Application Of Chemometric Method On Extracting Hyperpectral Characteristics Information And Assessing The Growth Status Of Winter Wheat

The winter wheat is one of the most important crops in our nation. It is widely accepted that real-time, accurately and rapidly monitoring the growth status of winter wheat will facilitate the fertilizer application, irrigation management, regulation of the yield and quality. Currently, the hyperspectral technology of remote sensing has been widely used in agriculture field. The inaccurate spectral information and low predictive accuracy of monitor model limit its further application.The experiments were conducted under different varieties of winter wheat, diverse levels of nitrogen application and various planting years to study the application of hyperspectral remote sensing on monitoring the growth status of winter wheat by using the spectral analysis, multivariate method and chemometric theory. The main conclusions were made as follows:1 The data obtained from the experiment had an obvious difference and could represent the common growth status of winter wheat under different conditions. The significant relationships between other growth indicators and CGI (comprehensive growth indicator) constructed with these growth indicators by using the principle component analysis had been achieved and it indicated that CGI was feasible to character the growth status of winter wheat2 The phenomenon that the standard NDVI approached a saturation level under high above ground biomass (AGB) and leaf area index (LAI) was found in study. These wavelengths which were always used to construct the vegetation index suffered saturation at high vegetation and the saturated threshold of LAI and AGB were 3 and 1 kg m-2, respectively. It was proved that the spectral reflectance corresponding with the vegetation resulted in the saturation of NDVI and other Vis. The sutyd also demonstrated these VIs computed with the visible bands and near infrared bands also suffered the saturation. Furthermore, the performances of seven different VIs were assessed on overcoming the saturation. The results suggested that these VIs computed with the wavelengths in the red edge region as much as possible and had no upper limit might have a better predictive ability of LAI, AGB when the saturation happened. Moreover, the model constructed with the multivariate method might be an effective way to avoid the spectral saturation.3 The 9 points smoothing (SM9), the transformation of square root (T4) and the noise correction were selected as the optimal spectral preprocessing procedures.4 These wavelengths (400,512,536,555,680,700,735,760,816,890,920,1130,2040,2430 nm) extracted by using the multivariate analysis had been validated to be significantly sensitive to the growth status indicators of winter wheat.5 Non-linear model of SVM had a better performance than linear model of PLSR, and the linear model of PCR was the worst comparing with these models based on the whole spectroscopy. Among five different kinds of chemometric models, it showed that the CGI performed best and it also indicated that the CGI constructed with the PCA (principle component analysis) method could comprehensively represent the growth status of winter wheat. These models based on the non-linear SVM method were commonly accurate and robust, however, the model based on the PLSR and the T4 preprocessing method was better (Rc2=0.768, RPDC=1.973; Rv2=0.724, RPDV=1.693) than SVM model (Rc2=0.813, RPDC=1.945; Rv2=0.715, RPDv=1.554) when monitoring the CGI. Actually, the predictive accuracy for these models of PLSR+SMLR based on sensitive wavelengths had potential application in practice. Moreover, most of the monitor models based on the pre-processing spectra performed better than the raw spectra and it indicated that the pre-processing spectral technology could obviously improve the model performance. Comparing the effect of pre-processing methods on model accuracy, it was found that the transformation of square root (T4) was superior to other pre-processing methods.6 The procedure recommended to monitor the growth status of winter wheat using the hyperspectral remote sensing technology is:the raw spectral reflectance was preprocessed with transformation of square root algorithm→the correlation analysis between the preprocessed spectra and growth indicator→extracting the feature information by using the combination of PLSR and SMLR, and validating the extracted hyperspectral information→constructing the predictive models with the methods of PLSR, PLSR+SMLR and SVM→assessing the performance of these models and selecting the model with high accuracy and robustness as the optimal model to monitor the growth status of winter wheat.