|
Crop Nitrogen rapid, non-destructive estimation to improve the yield and quality improvement is important. NIR spectroscopy is non-destructive crop monitoring and accurate access to information provides an effective means. The purpose of this study is based on different years, different species, different nitrogen levels in wheat field trials, based on Fourier transform near-infrared spectrometer for major growth stages of wheat fresh, dry leaves and maturing grain spectral information, the use of partial least squares (PLS), BP neural network (BPNN) and wavelet neural network (WNN), to analyze and quantify it with leaf nitrogen content, sugar nitrogen ratio, the relationship between grain protein content, in order to achieve growth and quality of wheat NIR information Quick prediction model. First, using a variety of spectral pretreatment technology, the fresh leaves in 1155 ~ 1803 nm and 2118 ~ 2500 nm spectral region, the dry leaves in the 2046 ~ 2155 nm ,2297-2339 nm and 2343 ~ 2378 nm spectral region, with the PLS, BPNN and WNN methods are established the nitrogen content in wheat leaves quantitative monitoring model. The results showed that the optimum fresh and dry leaf pretreatment methods were MSC Savitzky-Golay second derivative and MSC Norris first derivative. For fresh samples, PLS model predictions root mean square error (RMSEP) and the coefficient of determination (R2) were 0.216% and 0.841; BPNN model RMSEP and R2 were 0.175% and 0.894; WNN model RMSEP and R2 respectively. 0.169% and 0.901. For samples of powdered dry leaves, PLS model RMSEP and R2 were 0.147% and 0.910; BPNN model RMSEP and R2 were 0.101% and 0.960; WNN model RMSEP and R2 were 0.094% and 0.978. From the model accuracy and robustness of view, the neural network method is relatively better than the PLS method; powdered dry leaves model outperforms the fresh model. Using the above method to establish a similar fresh and dry leaves of wheat sugar nitrogen ratio prediction model. The results showed that leaf spectral model to predict poor performance; dried leaves in 1655 ~ 2378 nm spectral region-wide adoption of MSC Norris first derivative pretreatment methods build models, better performance; based PLS, BPNN and WNN constructed of sugar nitrogen ratio estimation model, the root mean square error of prediction (RMSEP) were 0.332%, 0.292% and 0.288%, the coefficient of determination (R2) were 0.853,0.865 and 0.870. Further grain-based near-infrared diffuse reflectance spectroscopy Whole Wheat Protein build predictive models. The results show that for spectra multiplicative scatter correction combined with Norris first derivative in 1242 ~ 2230 nm spectral region to build models performed better; tests showed, PLS model RMSEP and R2 were 0.848% and 0.794, BPNN model RMSEP and R2 is 0.770% and 0.814, WNN model RMSEP and R2 are 0.761% and 0.816; neural network is better than partial least squares regression method. Finally, we discuss both fresh and dried leaves of wheat estimation of total soluble sugars and nitrogen feasibility. The results showed, PLS, BPNN and WNN three methods are not accurate for the simultaneous determination of total nitrogen and total wheat leaf sugar content, but the building of dry leaves WNN model predicts better, its root mean square error of prediction (RMSEP) were 0.101% and 0.089%, the coefficient of determination (R2) of 0.957 and 0.941, respectively; And, in convergence speed and prediction accuracy, WNN models are much better than BPNN and PLS model, its spectral pretreatment methods are: fresh leaves MSC Savitzky-Golay second derivative; dried leaves MSC Norris first derivative; modeled spectral region of 1100 ~ 2500nm.
|