Abstract:In order to enhance the generality of near infrared spectroscopy model and solve the problem that the direct orthogonal signal correction algorithm may be over-fitting and unstable in the spectral processing process,a model transfer method,Random Forest-Direct Orthogonal Signal Correction(RF-DOSC),is proposed for regression analysis.The proposed method firstly uses the random forest algorithm to screen the near-infrared spectrum wavelength points,then conducts the direct orthogonal signal correction approach to perform spectral processing and establishes the regression equation.The regression coefficient is calculated by PLS to obtain the model transfer matrix.In the experiment,the near-infrared spectral data of corn(i.e.,D,D1,and D2 datasets were measured by S,S1 and S2 spectrometers,respectively.) was used to establish the transfer model.The prediction standard deviation(SEP) of water,oil,protein and starch components in the D1 set were 0.1267,0.0982,0.1569 and 0.4051,respectively;The SEP of the four components of the D2 dataset were 0.1548,0.0819,0.1366,and 0.3836,respectively,which were better than the results of conventional methods.It is clearly illustrated that the proposed model transfer algorithm could effectively eliminate the spectral noise;reduce the difference between the master and slave spectra.Meanwhile,improving the stability and accuracy of the model and realizing the sharing model between different instruments,as shown by the above experiment results.