Extracting phenotypic parameters of camellia oleifera using neural radiation field and path analysis

Camellia Oleifera has been widely cultivated as a vital economic tree species for its oil-rich seeds. Accurate phenotypic analysis is essential to the conservation, breeding, and optimization. Precise measurement of phenotypic parameters can also advance the breeding programs for the high crop yields and optimal genetic traits. However, the complex canopy structure of oil tea trees has presented the significant challenges to extract the phenotypic data under field conditions. Manual measurements and multi-view stereo (MVS) reconstruction can often limited to the accuracy, efficiency, and scalability, particularly when applied to the dense and intricate tree canopies. In this study, an innovative approach was introduced to combine the neural radiance fields (NeRF) and path analysis. The high-precision extraction of phenotypic parameters was also performed on the Camellia Oleifera trees in fields. Multi-view images were captured from the fixed positions around the tree. A NeRF model was trained to reconstruct an accurate three-dimensional point cloud of the tree. Both structural and high-frequency were preserved after reconstruction. Path analysis was subsequently employed to segment the tree stem and leaf clusters within the point cloud, thus enabling the precise extraction of key morphological and phenotypic parameters. Experimental results show that the NeRF significantly outperformed the traditional MVS, in terms of both efficiency and reconstruction quality. Specifically, the NeRF approach was reduced the point cloud reconstruction time by approximately 90%. While the reconstruction quality was improved with the peak signal-to-noise ratio (PSNR) by 10%. Furthermore, the path analysis-based segmentation demonstrated that the superior performance was achieved to separate the stems from leaves. There was also the increase in the recall, precision, F1 score, and overall accuracy of segmentation, compared with the conventional geometric feature-based and region-growing approaches. Respectively, compared with the geometric feature and the region growing approaches. The high accuracy was achieved to extract the phenotypic parameters with the minimal error, compared with the manual measurements. Specifically, the relative errors of tree height, crown width, canopy height, and trunk diameter were 0.519%, 0.325%, 0.364%, and 4.491%, respectively. In leaf parameters, the coefficients of determination (R²) for the leaf length, leaf width, leaf area, and leaf shape index were 0.98, 0.94, 0.97, and 0.93, respectively, indicating the excellent agreement with the manual measurements. As such, three-dimensional point clouds of Camellia Oleifera were accurately reconstructed to extract the precise phenotypic parameters, particularly under the challenging field conditions, where the complex canopy structures were involved. A promising solution was offered to extract the phenotypic parameters of plants with the dense and complex canopies, indicating the rapid, reliable, and cost-effective approach alternative to traditional measurement. New possibilities were opened up to rapidly and accurately obtain the phenotypic data for the large numbers of trees in field. Breeding strategies were improved to support the plant phenotyping in precision agriculture. Additionally, the implications were gained for the species with the similarly complex canopy structures. A broader range of plant species can be extended to improve the computational efficiency, crop management, yield prediction, and precision farming.