3.1. Spectral characteristics for the freshness detection
3.1.1. Prediction of freshness based on whole spectral wavelength
In this study, PLSR and SVR were used to build calibration models
to assess egg freshness using a full spectral range. The calibration
and prediction results obtained from various pre-treatment
routines were compared in Table 1. When PLSR was used to establish
a calibration model, the pre-treatment routines were
R2
p = 0.78, and RMSEP = 5.30% using ‘Autoscale’, which improved
predictive capacity compared to the original data (R2
p = 0.68,
RMSEP = 6.49%). When SVR was used to establish a calibration
model, the pre-treatment routines were R2
p = 0.85,
RMSEP = 4.33% using ‘PCA’. The pre-treatment processes improved
little predictive capacity compared to the original data, however,
the discrepancy between R2
c (1.00) and R2
p (0.85) in the original
data indicated that the model was over fit for the training samples,
resulting in a poor generalization ability. The above analysis suggests
that the SVR calibration model delivers better outcomes,
and the pre-treatment is effective and necessary to generate a suitable
set of data. To evaluate how accurate the model predicts the
value of a component, the root mean standard error value should
be compared with the standard deviation of the reference data
(Kamruzzaman et al., 2012). Therefore, calibration models were
evaluated according to the RPD. According to the reports by
Williams (2001), RPD value of 2.60 in SVR model suggesting that
the model can screen the egg freshness roughly. Kemps et al.
(2006) and Giunchi et al. (2008) reported the similar results but
with lower correlation coefficients for the measuring and the