关键词: 近红外光谱, 蛋白质, 方竹笋, 间隔偏最小二乘法, 反向区间偏最小二乘法, 波段优选

Abstract:

Objective: To select the characteristic wavelength bands for non-destructive analysis of proteins in
Chimonobambusa quadrangularis shoot by FT-NIR spectroscopy combined with interval partial least squares (iPLS) and
backward interval partial least squares (BiPLS). Methods: The full wavelength range was divided into 12 or 17 bands. PLS
models were developed based on the full wavelength range or on each band and then BiPLS models were built by optimized
combinations of various intervals. The models were evaluated by the mean square error of cross validation (RMSECV) and
root mean square error of prediction (RMSEP). Results: The iPLS and BiPLS models were both more effective than the PLS
models based on the full wavelength range and the BiPLS model was the best one. When 12 intervals were adopted, the
combination of 7 ([5, 3, 6, 12, 4, 2 and 11]) spectral intervals provided the best model showing a 0.321% RMSECV and a
0.218% RMSEP. The proposed method allowed rapid and non-destructive determination of proteins with less variables used
in the modeling and thus improved prediction accuracy.

Key words: near infrared spectroscopy, protein, Chimonobambusa quadrangularis, interval partial least squares, backward interval partial least squares, wavelength selection