Abstract:
Abstract: Tree peony seed has recently been introduced to produce a high-quality edible oil, rich in the green and organic nutritional ingredients. This study aims to explore the rapid detection for the content of moisture with the near-infrared spectroscopy (NIRS) in oil tree peony seed, and thereby to improve the accuracy of hyper-spectral estimation for the moisture content of peony seed oil. A specific modeling was developed to evaluate the moisture content in oil tree peony seeds, using the advanced hyper spectra technology. The near-infrared spectral reflectance measurements were used to collect the data in the wavelength of 350 to 2 500 nm using the spectrometer (SVC HR-1024i). An oven drying method was selected to obtain the moisture content of seeds. 156 samples were collected in total, two thirds of which were marked as the training set, and one third as the validation set. The constructed model was verified, according to the training set and the validation set. A systematic analysis was performed on the correlation between near-infrared absorption spectra, first derivative spectra, characteristic parameters of moisture absorption, and moisture content. The Single Linear Regression (SLR) models were established to evaluate the moisture content, according to the characteristic wavelength of absorption spectra, characteristic wavelength of first derivative spectra, and characteristic parameters of moisture absorption. Taking 2 characteristic wavelength first derivative spectra and 3 characteristic moisture absorption depth parameters as the input parameters, a BP Neural Network (BPNN) model was built, where the measured moisture content values were set as the output parameters. A Stepwise Multiple Linear Regression (SMLR) and Partial Least Squares Regression (PLSR) were used to simulate the moisture content, using the same input parameters. The predictive powers of SLR, SMLR and PLSR models were compared with that of the BPNN model. The results showed that: 1) The characteristic wavelength of moisture content absorption spectrum was located at 1410, 1900 and 1990 nm, and that of first derivative spectra was located at 1 150, 1 950 and 2 080 nm. 2) The moisture absorption characteristic parameters with the correlation coefficient greater than 0.9 were AD1930, AD2140 and AD1440. 3) The spectral characteristic variables for DF2080 (R=0.945) and AD2140 (R=-0.956) were significantly related with the moisture content values, and their linear models were achieved optimal for the better estimation models of moisture content. 4) In building BPNN model, the input parameters was set as the selected spectral characteristic variables during the single linear regression model using the hyper-spectral characteristic parameter variables, whereas, the moisture content values as the output parameters. The calibration and validation R2 of BPNN model for predicting moisture content were 0.978 and 0.973, the RMSE of 0.220% and 0.242%, the RPD of 6.478 and 5.889, respectively. Compared with other regression models, the BPNN model had the highest calibration and prediction accuracy. The SMLR model based on the selected spectral characteristic variables performed second to BP neural network. Furthermore, the SLR model was the simple method easy to operate. As such, the SLR model can be an optimal selection method under the condition of accurate moisture estimation, indicating that a real-time and high efficient method for the evaluation on the moisture content of oil tree peony seed. The finding can provide a sound theoretical basis to improve the remote sensing inversion accuracy of seed moisture content.