Recognition and phenotypic detection of maize stem and leaf at seedling stage based on 3D reconstruction technique

Abstract

As one of the major global food crops, rapid detection of seedling maize phenotypic traits is important for maize cultivation, management and variety selection. Due to the lack of a systematic approach for the morphological-physiological phenotypic profiling of maize growth stages, it is urgent to overcome the challenges of multi-view 3D reconstruction and phenotypic detection in seedling maize. In this paper, recognition and phenotypic detection of maize stem and leaf at seedling stage was proposed based on 3D reconstruction technology. First, a maize heterogeneous data collection system was constructed using three Kinect v2 sensors to acquire 810 sets of color images and depth data for the maize plant. Second, maize plant data were obtained through filtering, radius outlier removal, and Euclidean distance segmentation algorithms. Third, an improved random sample consensus − trimmed iterative closest point (RANSAC-TrICP) algorithm was employed for 3D registration of multi-view maize point clouds, achieving an average registration error of 0.0030. On this basis, a maize stem and leaf recognition method was established, which integrated eigenvalue decomposition and normal analysis techniques, achieving an accuracy of 0.9897. In addition, the density-based spatial clustering of applications with noise (DBSCAN) clustering algorithm was used to identify individual leaves, with an accuracy of 0.9516. Finally, 3D image processing and mathematical statistical algorithms were used to establish the plant height algorithm based on3D Euclidean distance, the leaf length algorithm based on fitting the single-leaf axis, the canopy width algorithm based on the external rectangle, and the stem thickness algorithm based on the least-squares method of fitting a circle. The results showed that the R² values for plant height, canopy width, leaf length, and stem thickness, were 0.9723, 0.9788, 0.9796, and 0.9876, respectively, comparing the calculated values with the measured values. This method effectively addressed the challenges of high-throughput phenotypic detection technology in monitoring maize growth state, providing a quantitative basis for the scientific regulation of phenotypic traits in maize cultivation, management, and breeding.