| Nowadays, China has the world’s largest area of facility agriculture, with a total of3.3million ha. The yield and economic profits are serious affected because of nutritional stress and imbalance, such as nitrogen(N), phosphorus(P) and potassium (K) stress. Accurate monitoring and diagnosis of nutrient state in facility crops during the growth process has very important significance. Currently nutrition examinations at home and abroad are using only the strength of lightwave, usually focus on reflectance or reflection intensity. Reflectance spectroscopy technology or imaging technology is widely used in crop nutrition status diagnosis. In order to overcome the inadequacies of models build with single feature sauce, which cannot fully obtain the internal components and external morphological changes caused by crop nutrients stress, polarization information of light is add into this study using tomato as the object. A new method of polarized spectra-hyperspectral is proposed on nutrition stress of facility crops.The main contents of this paper are as follows:(1) Polarized spectra-hyperspectral multidimensional information detecting platform for crop nutrition comprehensive testing was developed. In order to accurate extract and research the polarized features of facility crops, a new type of polarized spectra acquisition and analysis system was designed, which can comprehensive detect crop leaf and obtain polarized response on different incident zenith angle and different detection zenith angle. Combined polarized spectra acquisition system with hyperspectral imaging system, polarized spctrum-hyperspectral detecting platform was constructed to realize multidimensional information detection of crop nutrition. N, P, K nutrient stress and control groups tomato plants were cultivated respectively via soilless cultivation method in greenhouse. N, P, K element contents were tested adopting Kjeldahl method, spectrophotometry, flame photometric analysis respectively to analyse nutrition distribution in the different growing stage. The results showed that the sample cultivation reached the desired effect. Successful cultivation of nutrition stress samples plants had laid a solid foundation for subsequent research.(2) Tomato N, P, K nutrient stress detection models based on hyperspectral image features were studied. Hyperspectral image of tomato leaves were obtain on experiment followed by image segmentation, filtering preprocessing, and principal component transform which effectively reduce dimensionality. Then weighting coefficient curves were drawn after analysis the correlation between the first five principal components and N, P, K nutrition contents respectively. Four sensitive wavelengths were identified, of which there were common wavelength566.29nm,693.71nm,733.71nm; and N, P, K specific sensitive wavelengths corresponding sequence is:464.91nm,474.85nm,762.24nm. Adopting the method of second-order probabilistic statistical filtering based on GLCM to extract texture feature under sensitive wavelengths. Though correlation analysis to find that the following image features showed high correlation with N, P, K, that were VAR.693.71,CON566.29,DIS693.71,ENT733.71,ASM566.29,COR733.71; as well as unique high correlation image features of N, P, K, that were:nitrogen ASM464.91,COR464.91; phosphorus HOM693.71,ENT474.85; potassium HOM762.24.ENT762.24.In the process of establishing a quantitative model, three modeling methods MLR, PCR and PLS were comparatively used. The results showing that PCR model of N elements was best, RMSECV=0.3846%, Re=0.9630, Rp=0.9205, RMSEP=0.4486%, PLS model of P element get optimal performance, RMSECV=0.5704%, Re=0.8864, Rp=0.8713, RMSEP=0.5420%. MLR model of K elements performance best, Re=0.9109, RMSECV=0.4163%, RMSEP=0.5047%, Rp=0.8547.(3) Tomato N, P, K nutrient stress detection models based on reflectance spectrum features were studied. Reflectance spectra were collected from hyperspectral of tomato leaves followed by SNV, MSC preprocessing, etc., and feature band screening methods such as iPLS, SiPLS, BiPLS and iPLS-GA were using to select sensitive wavelength of N, P, K respectively to build detection models. iPLS-GA model of N elements is best, Re=0.9156, RMSECV=0.595%, Rp=0.9048, RMSEP=0.632%. SiPLS model of P elements is best, Re=0.8765, RMSECV=0.592%, Rp prediction set=0.8740, RMSEP =0.512%; SiPLS model of K elements is best, Rc=0.9116, RMSECV=0.598%, Rp=0.9075, RMSEP=0.835%.(4) Tomato N, P, K nutrient stress detection models based on degree of polarization features were studied after polarized reflectance characteristics analysis. For a specific object of study, identify the optimal combination of detection angle:incident zenith angle60°, polarizer rotation angle45°, Detect polarize zenith45°, azimuth180°. Difference between different nutrient stress levels and tomato growing stage was compared after optimal combination of detection angle. To quantify the difference, degree of polarization features of different N, P, K nutrient levels of tomato leaf samples were calculated by Stocks formula, the sensitive wavelengths were extracted after analysis the correlation between degree of polarization features and the reference measurement of N, P, K respectively. Though correlation analysis to find that the following sensitive wavelength showed high correlation with N, P, K degree of polarization features, that were655.408nm,744.482nm,850.578nm, and N, P, K specific sensitive wavelengths that N, P, K were380.487nm,914.562nm,556.664nm successively. Degree of polarization sensitive wavelength characteristics as independent variables, the value of the chemical reference measurement as dependent variable to established MLR, PCR and PLS models. PLS model of N elements is best, Rp=0.9145, RMSEP=0.7299%; PLS model of P element is best, Rp=0.7846, RMSEP=1.1021%. MLR model of K is best, Rp=0.9009, RMSEP=0.7982%. This study provides a new idea for the rapid detection of tomato nutrient content.(5) Tomato N, P, K nutrient stress detection models based on polarized spectra-hyperspectral multidimensional information were first studied. The image features, spectral features and the degree of polarization features were full and accurate extracted as well as feature level fusion based on linear and nonlinear methods after normalized of the three sources of data information. Only the accuracy of N elements is higher than each single source feature model among the linear fusion models of N, P, K with MLR and PLS methods. Then BP-ANN and SVR methods were using to establish the nonlinear fusion models. The results showing that, the effect of BP-ANN model of N elements is best when principal component factor equal to6, Rp=0.9400, RMSEP=0.1995%. SVR-PSO Model of P element is best, Rp=0.8998, RMSEP=0.1912%; the effect of SVR-GS Model of K element get optimal accuracy, Rp=0.9101, RMSEP=0.1417. The results show that the polarized spectrum-hyperspectral multidimensional information detecting method to judge the tomato nutrient stress conditions is feasible.(6) Tomato N, P, K nutrient stress detection models considering interaction effect of the three elements based on polarized spectra-hyperspectral multidimensional information were first studied. The general linear fusion equation were modified via calculation of weight coefficient matrix and interaction effect coefficient matrix. Then tomato N, P, K nutrient stress detection models were established which considering interaction effect of the three elements. Take flowering and mid-fruiting stage data for example to calculate the predicted values of N, P and K nutrition content. On flowering stage, the results of N model were Rp=0.9585, RMSEP=0.2609%; the results of P model were Rp=0.9201, RMSEP=0.1739%; the results of K model were Rp=0.9194, RMSEP=0.2263%. On mid-fruiting stage, the results of N model were Rp=0.9461, RMSEP=0.2452%; the results of P model were Rp=0.9183, RMSEP=0.2616%; the results of K model were Rp=0.9144, RMSEP=0.2436%. The results showed that this method obtain higher accuracy models, especially the accuracy of P element was enhanced relatively large, which proved the existence of the interaction effect between the three elements. Decoupling of interaction effect of N, P and K can further improve prediction accuracy of the multi-information fusion models, and provides a new idea for the rapid detection of crop nutrient content in the case of considering interaction effect of nutrients. |
PDF Full Text Request