Key Research On Extraction Of Rice Panicle Traits

Food security is an important foundation of national security and the top priority of state governance.Rice is one of the three major grains in the world,and it is the main grain crop in China,which plays an extremely important role in grain production."Agriculture puts seed first".Cultivating high-yield,high-quality,and stress-resistant rice varieties is an important goal of rice breeding,and yield traits are the focus of breeding research.The whole process of rice breeding needs to obtain a large number of sample yield traits as analysis indicators.Rice panicle is the organ of fruit growth,and the panicle traits such as the number of spikelet per panicle,the number of filled spikelet,and the grain type are directly related to yield.Therefore,the accurate extraction of rice panicle traits is of great significance for rice breeding and is an indispensable link.The traditional measurement method is mainly manual,which is complicated,low efficiency and low precision,and has become the bottleneck restricting rice breeding research.The digital measurement generally requires threshing first,which will inevitably cause grain residue and damage.This affects the accuracy of the measurement,and it is difficult to accurately distinguish between filled and unfilled spikelets.Therefore,there is an urgent need for a new method that can provide rapid,reliable,and comprehensive nondestructive detection and intelligent analysis for the extraction of rice panicle yield without threshing,so as to break through the limitations of current rice panicle trait extraction technology.But at present,there are three problems in the non-destructive detection of rice panicle traits that need to be solved:(1)The existing research work mainly uses single mode imaging,which makes it difficult to obtain the information of the kernel inside the spikelets;(2)Overlapped spikelet are difficult to segment and the process of existing methods is complicated and the error is large;(3)The natural structure of rice panicle cause grain occlusion,resulting in image missing.In view of the above three problems,the main research content of this dissertation includes:(1)Setting up a dual-mode imaging system with RGB/X-ray methods to obtain the reflected light image of the outer contour of the rice panicles and the transmission projection image of the internal structure.By using digital image processing,a total of 45 parameters were extracted from three categories including spikelet traits,kernel traits,and panicle structure traits.Thus a comprehensive set of non-destructive extraction methods for rice panicle traits was constructed;(2)Using the object detection technology of deep learning non-destructively,the panicle spikelets can be identified and counted end-to-end,by the means of network redesign,dataset construction and parameter optimization.Among them,the Faster R-CNN combined with the feature pyramid network achieved 0.75 m AP,and the correlation coefficient R2 between the prediction and the true value is 0.99.Also it has good robustness to different varieties,different lighting modes,and different panicle spreading types.In addition,it has achieved a breakthrough in the instance segmentation of spikelets,with 0.67 m AP;(3)Aiming at the problem of image missing caused by mutual occlusion of panicle spikelets,the images with occluded grains of different varieties were restored using generative adversarial network,through the construction of occluded grain datasets and network optimization.The grain traits after restoration are compared with the real values,which indicate that the relative error is between 1.6% and 2.2%,and the correlation coefficient R2 is between 0.86 and 0.97.From the perspective of scientific significance,this study overcomes the shortcomings of difficult threshing and poor separation of the filled and unfilled grains in traditional destructive measurement of rice traits,and can provide a new method for non-destructive measurement of rice panicle traits for rice breeding research.From the perspective of application prospect,the methods proposed in this study do not need threshing and shelling,and do not need to separate filled and unfilled grains.It provides a fast and accurate new way to acquire rice panicle traits,which can save a lot of labor cost and time cost of yield traits scorering,and has certain market application prospect and economic value.