Abstract: Three-dimensional (3D) reconstruction of potted plants remained challenging, due to complex morphology and self-occlusion. Existing reconstruction cannot accurately recover fine structures and geometric completeness in such scenarios. It is often required for sufficient structural representation and semantic constraints. In this study, a self-Repair Gaussian reconstruction was proposed, named Semantic-Structural Synergy-based Self-repairing gaussian reconstruction for potted plants (4SPP). Two modules included: 1) The Gaussian semantic assignment stage projected 2D organ-level segmentation into the 3D Gaussian space and labeled each Gaussian by aggregating multi-view mask responses. Uncertain points were refined via a skeleton-guided proximity check and K-nearest neighbor propagation, particularly for local semantic consistency. 2) A progressive structure repair stage detected persistent high-error regions via temporal rendering error analysis. Candidate positions for new Gaussians were estimated from the local neighborhood depth in each detected region, then refined toward the spatial center of semantically consistent neighbors. The new gaussians inherited average geometric and appearance attributes from their neighbors, leading to smooth integration into the existing structure with less risk of artifacts. A series of experiments was conducted on nine varieties of potted plants from the real world, each of which shared structural characteristics. Results showed that the improved model achieved an average improvement of 0.33 decibel (dB) in the 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). An average PSNR of 23.884 dB was obtained on a public plant dataset, which was improved from 0.50 to 0.75 dB, compared with the multiple baselines. Visual comparisons demonstrated that the reconstruction artifacts were effectively alleviated under occlusion and weak texture, such as broken branches, structural gaps, and floating noise. Different plant types shared the structural continuity and a more coherent organ-level distribution after reconstruction. The efficiency test showed that the training time increased by 14.9 percent in the repair module, while memory consumption increased by about 10.4 percent, which remained acceptable relative to the improvements. Ablation studies further verified that both the semantic assignment and the structure repair module contributed significantly to the overall performance. Their integration also produced the most robust performance. The framework effectively improved the geometric completeness and semantic consistency of potted plant reconstruction. Multi-view semantic information was incorporated into the 3D Gaussian representation, with the rendering error–driven progressive structure repair mechanism. Geometric deficiencies caused by insufficient structure from motion (SfM) initialization were alleviated in the branch connections and occluded areas. Meanwhile, the organ-level semantic constraints were introduced to regulate the distribution of Gaussians, thereby reducing structural artifacts for the overall reconstruction fidelity. This work can provide a reliable technical solution for high-fidelity plant reconstruction, particularly for promising potential applications, such as digital agriculture, plant phenotyping, and ecological analysis. Reconstruction completeness can also be enhanced to integrate generative models with botanical constraints for the structure recovery in severely occluded regions.