| Tobacco is one of the important economic crops in china, with the development of the top ten scientific and technological projects for China’s tobacco industry, advanced science and technology has played a huge role in tobacco production and scientific research. In this process, the quality evaluation of tobacco has attracted the great attention of tobacco enterprises and research institutes. All along, the chemical composition of tobacco is the subject of the chemist of tobacco’s research, now the chemical ingredients in tobacco have been identified that is more than 3800. Therefore, timely and accurately grasp of the content of tobacco in the biochemical composition, has important practical significance for the guidance of tobacco production.In this study, the Guizhou karst mountain and Da Mo village of Luliang County in Yunnan province as the experimental area, the "Nanjiang No. 3" and the "Yunyan No.97" were viewed as the research objects, which is the significance cultivars of Guizhou and Yunnan. Based on the field experimental design, the spectrum were collected by a Field Spec 3 portable spectrometer under the leaves non-vitro condition, and the biochemical parameters in corresponding leaves were determined; On the basis of pre-processing for the measured spectral data, using spectral feature extraction technique, the spectral characteristics of different transformation were extracted; Firstly, the correlation of the biochemical parameters in tobacco leaves with themselves and the different spectral transform were analyzed, and the significance variable of high spectral were selected as independent variables, with tobacco main biochemical parameters as dependent variables,then an estimate model for tobacco main biochemical indicators was established by multivariate stepwise regression, the accuracy of the model was tested. Secondly, The dimensionality of hyperspectral data were reduced by using principal component analysis technique, and the principal component that contains the original information were extracted as independent, then an estimate model for tobacco main biochemical indicators was established by partial least-squares regression and the accuracy of the model was tested. Thirdly, in this study, the hyperspectral data were preprocessed, the nicotine & nitrogen content estimation model were established after resampling of sensors OLI of Landsat8 satellite, the combination the non-imaging hyperspectral with multispectral imaging sensor were tried. The main findings are as follows in the thesis:(1) In this study, the hyperspectral reflectance curves of tobacco leaves were consistent with the hyperspectral feature of the normally green vegetation leaves; after the preprocessed spectral data, such as removal of outliers, averaging, smoothing and denoising, can be better represent the actual spectral characteristics of tobacco.(2) Analyzed and compared for the spectral characteristics of tobacco found that, for different types of flue-cured tobacco, spectral reflectance in the visible band were closer, in the near-infrared band showed the Nanjiang No 3> Yunyan No 97; for the same tobacco, the spectrum presented the different rules for different growing parts for upper, middle, lower; for the spectral reflectance from different position(base, middle and tip) of leaves, due to the inhomogeneities of chemical composition of tobacco in the spatial distribution showed some differences, the specific performance is base> central> tip in this study; the spectral characteristics of different growth stages for the same tobacco varieties, that was maturity > growing in the growth period; if the external conditions was certain(such as light, air humidity, etc.), the minimum changes for the spectral reflectance with different time in a day were showed.(3) The correlation of biochemical parameters of the leaf showed: significant correlation were founded between total sugar and reducing sugar, nicotine and chlorine content was significantly correlated, potassium content and nitrogen content was significantly correlated.(4) According to the result of the correlation between the nicotine content & nitrogen content and the spectral data, the spectral data that corresponding the positive and negative correlation with the strongest were extracted, the estimation models of hyperspectral remote sensing were established by multiple regression, and the accuracy for each model were tested. The results showed that: the nicotine content & the nitrogen content in the multiple regression model, the regression model the coefficient of determination R2 based on logarithmic differential transform was the largest, respectively were 0.775, 0.746, with the high accuracy in estimating model. the estimation accuracy based on Hyperspectral Feature variables is not ideal, it was poor for nicotine content, for nitrogen content, the coefficient of determination was 0.545.(5) The principal component were extracted based on the hyperspectral data, and it were choosen that its cumulative variance contribution rate was over 85%, the estimating model were established for each spectral transform by partial least squares regression. The results showed that: in the all estimation model, the accuracy based on the logarithmic first-order differential transform for the nicotine & nitrogen content was the highest.(6) Matched the preprocessing spectrum to filter function of OLI sensor of Landsat 8 satellite, the average values of reflectance all the wavelength of the 8-band were extracted. Then used the correlation analysis between nicotine &nitrogen content and them, the results showed that: the correlation that nicotine content with each extracted band from Landsat 8 was poor, and failed to establish its multiple regression model, the estimation accuracy of PLS was poor. For the nitrogen content, the estimation accuracy based on stepwise multiple regression model and PLS was closer, the coefficient of determination was 0.779, 0.785. |
PDF Full Text Request