Phenotype Extraction Of Potato Based On Multi-Source Data

Potato is an important cash crop that has high nutritional value,strong resistance to adversity,and great potential for increasing yield.It has received widespread attention from countries around the world.The growth monitoring,pest and disease detection,and yield prediction of potatoes often require multi-dimensional phenotypic parameters.In this study,two sensors,a laser scanner and an RGB camera,were used to obtain the three-dimensional point cloud and RGB image of potato plants.A one-to-one correspondence between the 3D point cloud and RGB image was established,and then various features,including three-dimensional morphological features,two-dimensional morphological features,color features,and texture features,were extracted from the same potato plant.The method used in this study fully utilizes the advantages of laser point clouds with rich volume features and RGB images with color and texture features,overcomes the limitations of a single sensor,and obtains diverse types and high-precision phenotypic parameters.This research is of great significance for the cultivation and breeding of potatoes.The main research contents and conclusions of this article are as follows:(1)To collect data efficiently and flexibly,this study used an asynchronous approach to acquire the laser point clouds and RGB images of potato plants.Six pots of potatoes were selected as the subjects of this study,and a comparative experiment was designed to determine the optimal laser scanner scanning scheme based on the point cloud’s effectiveness under different parameters.This study used the automated conveying device of the crop performance platform of Huazhong Agricultural University to transport potato plants efficiently to the bottom of the RGB camera and collect high-resolution RGB images in numerical order.This study designed a comparative experiment using six pots of potatoes,determined the optimal scanning scheme of the laser scanner according to the effect of point clouds under different parameters,and planned the site settings of the laser scanner to ensure the integrity and efficiency of the collection process.(2)In terms of 3D point clouds,this study proposes an improved Mean shift clustering algorithm that achieves high efficiency and accuracy in segmenting point clouds.The stem and leaf of single-plant potato plant point clouds were accurately segmented by cleverly combining the Euclidean clustering and K-Means clustering algorithms.At the same time,this study also discussed the feasibility of using the Stratified Transformer point cloud segmentation network based on deep learning in organ segmentation,compared traditional point cloud segmentation methods with deep learning-based point cloud segmentation methods,and summarized the advantages,disadvantages,and applicable scenarios of the two methods.(3)In terms of RGB images,this study improved the loss function of OCRNet and analyzed and compared five deep learning semantic segmentation networks:Deep Lab v3+,UPer Net,PANet,OCRNet,and the improved OCRNet.It was demonstrated that the improved OCRNet network has high accuracy in segmenting top-view RGB images of potatoes.The network was selected to perform semantic segmentation on RGB top-view images of potato plants and the results were used for subsequent phenotype parameter extraction.(4)This research proposes a strategy for establishing a one-to-one correspondence between individual potato RGB images and laser point cloud data using numbering to address the correspondence problem.This strategy improves the flexibility of data collection and combines the advantages of RGB images,which carry texture and color features,with the advantages of three-dimensional point clouds,which have volumetric features,enriching the types of phenotype parameters.Based on the one-to-one correspondence between RGB images and laser point clouds,this study extracts 23 two-dimensional phenotype parameters and 10 three-dimensional phenotype parameters of the same potato plant.(5)This study conducted an accuracy evaluation of three phenotypic parameters of potato plants,including leaf number,plant height,and maximum crown width.Compared with manual counting and measurement,the MAPE of leaf number,plant height,and maximum crown width in this study were 8.6%,8.3%,and 6.0%,respectively,with RMSE of 1.371 leaves,3.2 cm,and 1.86 cm,respectively,and R~2 of 0.93,0.95,and 0.91,respectively.The results of the accuracy evaluation show that the extracted phenotype parameters can reflect the growth status of potatoes accurately and efficiently.The combination of RGB image data and 3D point cloud data of potatoes can fully utilize the advantages of rich texture and color features of RGB images and volumetric information provided by 3D point clouds.This achieves high-precision and non-destructive extraction of two-dimensional and three-dimensional phenotype parameters of potato plants.