Phenotypic Parameters Estimation Of Lettuce In Two-dimensinal And Tree-Dimensinal Space

Lettuce is an important economic vegetable crop in the world,with a wide range of varieties and high nutritional value.The phenotypic study of lettuce is the key to screen lettuce varieties with high quality and high yield.At present,the acquisition of phenotypic parameters of lettuce mainly relies on manual visual inspection and manual measurement,which has disadvantages such as high labor cost,low efficiency and lack of real-time performance.Therefore,in the process of crop breeding and cultivation,there is an urgent need for non-destructive,rapid,low-cost,and automatic measurement of crop phenotypic parameters.In this study,the main phenotypic parameters of lettuce in two-dimensional and three-dimensional space were obtained by image processing technologies such as learning-based image semantic segmentation and multi-view stereo aiming at the non-destructive and automatic measurement of phenotypic parameters.The main research contents and conclusions are as follows:(1)Research on lettuce image segmentation algorithm and measurement of phenotypic parameters in two-dimensional space.Firstly,the performances of these several semantic segmentation networks,including PSPNet,Deep Lab V3+,CGNet and OCRNet network,were compared in terms of operational efficiency,convergence performance and segmentation accuracy.The experimental results shown that the OCRNet segmentation algorithm in this paper had the best performance in the task of semantic segmentation of lettuce images.So OCRNet network was selected for lettuce image segmentation.Then,for the data imbalance problem of the lettuce dataset,the loss function of OCRNet used Dice Loss to optimize the network training.And its average pixel accuracy(MPA)reached 98.92%,and the average intersection ratio(MIo U)reached97.97%.The Height and area of lettuce extracted by image processing algorithms were hightly correlated with the height and area of lettuce measured mauually,and their the coefficients of determination(R~2)were more than 0.99.Based on the image segmentation data,image processing technology was used to automatically measure 37 plant phenotypic characteristics of 1600 lettuce images of 200 lettuce varieties at seven time points.And the obtained plant height and projected area were used to detect the growth of lettuce.(2)Research on the 3D reconstruction algorithm of lettuce and measurement of phenotypic parameters in 3D space.In order to improve the network feature representation ability,this study proposed to insert the channel attention SE(Squeeze-and-Excitation)module into the learning-based multi-view stereo Cas MVSNet network.The experimental results shown that under the same equipment conditions,the improved Cas MVSNet method in this paper was better than that of Cas MVSNet on the DTU dataset.Its mean overall accuracy(Overall)reached 0.364,and which was 5%higher than Cas MVSNet.Then tested on the Tanks and Temples dataset,the average f-score value of improved method was 1.37%higher than Cas MVSNet.The point clouds of rapeseed,tomato,and lettuce were reconstructed using traditional multi-view stereo vision algorithms such as Visual SFM,Colmap,and Agisoft Photoscan,the Cas MVSNet algorithm,and the improved Cas MVSNet method.Comparing the point cloud results of all methods,it was found that the point cloud reconstructed by the improved algorithm in this paper is relatively dense,fine and less noisy.The original point cloud data obtained by the 3D reconstruction algorithm had the characteristics of disorder,redundancy and large amount of data.In this study,the original point cloud data was firstly processed by the steps of point cloud segmentation,point cloud downsampling and point cloud denoising.After then,the measurements of plant height,surface area,convex hull volume and concave hull volume of two lettuce plants at three time points were completed.In order to verify the advantages of the proposed algorithm in terms of accuracy,the improved Cas MVSNet method was used to reconstruct the point cloud of 21 rapeseed plants,and the phenotypic parameters such as plant height,leaf length,and leaf width of the rapeseed point cloud were measured by the algorithm.And the phenotypic parameters such as plant height,leaf length,and leaf width of the rapeseed point cloud were measured by the algorithm were compared with the manual measurements.The mean absolute percentage errors(MAPE)of the plant height,leaf length and leaf width were0.04%,0.03%,and 0.03%,respectively,the root mean square errors(RMSE)were0.31cm,0.18cm,and 0.09cm,and the coefficients of determination were 0.98,0.97,and0.97.It shows that the algorithm in this paper has high accuracy.In summary,this study used learning-based image processing technology to measure the 2D and 3D phenotypic parameters of lettuce,which provided a feasible approach and data support for the phenotype grouping and genetic breeding research of lettuce.