Dr. Md. Kamruzzaman
Associate Professor
Department of Food Technology and Rural Industries Bangladesh Agricultural University, Mymensingh-2202
Bangladesh, *E-mail : enamul@bau.edu.bd
M. R. Ali
Department of Farm Power and Machinery, Bangladesh Agricultural University, Mymensingh-2202, Bangladesh
M. E. Haque
Department of Computer Science & Mathematics, Bangladesh Agricultural University, Mymensingh-2202
Adulteration, Hyperspectral imaging, Wavelength Selection, Minced meat, Offal, PLSR
University College Dublin (UCD)
Quality and Nutrition
Minced lamb meat and lamb heart were purchased from local retail stores and transported to the laboratory of Biosystems Engineering, University College Dublin (UCD) for image acquisition. Firstly, lamb heart was cut into small pieces and then fully minced. Then, the minced lamb was weighed and evenly mixed with minced heart at a certain proportion to guarantee the adulteration percentage ranged from 2% to 36% with around 2% increment. Finally, the mixture was placed on a round metal lid for image acquisition. One sample for each adulteration level was prepared. Therefore, a total number of 18 samples were used in the study.. The line-scanned hyperspectral imaging system in the reflectance mode was used in the experiment . The imaging unit was composed of a 12-bit CCD camera, a high performance spectrograph (ImSpector N17E, Specim, Spectral Imaging Ltd., Finland) ), an illumination unit of two 500-W tungsten halogen lamps (V-light, Lowel Light Inc, USA), a translation stage (MSA15R-N, AMT-Linearways, SuperSlides & Bushes Corp., India), a data acquisition software (SpectralCube, Spectral Imaging Ltd., Finland). Each sample was placed on the translation stage and a hyperspectral image of the sample was acquired line by line. The acquired images were stored in a raw format before being processed. The image acquisition process was controlled by the SpectralCube data acquisition software (Spectral Imaging Ltd., Finland). Each image was recorded in the NIR region of 910-1700 nm with a spectral increment of about 3.34 nm between the contiguous bands, thus producing a total of 237 bands. After image acquisition, reflectance calibration was firstly performed on the obtained hyperspectral images of adulterated minced lamb using white/dark reference, as described by Kamruzzaman et al., (2013). Spectral data were extracted from the minced meat by ignoring background of the sample. A binary mask image was created by thresholding the image at 1300 nm with a value of 0.25. The mask was used as the main region of interest (ROI) to extract spectral data from the calibrated hyperspectral image. The reflectance spectrum from the ROI was computed by averaging the spectral value of all pixels in the ROI to produce only one mean spectrum for each sample. The same procedure was repeated to obtain the spectrum for all the tested samples. The extracted spectral data of each subsample were then arranged in spectral matrix (X) where the columns of this matrix represent the wavelengths (237 variables) and the rows of this matrix represent samples (18 samples). Partial least squared regression (PLSR) was applied on the full-wavelength range (910-1700 nm) for calibration and leave-one-out cross-validation (LOOCV). Partial least squares regression (PLSR) is a method for constructing predictive models where the factors are many in number and highly collinear. The core emphasis of this method is to predict the responses rather than identifying the underlying relationship between variables. PLSR method employs linear algorithm as there is a liner relationship between spectra and object properties (Kamruzzaman et al., 2013). Thus PLSR builds a liner model to predict a set of dependent variables, y (i.e. concentration of chemical attributes) from a large set of independent variables, X (i.e. predictors or wavelengths). Among many variables, there may be only a few underlying or latent factors that account for best predictive response (Wu et al., 2013). These orthogonal factors are called latent variables (LVs). These LVs are designed appropriately to capture most information in X as well as y. The PLSR analysis was carried out using the Unscrambler software v9.5 (CAMO AS, Trondheim, Norway). Performance of the regression models was evaluated using the root mean square errors of calibration (RMSEC), the coefficient of determination in calibration (R2 C), the root mean square errors estimated by cross-validation (RMSECV), and the coefficient of determination in cross-validation (R2CV). To reduce data redundancy and speed up analysis, the most pertinent wavebands were selected using successive projection algorithm. Successive projections algorithm (SPA) was proposed in (Araújo et al., 2001) for the optimum wavelengths selection. This method is a novel variable selection algorithm designed to solve the co-linearity problems by selecting variables with minimal redundancy. Nowadays SPA method has found its way with different modelling (Wu et al., 2012). Basically, SPA comprises three main stages. Initially, the algorithm builds candidate subsets of variables with minimum colinearity as the result of projection operations applied to the columns of the spectral matrix available for the training data. The second stage involves the evaluation of candidate subsets of variables based on the value of root mean square error (RMSE) obtained from the resulting multi linear regression (MLR) model assessed by cross-validation or by applying a separate validation set. The final stage consists of a variable elimination procedure aimed at removing uninformative variables without significant loss of prediction ability. Though SPA’s simple projection operations is quite effective, there are some drawbacks also. Variables selected by SPA may have a low signal-to-noise ratio (S/N) or be insufficient for multivariate calibration which can affect the precision of model prediction. When the developed model is optimized by selecting several important wavelengths/variables, the spectra can be applied in a pixel by pixel manner to obtain the distribution map where the components of the samples are clearly visualized and easily interpretable. The distribution map can be visually improved by employing post-processing routines.
J. Bangladesh Agril. Univ. 12(1): 189–194, 2014
Journal