Research On Prediction Model Of Soil Organic-Matter Based On The Near-Infrared Spectroscopy Technology

In this paper,we use partial least square method(PLS),principal components analysis(PCA)and support vector machine regression(SVR)to establish soil organic content of the near infrared model and research on the calibration selection,spectral preliminary method,music pick,model parameter option.The results are as follows:(1)We established the models with the calibration from 146 soil samples that we collected and 79 samples which selected from spectral principal component analysis(PCA).By comparison of two model parameters,we found that it is very close between the correlation coefficient(R)and RMSEC of the two models.The conclusion is also suitable for cross validation.This fully shows that soil specimen is representative and may be able to build the sample of the calibration.(2)Through the comparison of the six different spectral preliminary combinations,we can find that the original spectra by “a derivative,SG convolution smoothing and multiplicative scatter correction ” of preliminary method to build the model prediction effect is best which is also better than untreated model.Using this preprocessed method,we can get the correction model's R is 0.9771 and RMSEC is 0.055.The R of the verification model parameters is 0.9508,RMSE is 0.084.(3)When we handle the spectral data by "a derivative,SG convolution smoothing and multiplicative scatter correction",we can establish soil organic content by using partial least squares regression(PLS)and principal components regression(PCR)method in the preliminary model.The model prediction effect is the best.When we take PLS and the principal component scores are 7.we can get the principal component of the standard model(R)is 0.9793 and the RMSEC is 0.0529.The correlation coefficient of the verification model(R)is 0.9653 and the RMSEP is 0.0695.The correlation coefficient of the cross validation(R)is 0.9268 and the RMSECV is 0.0986.Using principal component regression(PCR)build the predictive model.Its principal component scores is 13,the correlation coefficient of the calibration model(R)is 0.9560 and the inaccurate RMSEC is 0.0781.The correlation coefficient of the model validation(R)is 0.9229 and the RMSEP is 0.101.The correlation coefficient of the cross validation(R)is 0.9163 and the RMSECV is 0.107.In the research,we find that using partial least squares regression model is better than using principal component regression to build the model.(4)We disposed the original spectra by "a derivative,SG convolution smoothing and multiplicative scatter correction",then deal with the data between-1 and 1 by normalizing method.Through the research and analysis,we can select the kernel function for the radial basis kernel function.The optimal parameters that c is 8 and g is 0.00097656.Then the parameters of a support vector machine regression(SVR)model which has high precision is established.By using the model to predict training set and test set,we can conclusion that the training set predicted value and the measured value of the correlation coefficient is0.9709 and the root-mean-square inaccurate is 0.06907.The test set predicted value and measured value of correlation coefficient is 0.9460 and the root-mean-square inaccurate is0.1111.It shows that the nonlinear modeling method can also be used to precise analysis the soil organic matter,and the fitting curves are well consonant with the soil organic matter.