Study On Rice Nitrogen Nutrition Monitoring Based On Digital Image Methods

Rice(Oryza sativa L.)is one of the three major food crops in the world.Its planting area and yield level significantly affect global food production security.In recent years,several environmental problems caused by experiential fertilization in the process of rice cultivation have become more and more prominent.Precision fertilization under the digital agriculture is more consistent with our national conditions,which embodies the goal of green agriculture development.At present,the identification and diagnosis of rice nitrogen nutrition are still limited to the observing rice leaf color growth and the experienced judgment of growers in Jiangxi.Therefore,it is necessary to obtain the basic information of crop nutrients quickly and accurately,and it is urgent to study the method of portable non-destructive analysis to assist farmers in nutrient management.Digital image technology is one of the mainstrem technologies to realize non-destructive monitoring of crop nutrients,and machine learning(ML)is a meaningful approach to learn intelligent monitoring of image.Aiming at some problems of rice nutrient monitoring based on the digital image,such as the error and omission of rice image segmentation in complex background,the difficulty in obtaining correct high rate of nutrient identification on small sample data sets,and the poor universality of the rice nutrient monitoring model.Based on the digital camera(Canon EOS 100D,Resolution72DPI)image data set and unmanned aerial vehicle(Mavica Air 2)image data set collected in Jiangxi field trials from 2019 to 2021,the key technologies of rice nutrient monitoring were studied,the main research work completed in this paper is as follows:(1)A method of rice canopy image segmentation based on threshold and improved UNet algorithm in complex background was proposed,which provided high quality image input data for building nutrient classification recognition model and monitoring model.Compared with the conventional threshold segmentation and K-means clustering segmentation algorithm,the segmentation accuracy of proposed method could reach0.969 under the same data set.Compared with the segmentation results of feature pyramid networks(FPN)and UNet++,the crop segmentation accuracy with improved UNet algorithm proposed by this paper had been improved,which provides support for further improving the precision of rice nitrogen monitoring based on digital image method.(2)A new method based on stochastic gradient descent(SGD)Optimizer+Efficient Net V2-S neural network was proposed to accurately classify and identify rice nutrient status.Firstly,the classification basis of the critical nitrogen concentration curve in rice canopy image was obtained by constructing the monitoring agronomic model NNI of rice nitrogen nutrition index.Secondly,the classification results of the nutrient status classification in the whole growth period of rice based on the Efficient Net V2-S neural network with digital image features showed that the average classification accuracy in the whole growth period was 0.88,which overcame the shortcomings of the traditional ML for classification and recognition by stages,and was significantly improved compared with the 0.71 of the Gradient Promotion algorithm with the highest classification accuracy in the traditional ML algorithms;Thirdly,the comparison of classification and recognition results showed that the classification accuracy of the recognition model based on Efficient Net V2-S network was higher than that of Alexnet,VGG19,Goog Le Net and Res Net50,and reaches 0.908 at the jointing stage;Finaly,based on the features of the optimizer,a recognition model of SGD Optimizer+Efficient Net V2-S network was proposed.The results showed that the recognition accuracy increased to 0.921 in the whole growth period.The experimental results verified the feasibility of the classification and recognition of rice nutrient status based on digital image,and provided a reference for the subsequent quantitative monitoring of rice nitrogen.(3)A nutrient estimation model of rice in the whole growth period was proposed,which combined features and improved Res Net50.By extracting the image features and fusing the color and texture features,the improved depth residual network Res Net50 was used as the regression model.The rice nitrogen quality sub-parameters are used as the output monitoring method.The prediction accuracy R~2of the proposed algorithm is 97%,and the RMSE value was 0.02 on the whole growth period data set.Compared with the traditional machine learning algorithm,the prediction accuracy of random forest was improved by 0.13,the prediction accuracy of LSTM is improved by 0.08,and the prediction accuracy of the Res Net50 algorithm was improved by 0.12.The data sets of different years were used for verification,and the verification results showed that the correlation coefficient between the predicted value and the measured value reaches 0.95.All the above results showed that the predicted results were basically consistent with the actual situation,and the proposed method was feasible for the quantitative monitoring of rice nutrients.(4)The feasibility of applying the method of building a rice nutrient monitoring model based on digital camera images to regional-scale UAV rice nutrient monitoring was verified.The application results showed that,by improving the image fusion features,the application of digital camera fusion features and improved Res Net50 rice nutrient monitoring model in UAV could be realized on a regional scale.The prediction accuracy of the proposed model was 93%,and the RMSE value was 0.04 for the whole growth period.Through the research of this thesis,the segmentation accuracy of rice canopy image under complex background was improved,and the qualitative and quantitative monitoring of rice nitrogen nutrient in whole growth period were realized for individual and regional scale,it can provide an important reference for the intelligent monitoring program of rice nutrients,and has a good application prospect.