Research On Methods Of Rice Yield Prediction And Weed Identification Based On Deep Learning

Rice is one of the most important food crops in the world and also the main food crops in China.Accelerating rice breeding is of great significance to ensure national food security and the formulation of relevant policies,and timely forecasting of rice yield can speed up the process of rice breeding.At present,breeding methods usually involve planting multiple varieties in small plots,and yield estimation usually requires manual completion,which is labor-intensive and inefficient.Therefore,it is necessary to explore methods for yield measurement in small plots.Grass damage is closely related to rice yield,so weed control and grass damage monitoring are very necessary to ensure rice yield.To solve the above problems,this study took the rice planted and managed in the rice experimental field as the research object.Four kinds of rice,including Jifeng 105,Shengxiang66,Xudao 9 and Jifeng 103,were planted in the experimental field according to three different densities.The rice yield data and grass damage in the field in 2021 and 2022 were monitored.This paper developed the prediction method of rice yield and the identification method of paddy weeds,and integrated the above methods based on MATLAB Simulink toolbox to develop the intelligent detection system of yield and grass damage on Android mobile phone.Finally,the detection system was tested in the field.The main research contents and conclusions are as follows:(1)The number of rice panicle per unit area is one of the key indicators to estimate yield.Therefore,this study developed a method to predict plot yield based on the number of rice panicle per unit area.Canopy images of rice at filling stage were collected in October 2021,and an ear detection method was developed based on the improved target detection model YOLOv5.The model detection performance was further improved by improving the model and introducing attention.Model test results show that the average precision(AP)of the optimal model CBAM-YOLOv5s can reach 88.4%,and the robustness test shows that CBAM-YOLOv5s can detect rice ears in complex working scenarios.The yield prediction model was established using panicle number per unit area predicted by rice ear detection method.The results showed that the mean absolute percentage error(MAPE)of yield prediction was up to 13%,which indicated that the method of predicting rice yield based on canopy panicle number was feasible.(2)In order to further improve the prediction accuracy of rice in the plot and realize the prediction of rice yield at multiple growth stages,this study monitored rice at tillering stage,jointing stage,booting stage,filling stage and maturity stage respectively from June to October 2022.The paddy in the plot with a unit area of 0.25m~2 was taken as the research object.Ten color indices such as normalized green and red difference index and 14 texture features such as large gradient dominance were extracted based on visible canopy images.Principal component analysis was used to reduce the dimension of color indices and texture features.The extracted principal components were used to establish and test rice yield prediction models for multi-growth periods based on multiple linear regression and BP neural network.The results showed that compared with the multiple linear regression model,the mean absolute percentage error(MAPE)between the estimated yield and the actual yield of BP neural network was 17.2%,12.7%,12.3%,9.7%and 8.9%,respectively.The above results indicated that the method of using visible canopy images to predict rice yield was effective,and had better results in grain filling stage and maturity stage,and the accuracy was significantly improved compared with the method using only ear number to measure rice yield.(3)After monitoring grass damage in the field,this study found that grass damage is closely related to rice yield,and intelligent weed identification method can guide precise weeding,which is of great significance for ensuring rice yield.To this end,this study monitored grass damage in paddy fields and obtained images of six major weed species from June to October 2022,and then obtained RE-SNet,an optimal weed recognition model,based on Shuffle Net v2 model trained by transfer learning.The test results show that RE-SNet has high robustness to illumination,shielding and different growth periods,and has high recognition accuracy,fast recognition speed and small size,so it has the potential to be applied to practical agricultural production.(4)To help farmers and breeders achieve rice yield prediction and grass damage monitoring in the field.This study first developed wechat mini program based on the rice ear detection model and Django Web framework and conducted field tests.However,from the field tests,this study found that wechat mini program had high requirements on the field network environment.In order to solve this problem and improve its functions,Based on MATLAB Simulink toolbox and combined with rice yield prediction model,rice ear count model and weed recognition model,this study developed an intelligent detection system of yield and grass damage on Android mobile phone.The processing and calculation data of the detection system were based on the CPU of the mobile phone itself,without network.Field tests showed that the detection system achieved good results in rice yield prediction,ear number prediction and weed identification at maturity stage.The prediction accuracy of yield prediction could reach 83.7%,ear number prediction accuracy could reach 79%,and the average recognition accuracy of weeds was 88.4%.