Estimation of external phenotypic parameters of Bunting leaves using FL-DGCNN model

Plant leaves blocked under natural growth conditions cannot fully acquire the complete external phenotypic data. Therefore, this study aims to estimate these parameters in low-cost and automated batches using Dynamic Graph CNN with Feature Layering (FL-DGCNN). A stack encoder model was also used for Bunting (Epipremnum aureum) leaves. A camera was selected to shoot at a certain angle between two images. Further, motion recovery was utilized to reconstruct the three-dimensional model of the plant after feature points matching. Straight-through filtering and clustering segmentation were used to obtain a single chip Point cloud data. Specifically, the geometric model of the blade using surface parameters was discretized into a point cloud, but the point cloud was incomplete to a certain proportion to simulate the natural growth state. An auto-encoder model was modified into a deep-structured stack encoder under a multi-layer combination, and then to reduce the distance between the input point cloud and the actual point cloud. As such, the incomplete geometric model point cloud achieved shape completion after the training. The determination coefficients of leaf length, width, and area estimated by the stack encoder decreased less, as the percentage of incomplete point clouds increased, while those estimated by the auto-encoder decreased by multiples. The robustness of stack encoder completion was better in the leaf point cloud, compared with autoencoders under the same incompleteness. The shape was also similar to the original point cloud after completion within 40% of incompleteness. There were great variations in the shape of the blade when the blade was incomplete or more than 50%. A better performance was also achieved in the occluded blades. The completed point cloud was input into the FL-DGCNN deep learning network, and the feature maps were then extracted at different scales in the image pyramid, thereby enhancing semantic and geometric information. The farthest point sampling was used to extract from the original point cloud. The extracted features were connected to obtain a vector after feature layering and fusion, particularly for point clouds with contour features at different scales. The basic neural network module of edge convolution structure was adopted to better capture the local structure. The point cloud structure was represented by the directed graph, where the edge feature was obtained from the neighbor nodes. The local features of each group were superimposed on the shallow and deep network for the multiple perceptions in the original edge convolution structure, and then the leaf length, width, and area were estimated for the external phenotypic parameters of leaves. The highest accuracy was achieved to estimate the leaf width and area, followed by that of leaf length with a relatively small error. The determination coefficient and root mean square error were better than before, indicating a relatively lower error and stronger ability of feature extraction, compared with multiple networks. Additionally, a total of 200 leaves of Epipremnum aureum were collected in the experiment to verify the model, where the estimated values were linearly fitted to the measured. The determination coefficients and root mean square errors of leaf length, width, and area were 0.92 and 0.37 cm, 0.93 and 0.34 cm, 0.94, and 3.01 cm2, respectively. The experiment demonstrated that the model is highly effective to estimate the external phenotypic parameters of plant leaves.