Studies On The Diagnosis Mechanism And Modeling Of Leaf Color-Nitrogen Nutrition In Rice And Rapeseed Plant By Computer Vision

The nitrogen fertilizer is one of the most basic and important impact factors for crop yield and quality, and its application depend on crop nitrogen status. Accurate, real-time, non-destructive acquisition of crop nitrogen nutrition information is key to relealize rational fertilization. Crop nitrogen status was decided by its intrinsic physiological conditions, it also manifested through the leaves color. Leaf color change information is an important basis for crop nitrogen nutrition diagnosis. The computer vision technology is an effective way to acquire field crop nitrogen nutritional status. This dissertation focuses on the diagnosis mechanism and modeling of leaf color-nitrogen nutrition in rice and rapeseed plant by computer vision, and studies this topic from four aspects:the illumination problem, image analysis technology problem, sensitive color feature and space selection problem and modeling problem. It should lay foundation for development of portable and non-destruction measurement instrument of leaf nitrogen and new remote computer vision diagnosis platform.Illumination problem is the chief concern in crop leaf color-nitrogen nutrition diagnosis, including the brightness incosistences problems and color cast problems. For the brightness inconsistences problems of images illumination, one new color space called ten color model is defined based on HSV color space. We propose three illumination compensation algorithms, which are called shades of grey algorithm, ten color model and grey world algorithm and combination algorithm, respectively. It can improve the relevance of 0.05-0.10. For the color cast problems, we propose a color cast estimation algorithm based on RGB values, the accuracy rate up to 95.9%. We also propose two color constansy algorithms based on ten color model and histogram and two color cast correctin algorithms based on histogram shift and crop canopy statistical features. It can improve the relevance of 0.02-0.22.Image analysis problem is another key problem in crop leaf color-nitrogen nutrition diagnosis, including image segmentation and analysis problems. For the background and crop structure complex problem, we propose several segmentation algorithms, which include ten color combination algorithm (TCC), excess green combination algorithm (EGC), and so on, and have better performance than that by conventional segmentation algorithm (such as excess green algorithm). For the enclosed rectangle calculation problem of damaged, irregular and curved leaf blades, we propose a new algorithm called most suitable enclosed rectangle (MSER). Compared with minimum enclosed rectangle (MER), minimum perimeter enclosed rectangle (MPER) and major axes rectangle (MAR) algorithm, MSER can reduce angular error of 10-50%.Sensitvie color feature and space region problem is the other key problem in crop leaf color-nitrogen nutrition diagnosis, including the sensitive color feature and space region inconsistence problems. For the sensitive color feature inconsistence problems in the literature, we propose a new algorithm based on the maximum correlation. The several sensitive color features at different levels, different imaging conditions, different growth stages, different varieties and different crops were filtered and selected as Si*I1,B-Y,Diff, G-B,12,R-B,b*,r,DGCI/V DGCL,DGCV,13,2G-R-B and (2G-R-B)/L*. For the color spatial distribution analysis of crop leaf and plant canopy, we propose grid analysis algorithm and concentric circles analysis algorithm. The sensitive leaf space region and plant leaf position were dicided by correlation analysis which was taken among these nitrogen indexes and the spatial sensitive color feature. The sensitive region in rice leaf surface is at 2/5 leaf space from leaf pulvini. The sensitive region in rapeseed leaf surface is at 1/5 leaf space from leaf apex. The sensitive leaf position in rice plant is the top 3 leaf. The sensitive leaf position in rapeseed plant is the base leaf. The sensitive region in rice plant canopy is at 1/2 space from the main stem of the rice plant.Model construction is the kernel problem in crop leaf color-nitrogen nutrition diagnosis. For the main diagnosis modeling problems, we focused on analyzing the present series problems of the quantitative regression and qualitative digonosis modeling and putting forward corresponding solutions. The stability and generalization performance of linear regression, robust linear regression, nonlinear regression, multivariate regression, stepwise discriminant analysis, support vector regression, extreme learning machine regression, and neural network regression analysis methods were analyzed and compared. The linear and binomial regression with Si*I1(rice) or (2G-R-B)/L*(rapeseed) color feature was selected to construct the unified nitrogen content (SPAD, chlorophyll and dry weight-based N concentration (Ndw)) quantitative estimation model at leaves, plant and canopy level, with RRMSE of 1%-33%. Linear discriminant analysis (LDA), extreme learning machine (ELM), support vector machine (SVMKM and libSVM) and extreme learning machine (ELM) methods with r and DGCI/V color feature were selected to construct rice, rapeseed and both mixed nitrogen status qualitative diagnosis model. The SVMKM model is excellent, with accuracy of 95.1%,97.8% and 95.8%, respectively.Finally, the crop image analysis and data processing software with all foregoing methods was developed at Matlab platform.