Based on Random Forest Regression NIR Wavenumber Selection and BP Neural Network Modeling

Abstract

The paper uses near-infrared spectral analysis to predict fat content of 99 corn samples. At first random forests regression is used to building the model, and its variable importance index(VIP) is used to filter wavenumber. Then selecting 20 from the 390 wavenumbers. In order to judge whether the wavenumber of the selected are applicable to other models, three methods of regression (decision tree regression and BP neural network and the random forest regression) are used to build models respectively in the whole spectrum and 20 optimal wave number. Finally 6 kinds of model are established, and after VIP selecting, by comparing the optimal models what is left is BP neural network model :r² is 0.985，RMSEP is 0.089.