Citrus accounts for more than 20% of global fruit production, yet accurate phenotypic analysis remains labor-intensive. Traditional manual measurements are slow and prone to bias, while orchard-scale three-dimensional (3D) reconstruction is complicated by manual tree circumvention, high planting density, complex and unstructured plant geometry, rapidly changing outdoor illumination, and sparse geometric features. Hardware scanners are costly, photogrammetry is sensitive to image quality and lighting, and simultaneous localization and mapping (SLAM), or Structure from Motion (SfM) pipelines can be computationally expensive. Neural rendering and multi-view learning improve realism but often demand substantial memory and processing power, limiting in-field deployment. Given these challenges, in-depth research is needed to develop scalable, robust, and robot-deployable 3D reconstruction frameworks for orchard environments.
Researchers at Huazhong Agricultural University (HZAU), including the College of Information, the College of Engineering, the Engineering Research Center of Intelligent Technology for Agriculture, and the National Key Laboratory for Germplasm Innovation & Utilization of Horticultural Crops, together with KermalMind Technology Co., Ltd., developed InspectGaussian. The study was published (DOI: 10.1016/j.plaphe.2026.100195) in Plant Phenomics in 2026 (Volume 8). It presents a coarse-to-fine Gaussian reconstruction pipeline for orchard inspection robots, combining RGB-D SLAM, YOLO-World detection, and improved 3D Gaussian Splatting (3DGS).
The InspectGaussian pipeline has four stages. First, an autonomous ground vehicle equipped with an Red, Green, Blue, and Depth (RGB-D) camera, LiDAR (Light Detection and Ranging), and an inertial measurement unit (IMU) collects synchronized images and depth maps while moving along tree rows. Second, ORB-SLAM3 (Oriented FAST and Rotated BRIEF Simultaneous Localization and Mapping 3) estimates camera poses and builds an initial sparse map; a dense mapping module then produces a globally consistent, colorized point cloud. Third, YOLO-World (You Only Look Once-World) detects plants and Quick Response (QR) code markers in two dimensions, and a 3D matching module uses poses and depth to associate each tree with its correct image sequence. Fourth, an improved 3DGS framework reconstructs each plant separately using a divide-and-conquer strategy. Region-aware masking focuses optimization on the target tree, while hard and soft depth regularization corrects Gaussian positions and opacity without disturbing rotation and scale parameters. In citrus orchard tests, the method achieved 96% average precision and 93% recall in plant view extraction. It reached 31.226 Peak Signal-to-Noise Ratio (PSNR), 0.915 Structural Similarity Index Measure (SSIM), and 0.067 Learned Perceptual Image Patch Similarity (LPIPS), with about 7 mm point-cloud error.
The authors said the framework was built around a practical robot workflow rather than manual scanning. They said row-based capture reduces labor and plant disturbance, while 3D matching helps recover correct plant identities even when two-dimensional detection misses a marker. According to them, combining depth regularization with region-aware refinement keeps Gaussian primitives focused on the tree, improving leaf-edge detail and point-cloud completeness. They added that the approach offers a scalable path toward automated, high-throughput orchard phenotyping.
The framework could support high-throughput phenotyping, intelligent orchard monitoring, and precision agriculture. In trials, plant height predictions reached a coefficient of determination (R²) of 0.802, with a root mean square error (RMSE) of 9.76 cm and a mean absolute error (MAE) of 5.76 cm. Primary branch angle predictions reached R² = 0.778, with RMSE = 4.92° and MAE = 4.26°. The authors noted that reliance on QR code markers adds labor. Future work will aim to remove marker dependency by improving natural feature recognition and pose estimation, enhancing adaptability, automation, and scalability.
###
References
DOI
10.1016/j.plaphe.2026.100195
Original Source URL
https://doi.org/10.1016/j.plaphe.2026.100195
Funding information
This work was supported by the National Key Research and Development Program of Chinaunder Grant 2023YFF1000100; the Fundamental Research Funds for the Central Universities in China under Grant 2662024SZ002 and Grant 2662024XXPY001.
About Plant Phenomics
Plant Phenomics publishes breakthrough research that integrates genomics, genetics, physiology, molecular biology, bioinformatics, statistics, mathematics and computer science to advance plant phenotyping—from the cellular level to entire populations. The journal tackles key scientific challenges in phenomics by showcasing cutting‑edge technologies for data acquisition, management, interpretation and modelling. Its scope spans practical applications across all plant sciences, aiming to translate novel phenotyping tools into real‑world solutions for breeding, crop management and fundamental plant biology. By fostering interdisciplinary innovation, Plant Phenomics serves as a hub for researchers developing high‑throughput, accurate and scalable methods to decipher complex plant traits, ultimately accelerating progress in agriculture and plant science.