Unscrambler x 10

The calibrated hyperspectral images were then imported to the Environment for Visualizing Images (ENVI) software (ITT Visual Information Solutions, Boulder, USA) for image analysis. The size 135 × 150 pixel at the center of an image was chosen by the region of interest (ROI) tool. The reflectance spectra curve of pixels extracted from ROI regions were averaged to represent each sample.
For the purpose of eliminating noise of the spectral data and improve the predictive ability of samples, the preprocessing methods of Savitzky-Golay smoothing was used. The spectral data were calculated and processed by Unscrambler X 10.1 (CAMO Software, Norway). The preprocessed spectral data were used to establish FW-PLSDA, CARS-PLSDA and RC-LDA respectively.

In order to compare and obtain robust and reliable performance, spectral data pretreatment is necessary prior to the development of quantitative models. In this research, a series of pretreatments including normalization, standard normal variate (SNV), multiplicative signal correction (MSC), detrending (detrend), first-order derivative, and second-order derivative (1st and 2nd derivative) were applied in addition to non-preprocessed spectra. Multiplicative interferences of scatter in the spectra can be effectively removed using SNV approach. MSC is a method like SNV which is effectively used in multiplicative variations elimination. Derivatives were usually used to remove baseline offsets and separate overlapping absorption bands. In this research, derivatives were calculated using second-order polynomial with Savitzky–Golay smoothing by a moving window size of 15 points. Detrending was implemented combined with SNV to suppress the baseline shifting and curvilinearity. Normalization was utilized to present the spectral differences caused by slight optical path variations. The pretreatment or the combinations were all implemented in the Unscrambler X 10.1 (Camo Software Inc., Trondheim, Norway).