Research On Wheat Canopy Leaf Egmentation And Nitrogen Nutrition Diagnosis Modeling Based On Deeping Learning

Recently,digital diagnosis technology have been well developed and widely used in our daily life,and monitoring crop nitrogen status with rapid,non-destructive and accurate way is a key part in the precision agriculture.So it’s import to build digital diagnostic system.In this study,canopy images taken by digital camera and nitrogen indicator were taken from 2 consecutive years of wheat field experiments,at the same period.The paper extracted color characteristic values of wheat leaf and analyzed the correlation of LNC,and did research on establishment method of modeling to estimate LNC.Extracting features is based on precise segmentation of canopy images.However,background and light in the field environment,and changes in the canopy blades during fertility,brought challenges of segmentation.For the above problems,the main contributions of the paper were as follows:(1)The net model based on U-net trained in small data set,improved segmentation accuracy of canopy images in the field.The first step was to manual mark the background and blades in canopy images token from 2013,using Classification 0 or l,to make the data set.The second step was to train 2 modelbased on U-net,of different inputs(256_unet and 512_unet),and also to use k-means clustering algorithm to segment the image of wheat canopy,based on H component.The third step was to compare accuracy of segmentation in testing data set.For background complex class,compared 256_unet,the mean intersection over union(MIoU)and frequency intersection over union(FIoU)of 512_unet were both increased by 4%;compared.(2)The images of winter wheat canopy taken from 2013 to 2014 were segmented using kmeans clustering algorithm based on H component and the monochromatic components of 3 color spaces(HSV、RGB and L*a*b*)were extracted as independent variables of the model estimating wheat LNC.The estimation models were constructed using multivariate linear regression(MLR)、support vector regression(SVR)and random forest(RF)algorithm.The models in 3 color space were studied through 10×10 nested cross-validation,taking coefficient of determination(R2)and root mean square error(RMSE)as evaluation index.In single color space,the fitting and generalization performance of 3 algorithm models were optimal in HSV color space,second in L*a*b*color space,and worst in RGB color space;RF had the best fitting performance,and also had a problem of overfitting,but the variance of the algorithm model dominated generalization error;the fitting ability of SVR was weaker than it of RF,which was better than it of MLR,the model had the best performance though;MLR has the weakest fitting ability,but the bias of the model dominated the generalization error and the model was underfitting and subject to the noise interference of data.In multi-color space involved 9 monochromatic components of 3 color spaces(HSV、RGB and L*a*b*),the fitting and generalization performance of 3 algorithm models were better than it in single color space and RF model had the best performance.Compared with the best single color space HSV,R2 of RF was increased by 2.67%and RMSE improved by 11.59%in training set and R2 increased by 7.57%and RMSE improved by 11.49%in testing set.In multi-color space,the fitting performance of RF was better than it of SVR.The generalization performance of RF was effectively improved,and the promotion ratio of RF was higher than it of SVR.The fitting and generalization performance of RF is best in 3 algorithm models using 9 monochromatic components of 3 color spaces,so RF can provide a reference for estimating LNC of wheat.(3)The improved model(my_unet)based on U-net,improved the accuracy of segmentation,and optimisedthe structure,compared U-net.The segmentation accuracy of 3 models(my_unet、512_unet andnew_unet)was compared in data set made from canopy images token from 2014.The result is that MIoU and FIoU of my_unet were improved by 1,6%、0.35%,compared with new_unet and also improved by 0.24%、0.29%,compared with 512_unet.Not only,my_unet incersed the accuracy,but also optimised the structure.(4)The model of double kernel based on SVR improved accuracy of estimating LNC,compared with model of signal kernel.The first step was to segment canopy images token from 2014 using ensemble model.The second step was to extracted the monochromatic components of 3 color spaces(HSV、RGB and L*a*b*).The third step was to construct estimation model using SVR with 3 different kernel(svr_rbf、svr_ploy and svr_dk).R2 of svr_dk、svr_rbf and svr_ploy were respectively 0.6752、0.6467、0.6348.RMSE of svr_dk、svr_rbf and svr_ploy were respectively 0.4560、0.4756、0.4835.The above-mentioned research results could provide the reference value for nitrogen nutrition diagnosis of wheat,also the model of segmentation could provide technical support for digital support for digital image processing technique applied in the actual agricultural extension.