Determining lamb’s lettuce postharvest age based
on visible / near infrared reflectance spectroscopy
Bert A.J.G. Jacobs1, Bert E. Verlinden1a, Els Bobelyn1, An Decombel2, Peter Bleyaert2, Joris Van
Lommel3, Isabel Vandevelde3, Wouter Saeys4 and Bart M. Nicolai1,4
1Flanders
Centre of Postharvest Technology, Leuven, Belgium; 2Inagro, Rumbeke-Beitem, Belgium;
voor de Groenteteelt, Sint-Katelijne-Waver, Belgium; 4KU Leuven, Department of
Biosystems, MeBioS, Leuven, Belgium.
3Proefstation
Abstract
Lamb’s lettuce (Valerianella locusta L.) which is presented to the market is not
always freshly harvested. The product can be stored up to 28 days and is
indistinguishable from fresh material by the human eye. However, due to the prior
storage period the shelf life potential is limited and this leads to losses in distribution
and a lower quality for the consumer. This work aims to develop a rapid and nondestructive methodology using visible/near infrared (Vis/NIR) reflectance
spectroscopy to detect and quantify the postharvest age. Vis/NIR reflectance spectra
were linked to the time in storage by partial least squares regression (PLS). Two
variable selection techniques, Genetic Algorithms PLS and Monte Carlo Uninformative
Variable Elimination PLS, were combined to improve the accuracy and robustness of
the prediction model while decreasing the number of wavelengths used. The final
model used only 10% of the original wavelength variables while the root mean
squared error of cross validation decreased from 6.0 to 3.6 days. The final model was
tested using 2 external test sets and had a maximum root mean squared error of
prediction of 3.7 days. Therefore, it was concluded that Vis/NIR reflectance
spectroscopy can be a valid rapid and nondestructive method for identifying and
quantifying the postharvest age of lamb’s lettuce.
Keywords: storage, quality, corn salad, lambs lettuce, near infrared, spectroscopy,
multivariate statistics
INTRODUCTION
Lamb’s lettuce is a popular greenhouse vegetable thanks to its ease of use and readyto-eat character. It is used both as a leafy salad and as an ingredient in ready to eat salad
mixtures (Enninghorst and Lippert, 2003). However, lamb’s lettuce presented to the market
is not always freshly harvested. Depending on the season, it can be stored up to three weeks.
Stored samples are by eye visually indistinguishable from fresh produce, but they have
impaired shelf life potential (Rico et al., 2007).
To detect a lamb’s lettuce storage period, a fast and nondestructive measurement setup is needed which estimates how long a batch of lamb’s lettuce has been stored before it is
commercialized. Nondestructive measurements using visible / near infrared (Vis/NIR)
spectroscopy may provide this nondestructive method. The potential of NIR to characterize
and analyze fruit and vegetables has been shown before (Nicolai et al. 2008).
Due to the nature of NIR having broad overlapping peaks and a high correlation of
adjacent wavelengths, it is necessary to use multivariate statistical techniques to extract
useful information from the NIR spectrum (Nicolai et al., 2007). This is usually done by
means of partial least square regression (PLS). This statistical method uses the independent
variables, the Vis/NIR spectrum, to predict a dependent variable of interest. PLS defines new
a
E-mail: [email protected]
variables called Latent Variables (LV’s) based on the covariance between the dependent and
independent variables. In this way, a PLS model captures the variation in the Vis/NIR
spectrum which explains the dependent variable (Naes et al., 2002; Wold et al., 2001).
An important factor for the successful application of multivariate calibration models is
their robustness (Zeaiter et al., 2004). A calibration model which is sufficiently robust for
the specific application is based on a calibration dataset which is sufficiently rich in
variation. To test for the robustness of a calibration model, which is essential for a
successful application, an appropriate external validation is of prime importance (Bobelyn
et al., 2010). Improving robustness and model accuracy can be achieved by preprocessing
the spectra and applying wavelength selection techniques to remove any irrelevant
information which cannot be handled properly by PLS (Xiaobo et al., 2010).
Therefore, the aim of this study was to develop a fast and nondestructive methodology
to estimate how long a batch of lamb’s lettuce has been stored before it was presented to the
market. The potential of Vis/NIR reflectance spectroscopy in combination with multivariate
statistics were tested for this purpose. To achieve robustness the calibration models were
trained on a diverse calibration matrix after pre-processing of the spectra and selecting the
most useful wavelengths.
MATERIALS AND METHODS
Plant material and storage conditions
Samples of nine cultivars (Agathe, Audace, Baron, Calarasi, Cirilla, Gala, Pulsar,
Trophy, Palace) of lamb’s lettuce (Valerianella locusta L.) were harvested between
September 2012 and November 2014. Batches harvested before 2014 were grown at the
experimental garden Inagro (Rumbeke-Beitem, Belgium). Batches harvested in 2014
contained more diverse plant material from commercial growers. Different treatments were
applied during the postharvest period to induce extra variation.
Vis/NIR reflectance spectroscopy
From each selected lamb’s lettuce rosette, only the largest leaf was used for Vis/NIR
reflectance spectroscopy measurements. Adaxial and abaxial reflectance spectra (380 - 1690
nm, wavelength increment 2 nm) of the leaves were acquired using a Zeiss Corona 1.7 (Carl
Zeiss, AG, Germany) diode array (Si - InGaAs) with a 0/45° reflectance set-up using a fiber
optics probe (Nicolai et al., 2007). For each measurement, the leaf sample was placed
between a polished white PTFE block and the measuring head. Spectra were acquired at
least once a week for a period of 3 to 4 weeks. Each measurement point during storage new
samples from the same batch were used to minimize the effect of sample handling on the
quality of the samples.
Prediction model
1. Data matrix
Based on preliminary results de data matrix of the independent variables consisted of
the concatenated adaxial and abaxial spectra. The dependent variable was storage time after
harvest (days). Regions in the combined spectra where the noise was too high (380 - 418
nm) and where the detectors change (962 - 992 nm) were ignored.
For possible practical implementations it would be more interesting to have a cheaper
measurement set-up. Preliminary analysis showed a rise in R² from 0.81 to 0.86 when the
wavelengths above 1100 nm were discarded. Also a decrease in the RMSEC and RMSECV of
respectively 0.5 and 0.7 days was observed with the exclusion of wavelengths above
1100nm. Therefore, the analyses in this study were limited to the theoretical range of a Si
detector (380 - 1100 nm).
The 706 spectra acquired in January 2013 (193), July 2013 (79), November 2013
(138), February 2014 (164) and November 2014 (132) were used as calibration data. A
cross validation (CV) was applied to evaluate the PLS model performance during
construction. Each harvest period was used as a separate CV group. The 190 spectra
acquired in September 2012 were used as a validation set for variable selection to prevent
over fitting. An external test set for validating the final prediction models existed out of 165
spectra acquired in March 2013 (81) and May 2014 (84).
2. Variable Selection
Two wavelength selection techniques were applied on the spectra to improve the
prediction potential and robustness of the PLS model by removing wavelength variables
which were not informative for predicting the dependent variable. The applied wavelength
selection techniques were Genetic Algorithms PLS (GA-PLS) (Lucasius et al., 1994) and
Monte Carlo Uninformative Variable Elimination PLS (MC-UVE-PLS) (Cai et al., 2008).
GA-PLS is a combination of Genetic Algorithms with PLS. Genetic Algorithms are an
optimization technique that is inspired by Darwin’s theory of natural selection. Different
combinations of subintervals of equal size of the spectrum, called individuals, are made by
coding their chromosomes as a sequence of zeros and ones indicating for each variable or
interval of variables whether it is used or not (genes). The PLS model performance for these
individuals is then evaluated in CV. The best performing half of the individuals are then
selected to ‘breed’ new individuals by creating pairs of parents and exchanging sections of
their chromosomes in a single or double cross over. This new generation of individuals
consisting of the best half of the individuals from the previous generation and the newly
bred individuals is then again evaluated based on its PLS model performance in CV. This
process is repeated until a pre-defined fraction of the individuals share the same genes or
until a certain number of generations has been reached (Mehmood et al., 2012).
In MC-UVE-PLS, many PLS models are built using every time a different selection of
spectra/samples using a Monte Carlo sampling. We used 1000 Monte Carlo runs with 75%
of the samples selected. Each model has a different beta coefficient for each wavelength. A
reliability index (RI) for each wavelength is then calculated as the mean of its beta
coefficients divided by its standard deviation. (Cai et al., 2008; Mehmood et al., 2012). The
absolute value of the RI was the basis on which wavelengths were retained. Wavelength
variables with a low absolute value of RI were considered less important for good
predictions. The selection of the number of variables to retain was based on the RMSECV
and RMSEV of different PLS models with different amounts of the best wavelengths
retained. The maximum number of LV's in MC-UVE-PLS was set to 13 which is the same
number of LV’s used by the initial full spectrum model.
3. Preprocessing
Initially, the combined spectra were preprocessed using a multiplicative scatter
correction (MSC). MSC is a pre-processing step that attempts to account for offset and
scaling effects (Geladi et al., 1985). After variable selection a Generalized Least Squares
weighting (GLSW) was applied after a full spectrum MSC to further reduce the number of
LV’s. GLSW identifies interfering signals in the spectra and down-weights them (Martens et
al., 2003).
RESULTS AND DISCUSSION
Evaluation of initial model
The initial full spectrum PLS model was based on the spectrum of both adaxial and
abaxial leaf sides. The RMSEC and RMSECV were inconclusive for indicating an optimal
value of LV’s. There was only a local minimum when 8 LV’s were used, but the RMSECV gave
lower values with increasing LV’s and no real minimum was reached. Therefore, the optimal
number of LV’s was selected based on the estimated signal to noise ratio (S/N). An
estimated S/N greater than or equal to 3 is considered good. As the S/N for the 13th LV was
still good, the prediction model with 13 LV’s was selected. This model had an R² of 0.75 and
a RMSEC, RMSECV and RMSEV of 3.6, 6.0 and 5.4 days, respectively (Figure 1).
Figure 1. (A) RMSEC (open) and RMSECV (solid) of the basic PLS model with limited
preprocessing and no variable selection. (B) Estimated signal to noise ratios for
different LV’s. The horizontal line is the threshold (S/N = 3). Solid and open
symbols represent good (S/N ≥ 3) and bad (S/N < 3) signal to noise ratios
respectively.
Variable Selection
4. GA-PLS
The importance of each wavelength in the combined spectrum was determined by the
presence of each wavelength in the resulting 6226 models after 25 runs. The optimal
number of wavelengths was selected at 60% which means that 390 wavelength variables
were used for the construction of the PLS model (Figure 2A) This model used 13 LV’s and
had an R² of 0.78. The RMSEC, RMSECV and RMSEV were 3.3, 4.5 and 4.8 days, respectively.
The RMSECV of all the 6226 models which were constructed using GA-PLS varied
between 4.6 and 3.5 days. Chances of an overrepresentation of bad performing wavelengths
was a possibility. To cope with this problem a second evaluation was conducted using solely
the 10% best performing models of each replicate run based on the RMSECV. The optimal
number of wavelengths was selected at 15% which means 98 variables were used for the
construction of the PLS model. This prediction model used 10 LV’s and had an R² of 0.82.
The RMSEC, RMSECV and RMSEV were 3.1, 4.3 and 4.1 days, respectively.
5. MC-UVE-PLS
Different prediction models were constructed using different numbers of wavelengths
based on the absolute value of RI. The optimal number of wavelengths was selected at 35%
which means 228 wavelengths were used for the construction of the PLS model (Figure 2B).
This model used 14 LV’s and had an R² of 0.84. The RMSEC, RMSECV and RMSEV were 3.0,
3.7 and 3.9 days, respectively.
Figure 2. Output and selection of wavelengths of (A) GA-PLS, and (B) MC-UVE-PLS. On the left and the middle pane, the dashed line is the
mean spectrum, the thick solid regions are wavelengths selected for constructing the PLS model and the horizontal dashed line is the
threshold which discriminates between useful and useless wavelengths. The thin solid line in A is the presence of each variable in the
models constructed by the GA-PLS. The thin solid line in B is the rescaled absolute value of RI. On the most right pane, the RMSEC
(solid circles), RMSECV (open circles) and RMSEV (diamonds) of PLS models constructed using different numbers of wavelengths are
shown. The vertical line denotes the selected model for (A) GA-PLS, and (B) MC-UVE-PLS.
Evaluation of the added value of variable selection
When the different prediction models with a reduced number of wavelength variables
were compared with the initial model, it became clear that both wavelength selection
methods improved the RMSECV of the PLS models. The prediction model based on
wavelengths selected by GA-PLS had an RMSECV of 4.3 days and an RMSEV of 4.1 days
which implies that both models were quite robust. Of both techniques, MC-UVE-PLS was the
most successful, resulting in a PLS model with the highest R² (0.84), the lowest RMSECV (3.7
days) and the lowest RMSEV (3.9 days). Although only 35% of the initial wavelengths were
retained for model construction 14 LV’s were used for the predictions which was higher
than the 10 LV’s of the GA-PLS model.
Combining variable selections
The GA-PLS and the MC-UVE-PLS model used 98, and 228 wavelengths respectively.
There is a high consistency in the wavelengths which none of these wavelength selection
methods selected. 32% (206) of the initial 650 wavelengths were discarded by both
techniques. A combination of the wavelengths selected by both techniques might give even
better results than any selection made by a single technique. Therefore, a combination of the
selected wavelengths was made by combining all the selected wavelengths of both
techniques or by keeping only the wavelengths on which the two techniques were
unambiguous. In Table 1 the performance of both models based on these combinations is
presented together with the number of selected wavelengths. C1 was the only combinations
which performed as good as MC-UVE-PLS, but the number of included wavelengths was
lower. While the selection based on MC-UVE-PLS used 35% of the initial 650 variables,
combinations C1 used only 10%. To reduce the number of LV’s GLSW was applied with an
optimal threshold of 1 and 0.4 for MC-UVE-PLS and C1 respectively. These final models were
tested using an extra external test set. The root mean square error of this extra external test
set (RMSEP) was an extra indication for model robustness. All of the selected variable sets
gave similar results in RMSECV, RMSEV and RMSEP which indicates that these selections
gave robust PLS prediction models (Table 2). As both models performed similarly, the
model based on the lowest number of wavelength variables (C2) was selected (Figure 3 and
4).
Table 1. Performance of PLS models constructed by using combinations of wavelengths
selected by GA-PLS and MC-UVE-PLS.
Wavelengths selected
Number of LV’s R²
RMSEC
RMSECV RMSEV
wavelengths
(days)
(days)
(days)
Included by GA-PLS
98 (16%)
10
0.82 3.1
4.3
4.1
Included by MC-UVE-PLS
228 (35%)
14
0.84 3.0
3.7
3.9
C1: Included by both
65 (10%)
13
0.82 3.3
3.7
3.9
C2: Included by GA-PLS
261 (40%)
12
0.81 3.3
4.4
4.2
or MC-UVE-PLS
Table 2. The performance of 2 models using different number of wavelengths tested on data
from September 2012 (RMSEV) and data from March 2013 and May 2015 (RMSEP).
Name
of Number of LV’s R²
RMSEC
RMSECV
RMSEV
RMSEP
combination wavelengths
(days)
(days)
(days)
(days)
MC-UVE-PLS 228 (35%)
5
0.85 3.0
3.5
3.8
3.7
C1
65 (10%)
7
0.83 3.2
3.6
3.7
3.3
Figure 3. Time in storage after harvest plotted against the predicted time in storage. The
solid and open symbols are samples from the calibration and external test set,
respectively. The dashed line is the optimal regression line and the solid black
line is the regression line for C1.
Figure 4. The wavelengths which were included in selection C1. The dashed line is the mean
spectrum and the thick solid regions are wavelength variables selected for
constructing the PLS model.
CONCLUSIONS
Vis/NIR reflectance spectroscopy in combination with PLS regression was evaluated
as a fast and non-destructive method for the determination and quantification of a prior
storage period for lamb’s lettuce. The accuracy and robustness of the predictions improved
vastly after wavelength selection with a RMSECV, RMSEV and RMSEP of 3.6, 3.7 and 3.3 days
respectively. The number of LV’s dropped from 13 to 7, the RMSECV and RMSEV decreased
with 2.4 and 1.7 days respectively while the R² increased from 0.75 to 0.83. The number of
used variables was minimized by combining the output of GA-PLS and MC-UVE-PLS
resulting in 65 essential variables which made up 10% of the initial 650 variables. These
wavelengths in the Vis/NIR spectrum contained the essential information related to the
time in storage after harvest and had a good signal to noise ratio. It is still possible that
certain wavelength variables were influenced by external factors, but this could then be
corrected by other wavelength variables to prevent an incorrect prediction.
AKWNOWLEDGEMENTS
This research was carried out as part of IWT project 100885 supported by the Agency
for Innovation by Science and Technology in Flanders (IWT) and LAVA, Belgium.
Literature Cited
Bobelyn, E., Serban, A.-S., Nicu, M., Lammertyn, J., Nicolai, B.M., Saeys, W., 2010. Postharvest quality of
apple predicted by NIR-spectroscopy: Study of the effect of biological variability on spectra and
model performance. Postharvest Biol. Technol. 55, 133–143. doi:10.1016/j.postharvbio.2009.09.006
Cai, W., Li, Y., Shao, X., 2008. A variable selection method based on uninformative variable elimination
for multivariate calibration of near-infrared spectra. Chemom. Intell. Lab. Syst. 90, 188–194.
doi:10.1016/j.chemolab.2007.10.001
Enninghorst, A., Lippert, F., 2003. Postharvest Changes in Carbohydrate Content of Lamb ’ s Lettuce,
in: International Conference on Quality in Chains. Acta Horticulturae, pp. 553–558.
Geladi, P., MacDougall, D., Martens, H., 1985. Linearization and scatter-correction for NIR reflectance
spectra of meat. Appl. Spectrosc. 39, 491–500.
Lucasius, C.B., Beckers, M.L.M., Kateman, G., 1994. Genetic algorithms in wavelength selection: a
comparative study. Anal. Chim. Acta 286, 135–153. doi:10.1016/0003-2670(94)80155-X
Martens, H., Høy, M., Wise, B.M., Bro, R., Brockhoff, P.B., 2003. Pre-whitening of data by covarianceweighted pre-processing. J. Chemom. 17, 153–165. doi:10.1002/cem.780
Mehmood, T., Liland, K.H., Snipen, L., Sæbø, S., 2012. A review of variable selection methods in Partial
Least
Squares
Regression.
Chemom.
Intell.
Lab.
Syst.
118,
62–69.
doi:10.1016/j.chemolab.2012.07.010
Naes, T., Isaksson, T., Fearn, T., Davies, T., 2002. A user Friendly guide to Multivariate Calibration and
Classification 352.
Nicolai, B.M., Beullens, K., Bobelyn, E., Peirs, A., Saeys, W., Theron, K.I., Lammertyn, J., 2007.
Nondestructive measurement of fruit and vegetable quality by means of NIR spectroscopy: A review.
Postharvest Biol. Technol. 46, 99–118. doi:10.1016/j.postharvbio.2007.06.024
Nicolai, B.M., Verlinden, B.E., Desmet, M., Saevels, S., Saeys, W., Theron, K., Cubeddu, R., Pifferi, A.,
Torricelli, A., 2008. Time-resolved and continuous wave NIR reflectance spectroscopy to predict
soluble solids content and firmness of pear. Postharvest Biol. Technol. 47, 68–74.
doi:10.1016/j.postharvbio.2007.06.001
Rico, D., Martín-Diana, a. B., Barat, J.M., Barry-Ryan, C., 2007. Extending and measuring the quality of
fresh-cut fruit and vegetables: a review. Trends Food Sci. Technol. 18, 373–386.
doi:10.1016/j.tifs.2007.03.011
Wold, S., Sjöström, M., Eriksson, L., 2001. PLS-regression: A basic tool of chemometrics. Chemom.
Intell. Lab. Syst. 58, 109–130. doi:10.1016/S0169-7439(01)00155-1
Xiaobo, Z., Jiewen, Z., Povey, M.J.W., Holmes, M., Hanpin, M., 2010. Variables selection methods in
near-infrared spectroscopy. Anal. Chim. Acta 667, 14–32. doi:10.1016/j.aca.2010.03.048
Zeaiter, M., Roger, J.-M., Bellon-Maurel, V., Rutledge, D.N., 2004. Robustness of models developed by
multivariate calibration. Part I: The assessment of robustness. Trends Anal. Chem. 23, 157–170.
doi:10.1016/S0165-9936(04)00307-3