Research On Field Rice Panicle Segmentation And Nondestructive Yield Prediction Based On Deep Learning

As the main food crops and important cash crops in China,sustainable yield increase has always been the primary goal of the rice breeding research.With the rapid growth of population in china,the demand for rice is more and more large,and the situation of land scarcity is difficult to alleviate in short time.Therefore,the breeding of high-yielding varieties of rice has become the focus part in the field rice research.Most of the traditional methods in field rice yield prediction are destructive methods.After rice harvesting,threshing,sun-drying and weighing,the researchers convert the final rice yield.This method is time-consuming,and it is very easy to introduce artificial error because of operation error.Most of nondestructive estimation methods extract the representative meteorological traits or spectral factor to construct reasonable yield forecasting models by using the field meteorological model or spectral index method.But there methods have large amount of calculation and need to rely on a large number of supportive data.So,the practical value of there methods are not high.Therefore,fast and accurate field rice yield estimation is still the key point of rice research.In this paper,a new method based on image to predict rice yield in field is presented.Panicle is the important reproductive organ of rice.The panicle length and spike number per panicle are closely related to the final rice yield.Therefore,the accurate segmentation of rice panicle is the most critical step in the prediction of rice yield based on image.In this paper,a digital camera is used to obtain the images of rice plot from different angles in the field,then we designed three different panicle segmentation algorithms,including Panicle-SEG,Panicle Net and Panicle Net v2 using deep learning technology.Five criterias,including precision,recall,F-measure,Io U,and segmentation efficiency are used to evaluate the final results.Compared with the traditional segmentation algorithms,the proposed algorithms using deep learning show excellent segmentation performance.For the segmented rice plot images,this paper extracted the traits of rice from three angles.Firstly,based on the scale of the plot,we extract the feature description from color,texture,shape and some scale invariable parameters.Then we investigate the details of a single panicle image,two traits including panicle length and panicle area were extracted.Finally,the grain traits for single rice were analyzed.In this paper,we use diverse perspectives to analysis the panicle.In this way,we can get a more comprehensive traits description of panicle.For a large number of panicle traits we have extracted,this paper discussed the building method of regression models based on the combination of multiple panicle traits.We attempt to build different regression models including linear and nonlinear models,and compare the performance of each regression model.From the final prediction results,nondestructive rice yield estimation based on panicle images provides a possibility for accurate early rice yield estimation,and it provides a successful case for the application of deep learning technology in agriculture.This early rice yield estimation method,that make farmers grasp the production of different fields in time,summarize the experience and deficiencies in rice planting process.In this way,they can adjust management and cultivation measures timely,analysis of the main factors affecting the yield,that provides a new thought for the research of rice yield and breeding.