3D crop phenotyping technology provides critical support for screening morphology-related plant genes and identification of germplasm resource. Organ segmentation or recognition is the first key step in 3D crop phenotyping, where inductive deep learning currently dominates as the mainstream methodology. However, the high requirement for data annotation in inductive learning paradigm has transformed the manual data labeling into a labor-intensive task, thereby in turn restricting the progress of inductive learning. This problem has inspired us to leverage Graph Neural Networks (GNNs) as the transductive learning tool to directly segment organs on sparsely annotated crop point clouds. We propose a Dual-branch Graph Convolutional Network (DBGCN) that only requires sparse labels to perform organ instance inference directly on plant point clouds that have featureless point features. Different from existing graph-based networks, DBGCN not only carries out the static-feature-space graph convolutions that are good at mining and aggregating on local information on the point cloud, but also incorporates dynamic graph convolutions that captures the potential changes of the graph manifold in deep feature space. Extensive experiments prove that the fusion of two types of graph feature convolutions brings a high node (point) classification accuracy, outperforming mainstream GNNs and even several popular inductive deep architectures. On the PlantNet sub-dataset, DBGCN achieves an mAcc (mean accuracy of node classification) of 93.00% under 1.95% manual annotation ratio. On the Soybean-MVS sub-dataset, DBGCN achieves an mAcc of 91.05% under 4.88% manual annotation ratio. Furthermore, our DBGCN not only works well on crop 3D data but can also serve other applications such as the segmentation of point cloud data for large-scale street view. Our dataset and code can be found at https://github.com/chinazhouzhaoyi/DBGCN/tree/master/

In order to address the challenge of non-destructive detection of tomato fruit ripeness in controlled environments, this study proposed a real-time instance segmentation method based on the edge device. This method combined the principles of phenotype robots and machine vision based on deep learning. A compact and remotely controllable phenotype detection robot was employed to acquire precise data on tomato ripeness. The video data were then processed by using an efficient backbone and the FeatFlowNet structure for feature extraction and analysis of key-frame to non-key-frame mapping from video data. To enhance the diversity of training datasets and the generalization of the model, an innovative approach was chosen by using random enhancement techniques. Besides, the PolyLoss optimization technique was applied to further improve the accuracy of the ripeness multi-class detection tasks. Through validation, the method of this study achieved real-time processing speeds of 90.1 fps (RTX 3070Ti) and 65.5 fps (RTX 2060 S), with an average detection accuracy of 97% compared to manually measured results. This is more accurate and efficient than other instance segmentation models according to actual testing in a greenhouse. Therefore, the results of this research can be deployed in edge devices and provide technical support for unmanned greenhouse monitoring devices or fruit-picking robots in facility environments.