| It is of great significance for wheat growth management and yield prediction to use digital image to detect nutrients in wheat growth period accurately,quickly and non destructively,which is also an urgent need for the development of agricultural mechanization and automation.To realize dynamic monitoring of wheat growth process and provide scientific guidance for rational variable rate fertilization,the wheat at jointing stage under the field environment is taken as the research object,and the accurate,rapid and nondestructive detection of wheat mineral nutrient content is taken as the research goal.By the experiment of different nitrogen application levels,the wheat images at two scales(leaf and canopy)and the corresponding chlorophyll content data are collected.Based on the digital image technology and machine learning algorithm,the models of chlorophyll content detection in two scales and the models of wheat nitrogen level discrimination in canopy scale are established.The main research achievements are shown as follows:(1)In order to solve the problem of target extraction of wheat leaf and canopy images in field environment,the characteristics of the two images are fully considered,and the target extraction methods of wheat leaf and canopy are proposed.In the aspect of leaf images segmentation,an iterative color segmentation method is proposed to realize the separation of background and leaf targets.The experimental results verify the effectiveness of this method.For the problem of uneven illumination and complex background of wheat canopy images in the field prototype,by analyzing the gray value characteristics of each component of pixels in RGB space of land,canopy leaves and shadow areas,a gray threshold segmentation method based on RGB space is proposed,which realizes the separation of wheat canopy and land.A large number of experimental results show that a good segmentation effect is achieved by the method on the images with uneven illumination and complex background which overcomes the over segmentation phenomenon of the small leaf area.(2)For the feature extraction of wheat images related to nutrient content in the field environment,a set of image evaluation index set is established,which fused multi-color spaces.In order to reduce the influence of natural environment on RGB color space,the rgb color space is constructed by a set of invariant moments,and the combined color feature index(CCFI)is constructed based on this space;A set of basic combination indexes,such as the mean values of RGB,rgb,HSI and La*b*color spaces are given;The color standard deviation features and canopy coverage that describe the growth characteristics of wheat canopy are introduced as image indicators,and the corresponding calculation formula is proposed.The effectiveness of the image evaluation index set is proved by experiments.(3)To detect the chlorophyll content of wheat leaves in the field prototype,a feature dimensionality reduction method of wheat leaf image evaluation index set is proposed based on Pearson correlation coefficient.In order to obtain the optimal detection model a correlation based stepwise input(CBSI)method is proposed,and the optimal detection model of chlorophyll content in wheat leaves is established.Compared with other models,Input11-LR is the best model,its R2 value is 0.727,and RMSE value is 5.005.The results of this approach provide a theoretical basis and method reference for the detection of chlorophyll content of wheat leaves in field prototype.(4)For the detection of chlorophyll content on the canopy scale of wheat in field prototype,considering the correlation between the image evaluation indexes with chlorophyll content and the effect of the lighting removal method,the influence of different shooting angles on each image evaluation index is analyzed,and 90 degree is determined as the best shooting angle.Then,based on the correlation degree between the evaluation index set of wheat canopy image and chlorophyll content,the linear regression,ridge regression,random forest and back propagation artificial neural network detection models of wheat canopy chlorophyll content are established,and the optimal detection model of wheat canopy chlorophyll content is determined by nested cross validation and parameter optimization.The experimental results show that,compared with linear regression,random forest and back-propagation artificial neural network models,Input8-RR has the highest R2 value of 0.838 and the lowest RMSE value of 2.825,which shows that the optimal model established in this paper can improve the prediction accuracy of canopy chlorophyll content detection model and effectively evaluate the chlorophyll content of wheat canopy.(5)For the problem of determining the level of nitrogen deficiency in the canopy scale of wheat in the field prototype,an feature selection and construction method based on genetic programming algorithm(GP)is proposed.Then,the wheat nitrogen deficiency level discrimination model is established by the new high-level image evaluation index set combined with C4.5,KNN and NB classification algorithms.The experimental results show that the feature selection(Fs)and feature construction(Fc)based on GP can improve the accuracy of wheat nitrogen deficiency classification model while reducing the dimension of inputs;the average classification accuracy of the optimal model Fc-NB is 75.35%,the highest classification accuracy is 85.42%,and the standard deviation is 5.77.It can meet the requirements of the evaluation of nitrogen deficiency in wheat. |
