Leaf Nitrogen Content Estimation Of Winter Wheat Based On Hyper-spectral And Satellite Imagery Data In Guanzhogn Region

Nitrogen is a major element for plant growth and yield formation in agronomic crops. Nitrogen deficiency leads to reduced photosynthetic capacity and yield of crops, and too much may cause farmland ecological environment pollution. Therefore, real-time monitoring and rapid diagnosis of crop nitrogen nutrition by remote sensing technique is a topic research in agriculture application research. In this article, field and plot experiments consisting of varied N fertilization levels and winter wheat varieties were set from year 2013 to 2015 in the National Agriculture High-tech Industrial Demonstration zone in Yangling, Shaanxi Province and its surroundings. Ground-based canopy spectral reflectance and leaf nitrogen content（LNC, %） of winter wheat were measured across the whole growth stages. The objectives of this paper were to explore the crucial technology and methods for canopy spectral data processing, sensitive band and feature selection, and method for model building. According to the sensitive bands and spectral parameters of leaf nitrogen content, kinds of LNC estimation models were proposed and compared with each other. Combining with multispectral satellite data, we tried to find the most effective approach to extract the planting area information of winter wheat from domestic high resolution satellite data, and exam the potential capabilities and limitations of multispectral satellite data to quantitatively estimate LNC over winter wheat fields at regional scale.The main results and achievements of this study listed as follows:（1） After the derivative transform, logarithmic transformation and continuum removal processing to original canopy spectral reflectance, the sensitive waveband regions and the most related band reflectance to LNC were obtained. The result showed that, the sensitive bands of original spectrum and logarithmic spectrum located at 400⁷20nm. The reflectances at wavelength 435⁴65nm,734⁷57nm,789⁷92nm, and 1176¹185nm were more sensitive to LNC changes in derivative spectrum. LNC could be best estimated by the reflectances at 400⁷60nm and 1181¹268nm in continuum removal spectra. Part of sensitive bands of logarithmic and continuum removal spectra were found to be more correlated with the LNC than other spectrums with the correlation coefficient great than 0.8. The exponential regression models were more suitable for the description of quantitative relationship between sensitive waveband reflectance of each transformed spectrum and LNC. Differences were existed in each regression equations and all the prediction precision were relatively poor.（2） Sensitivity of canopy spectral characteristics and parameters of winter wheat to leaf nitrogen content was systematically analyzed. The fitting equations were established to reveal the quantitative relationship of trilateral parameters, absorption features, spectral indices to LNC respectively. The results showed that, red edge position with modified linear extrapolation algorithm, ratio of red area and blue area, normalization of red area and blue area gave the most accurate estimation to LNC. The regression model derived from absorption features at 550⁷50nm performed well than the bands at 400⁵50nm. The total area of absorption peak at 550⁷50nm produced the lowest standard error of prediction on the test data among all the absorption features, also compared with trilateral parameters. But for high-values of LNC, the estimation precision of leaf nitrogen content still needed to be improved. Spectral indices significantly improved the accuracy of leaf nitrogen content estimation. The more prominent spectral indices included m SR705, SR705, GM, VOG3 and CI_{red edge}, which had the best fitting precision and small errors in models checking. All the possible combinations of bands within the 400¹350nm were selected to construct the ratio index, difference index, normalized spectral index and soil-adjusted vegetation index. The optimal spectral indices for LNC prediction were R⁷⁷⁰/R⁷⁰², FD⁷⁴⁴/FD⁵⁰⁴, LOG⁷⁵⁰-LOG⁷¹⁷and CR¹⁰⁵⁶/CR⁷⁰² of original, derivative, logarithmic and continuum removal spectrum respectively. The lowest prediction error（RMSE=0.06） and highest R2（0.89） were obtained with the ratio spectral index FD744/FD504, which significantly improve the estimation precision of high LNC values.（3） Estimates potential of wavelet transform analysis in leaf nitrogen content prediction was studied. The results showed that, continuous wavelet transform was not sensitive to the noise signal in canopy spectra. The relationship between the 4 transformed spectrum reflectivity and leaf nitrogen content was weak at lower scale than the higher scale, that is, the absorption characteristics reflected by detailed and high frequency signals may be caused by pigment content and other biochemistry composition in crop leaves, but not leaves nitrogen content. The 4th floor approximation coefficients of discrete wavelet transform based on partial least square regression model gave the best fit to LNC with variable counts 73. The R2 of regression equation with training dataset and test dataset, root mean square error（RMSE） values, the relative error probability（REP） were 0.94, 0.93, 0.04, and 2.72 respectively, which minimized the influence of soil and measurement backgrounds and thus enhanced the information of the sensitive wavebands, and gave the best fit result to LNC in crop plants.（4） Based on the ZY-1 02 C HR Image, three kinds of textures were extracted such as variogram texture, gray level co-occurrence matrix texture and spectral gradient texture and then a new SVM classification method with multi-source data by integrating the spectral information and 3 kinds of textures was presented to obtain the plant area of winter wheat. The study showed that, variogram texture and gradient texture involved in multi-source data effectively improved the image classification precision with an overall accuracy from 85.14% to 87.43% and Kappa coefficient from 0.82 to 0.85. As a limited spectral information, a little difference between textural features caused a notable classification difference. The variation function of absolute value form provided a theoretical basis for the optimal texture window analysis, and textural features based on the average lag distance could significantly improve the classification accuracy with an overall accuracy from 75.2% to 87.14% and Kappa coefficient from 0.7 to 0.87. Based on multi-source data, the SVM classification method of high spatial resolution remote sensing image could effectively overcome the fragmentation problem of traditional image classification. The result was significantly superior to MLC and DT with an overall accuracy of 89.14% and Kappa coefficient of 0.87. Different classification methods kept the extraction accuracy of winter wheat. The user accuracy and producer accuracy of winter wheat in SVM classification were 97.14%, and the Kappa coefficient was 0.96. The domestic resource satellite data had a certain stability and advantage over the extraction of thematic agriculture information. It could be wildly used in the identification of crop types and estimation of planted area.（5） Crop nitrogen content estimation by remote sensing technique had been being a topic research in remote sensing monitoring of agricultural parameters. Multi-spectral broadband reflectance data of six satellites were simulated using the measured hyper-spectral reflectances and spectral response functions of Landsat 8, SPOT6, HJ-1A, HJ-1B, GF-1 and ZY-3. Spectral indices derived from simulated broadband spectral reflectance data across the visible and near infrared bands were used to construct the LNC estimation models. The results showed that, there were no significant differences of simulated broadband reflectances and spectral indices between six satellite platforms. The spectral indices based on the simulated spectral reflectances were significant related with the leaf nitrogen content at 0.01 probability level with the correlation coefficients great than 0.6. A different pattern of the best combinations was found for six two-band spectral indices. The selection of 610⁶90 nm paired with 750⁹00 nm were the most effective two-band combinations in RVI index. It is also the center wavelengths of the red and near infrared bands for GF-1 satellite data. Considering the influence factors in evaluation the efficiency of a spectral index–LNC relationship: the stability of the spectral index（SI） with other perturbing factors, the sensitivity of the SI to a unit change of LNC, and the dynamic range of the SI, an improved sensitivity index-S was proposed based on the NE and TVI models. It indicated that all the regression models of selected spectral indices passed the significance test at 0.01 level. The TCARI/OSAVI and RVI indices linear related with LNC implied a stable response to the LNC changes. The TCARI/OSAVI and RVI indices performed well in accuracy test, and the RVI index was more excellent in remote sensing mapping based on the GF-1 satellite image. Taking all factors into consideration, The RVI index was regarded as the most suitable index for LNC estimation.