved structural continuity and more coherent organ-level distribution across different plant types. The efficiency test shows that after introducing of the repair module increased training time by approximately 14.9 percent and memory consumption by about 10.4 percent, which remained acceptable relative to the achieved improvements. Ablation studies further verified that both the semantic dividing module and the structure repair module contributed significantly to the overall performance, and their integration produced the most robust results. The proposed framework effectively improves the geometric completeness and semantic consistency of potted plant reconstruction. By incorporating multi-view semantic information into the 3D Gaussian representation and integrating a rendering error–driven progressive structure repair mechanism, the method alleviates geometric deficiencies in regions such as branch connections and occluded areas caused by insufficient structure from motion(SfM) initialization. Meanwhile, the introduction of organ-level semantic constraints helps regulate the distribution of gaussians, thereby reducing structural artifacts and improving overall reconstruction fidelity. This work provides a reliable technical solution for high-fidelity plant reconstruction and demonstrates promising potential in applications such as digital agriculture, plant phenotyping, and ecological analysis. Future work will explore the integration of generative models with botanical constraints to improve structure recovery in severely occluded regions, aiming to further enhance reconstruction completeness.