The 3-dimensional Plant Organs Point Clouds Classification for the Phenotyping Application based on CNNs.

Abstract

The rice breeding produces the high-throughput via a genotyping technology. It can rapidly test and analyze on a large number of samples while the performance of phenotypic evaluation is still very low because of the manually evaluation. Therefore, this is the main barrier retarding the new rice varieties development. This research is aimed to develop a method for classifying plant organs from 3D point cloud in order to analyze plant morphology or architecture automatically. The rice plant was scanned with a 3D laser scan machine. The points in the cloud were reduced by the skeleton skimming method because the number of points in each cloud group is too large. Thus, it is necessary to preprocess before importing into neural networks for classification. The PointNet was selected as the 3D classifier in this research. The first experiment was conducted in order to evaluate the proposed method. The result showed that the proposed method can classify rice organs, regardless of rice varieties, with accuracy of 87.04%. Then, the second experiment was conducted in order to obtain the accuracy of the network for each rice variety to demonstrate the influence of rice cultivars in the classification due to their different shapes. The results showed that the SPRLR, which had large numbers of leaves and yield, has the lowest accuracy of 51.61% while the other varieties with the greater leaf and panicle distribution have a much better accuracy. The Nieow dum had 91.16% accuracy while Jae hwa, Kaow lueng and Kam had 89.06%, 86.52% and 75.22% accuracy respectively.