Three-Dimensional Reconstruction of Potato Plants Based on Improved Neural Radiation Field Theory

Intelligent and reliable access to high-quality crop phenotypic big data, which can be achieved by three-dimensional (3D) reconstruction of crop plants, will greatly promote the research in digital breeding. However, efficient achievement of satisfactory accuracy in 3D construction has remained challenging. To address this, we developed a dataset preprocessing algorithm (referred to as Laplacian) as well as an efficient and effective 3D reconstruction algorithm (referred to as NGP). As a case study, the Laplacian algorithm was tested on video datasets sampled from potato plants. Results showed that 3D reconstruction algorithms including NGP as well as classical algorithm Colmap delivered higher PSNR (1.08%-1.62% increase) and SSIM (0.34%-0.74% increase) on optimized datasets than raw datasets, implying the effectiveness of the Laplacian algorithm. Comparing NGP and Colmap on the same (raw and optimized) datasets showed that NGP delivered significantly better in PSNR and SSIM. In the meanwhile, the tests showed that in terms of reconstruction speed, NGP reconstructs 10.2 times faster than Colmap when the number of iterations was set as 20000. The strategy proposed in this study provides a novel solution for achieving efficient, low-cost and high-precision 3D reconstruction of field crops, which paves the path for intelligent and reliable phenotyping.