Quantitative Analysis Using NIR by Building PLS-BP Model

It is necessary to compress the input data for Back Propagation network (BP) with linear arithmetic to avoid the model over tally with calibration set, when a quantitative analysis of biology samples using NIR by building BP model is employed. At present Principal Component Analysis (PCA) and Stepwise Regression (SRA) have been widely used. The PCA computes the scores for the Principal Components (PCs) and uses these scores as ANN input. Since PCA can compress thousands of spectral data into several scores and describe the body of spectra, the training time has been shortened significantly and the model's prediction ability can meet the needs. But the compressing does not concern the relationship between the input variables and target output. The SRA cannot transform the spectrum data, but chooses the same from them according to the sorts of components. Among the variables selected there exists relativity and the efficiency of them depends on the sequence of the samples in the calibration set. Thus it's difficult to obtain the best selection and the BP model's prediction ability cannot often meet the needs. Partial Least Square (PLS) can compute the scores for the principal components (PCs) according to the sorts of components. In this study the scores computed by PLS have been applied as BP input and this BP model has been used to predict the contents of protein in 30 wheat samples. Compared with PCA-BP the PLS-BP model's prediction deciding coefficient (R2 ) has increased from 92.50 to 97.10 with its training iteration times decreased from 12 000 to 4500.