Content uploaded by Arnaud Gucciardi
Author content
All content in this area was uploaded by Arnaud Gucciardi on Mar 14, 2022
Content may be subject to copyright.
Content uploaded by Arnaud Gucciardi
Author content
All content in this area was uploaded by Arnaud Gucciardi on Mar 14, 2022
Content may be subject to copyright.
Compact optical fluorescence sensor for food quality control
using artificial neural networks: application to olive oil
Arnaud Gucciardia,b, Umberto Micheluccia, Francesca Venturinic,a, Michela Spertid, Vanessa
M. Martose, and Marco A. Deriud
aTOELT llc, Machine Learning Research and Development, Birchlenstr. 25, 8600 D¨ubendorf,
Switzerland
bArtificial Intelligence Laboratory, University of Ljubljana, Ljubljana, Slovenia
cInstitute of Applied Mathematics and Physics, Zurich University of Applied Sciences,
Technikumstrasse 9, 8401 Winterthur, Switzerland
dPolitoBIOMed Lab, Department of Mechanical and Aerospace Engineering, Politecnico di
Torino, Turin, Italy
eDepartment of Plant Physiology, Faculty of Sciences, Biotechnology Institute, Campus
Fuentenueva s/n, 18071 University of Granada, Spain
ABSTRACT
Olive oil is an important commodity in the world, and its demand has grown substantially in recent years. As of
today, the determination of olive oil quality is based on both chemical analysis and organoleptic evaluation from
specialized laboratories and panels of experts, thus resulting in a complex and time-consuming process. This work
presents a new compact and low-cost sensor based on fluorescence spectroscopy and artificial neural networks
that can perform olive oil quality assessment. The presented sensor has the advantage of being a portable,
easy-to-use, and low-cost device, which works with undiluted samples, and without any pre-processing of data,
thus simplifying the analysis to the maximum degree possible. Different artificial neural networks were analyzed
and their performance compared. To deal with the heterogeneity in the samples, as producer or harvest year,
a novel neural network architecture is presented, called here conditional convolutional neural network (Cond-
CNN). The presented technology is demonstrated by analyzing olive oils of different quality levels and from
different producers: extra virgin olive oil (EVOO), virgin olive oil (VOO), and lampante olive oil (LOO). The
sensor classifies the oils in the three mentioned classes with an accuracy of 82%. These results indicate that
the Cond-CNN applied to the data obtained with the low-cost luminescence sensor, can deal with a set of oils
coming from multiple producers, and, therefore, showing quite heterogeneous chemical characteristics.
Keywords: Fluorescence spectroscopy; fluorescence sensor; olive oil; machine learning; artificial neural networks;
convolutional neural networks; quality control
1. INTRODUCTION
Olive oil contains many healthy components,1like vitamins, and has beneficial effects like antioxidant activity.2
Virgin olive oil, that is oil produced by the use of mechanical means only, is classified for commercial purposes
in three decreasing quality grades: extra virgin olive oil (EVOO), virgin olive oil (VOO), and lampante olive
oil (LOO). The best quality (EVOO) is of course the most expensive and beneficial, while the lowest grade
(LOO) should not be consumed without refining according to the European Union regulations.3,4Therefore, the
determination of the quality of olive oil is of the utmost importance. For this purpose, several approaches were
proposed in the past including, for example, crystal microbalance arrays,5metal oxide sensors6and spectroscopic
techniques in general.7–9Other approaches involve gas chromatography and mass spectrometry.10,11
Among the spectroscopic techniques, fluorescence spectroscopy is of great interest since it is a fast, cost-
efficient, and at the same time sensitive method to study the properties of vegetables, particularly olive oils.12,13
Send correspondence to F. Venturini: francesca.venturini@zhaw.ch
The presence in olive oil of natural fluorescent molecules such as chlorophyll and beta-carotene, phenolic com-
pounds, such as tocopherol, and their oxidation products make luminescence approaches very effective. The
most common techniques are the acquisition of excitation-emission matrices (EEMs) or the use of synchronous
scanning.14 Applications include (non-exhaustive list): classification of different quality grades of olive oils,15,16
adulteration detection,17–19 monitoring of the oxidation processes,7,20,21 deterioration monitoring,22 and geo-
graphical origin determination.23–25 The complexity in acquiring high-quality data (especially of EEMs) and in
data post-processing limit largely the accessibility of these methods.
This work presents an easy and non-destructive method to determine the quality of olive oil from a simple
single fluorescence spectrum. In particular, it describes an analysis of neural networks architectures to perform
olive oil classification when dealing with 1-D spectra obtained with the low-cost sensor for fluorescence spec-
troscopy described by the authors in a previous publication.26 The main contributions of this paper are three.
Firstly, a large dataset composed of 45 different olive oils of three qualities (EVOO, VOO, and LOO) from two
different producers and harvest years is presented. Secondly, three different neural network architectures are
discussed and their performance in classifying olive oils’ quality is compared. Thirdly, a new neural network
architecture is proposed to deal with the heterogeneous chemical properties of oils, even when having the same
declared quality.
2. MATERIALS AND METHODS
2.1 Oil Samples
In this study, 45 olive oils were investigated. 24 oils were provided by the producer Conde de Benal´ua, Granada,
southern Spain, from the 2019-2020 harvest and 21 from Producers of Eastern Mountains of the Province of
Granada, southern Spain, from the 2020-2021 harvest. The oils are of three qualities: extra virgin olive oil
(EVOO), virgin olive oil (VOO), and lampante olive oil (LOO). The quality assessment of all the olive oils was
performed according to the current European Union regulations3,4for the commercialization of olive oil. An
overview of the olive oils divided into the three quality classes and harvest year is reported in Table 1.
Quality class Number of samples 2019-2020 harvest 2020-2021 harvest
EVOO 20 10 10
VOO 13 8 5
LOO 12 6 6
Table 1. Overview of the olive oils samples in each quality class. EVOO: extra virgin olive oil, VOO: virgin olive oil, LOO:
lampante olive oil.
The oils of the two harvests have very different chemical and spectral characteristics, as it can be expected,
being the oils produced from two different harvests and different producers, although all from the same geo-
graphical region. This heterogeneity of the oils poses a learning challenge for the neural networks because the
spectra within one single quality class show very strong variations, sometimes stronger than between spectra
from different classes.
2.2 Experimental
The fluorescence spectra were taken with the portable and low-cost sensor described in a previous publication.26
The excitation light is provided by a UV LED with an emission at 395 nm. The fluorescence is collected by a
miniature spectrometer from Ocean Optics placed at 90◦with respect to the excitation LED. The resolution of
the spectrometer is 16 nm. The spectrometer has a 1024-element CCD array which acquires the entire spectrum
in one single measurement. The spectrum, therefore, consists of 1024 intensity values. The sensor can be seen
in Figure 1and was previously described in detail.26
All the measurements in this work were done on undiluted samples: for each oil, 4 ml were taken and placed in
clear glass vials (as visible in Figure 1), and then measured 20 times. From the acquired spectra the background,
or dark measurement, was subtracted, and the first 250 values, corresponding to wavelengths below 450 nm,
were removed since they did not contain any fluorescence signal but may contain spurious stray light from the
excitation LED. Finally, they were normalized to have an average of zero and a standard deviation of one.27
Figure 1. Photo of the fluorescence sensor an olive oil samples in the glass vials.
2.3 Machine Learning Models
In this work, three different types of architectures were investigated and compared to better understand the
variability between the oils and to improve the prediction accuracy. The architectures are 1) a feed-forward
neural network (FFNN), 2) a one-dimensional convolutional neural network (1D-CNN), and 3) a new type of
architecture called here conditional convolutional neural network (Cond-CNN).
The input to the FFNN and 1D-CNN is the spectra after the background subtraction and normalization
described in Section 2.2. Therefore, the input of the network is a one-dimensional row of 774 intensities. Both
the FFNN and 1D-CNN do not take into account if the oil sample is coming from the single producer Conde
de Benal´ua, harvest 2019-2020, or from one of the other producers, harvest 2020-2021. Therefore, they have to
try to extract features that are common to all samples, a task that may be very hard due to the heterogeneity
discussed in Section 2.1.
The new Cond-CNN architecture proposed here addresses this issue by adding as input also a class: 0 if the
sample is from Conde de Benal´ua and 1 if coming from other producers. The additional input branch of the
network is responsible for correcting the classification according to the producer of the input olive oil sample.
For all three neural network architectures, the cross-entropy loss function was chosen
L(ˆyi, yi) = −1
N
N
X
i=1
(yilog ˆyi+ (1 −yi) log(1 −yi)) (1)
where yiis the true expected class for the ith input spectra, ˆyiis the predicted class for the ith input spectra
from the network, Nis the size of the dataset used.
To evaluate the performance of the neural networks two additional metrics were monitored: the accuracy a
and the confusion matrix C.27 Defining the identity function I(x, y) as
I(x, y) = (1,if x=y
0,otherwise (2)
the accuracy aover a dataset of size Ncan be defined as
a=1
N
N
X
i=1
I(ˆyi, yi).(3)
The confusion matrix C∈R3×3can be also defined with the help of the identity function as
Ck,l =
N
X
i=1
I(ˆyi, l −1)I(yi, k −1) (4)
where l∈ {1,2,3}and k∈ {1,2,3}. The metrics reported in the results section are the mean of the accuracy
over the 10 splits indicated with ⟨a⟩and its standard deviation over the splits, indicated with σ(a).
To better judge the generalization capacity of the trained neural networks the dataset was split to include
the spectra of 35 oils (training dataset) and those of 10 oils (validation dataset ). Each model was trained on
the training dataset and its performance evaluated on the validation dataset, therefore, the validation is done
on spectra of unseen oils. To avoid overfitting an early-stopping strategy was adopted.27 The model that gave
the best performance on the validation dataset during the training was then saved and used. This process was
repeated 10 times (effectively performing 10 random splits) to be able to estimate the variance of the results on
different dataset splits. The number 10 was chosen as a compromise between running time and confidence of the
estimation. In the results section, the averaged confusion matrix over the 10 splits is reported, with the values
given in percent of the total number of inputs in the validation dataset.
For all the three neural networks, to deal with the light unbalance of the data, the loss function was minimized
by using class weights inversely proportional to the respective class frequencies.28,29 In all three cases the network
optimizer chosen was Adam.30 Note that overfitting was carefully monitored. In particular, by using early
stopping and drop-out,27 the architectures were optimised until the validation accuracy and training accuracy
were very close, meaning that the network outputs the same accuracy results on the training set and on the
validation test. In particular, all the results described here were obtained with models where the training accuracy
and validation accuracy are well within one standard deviation of each other. The details relevant to each of the
different architectures are given in Sections 2.3.1 to 2.3.3.
2.3.1 Feed Forward Neural Network Architecture
OutputDense Layers
.
.
.
.
Input 1:
Measured Spectra
.
.
.
.
.
.
.
.
Figure 2. Architecture of the FFNN used in this work. The yellow elements symbolize dense layers.
The architecture of the FFNN network used in this work is depicted in Figure 2, and consists of three layers,
each having 32 neurons. The ReLU activation function was chosen. A drop-out rate of 0.1 in each dense layer
was used. A mini-batch size of 64 was used and the network was trained for 500 epochs.
2.3.2 1D-CNN Architecture
The 1D-CNN network, depicted in Figure 3, is composed by the following sequence of layers:
Convolution
Layer
Max Pooling
Layer
Convolution
Layer
OutputDense Layers
.
.
.
.
Input 1:
Measured Spectra
Figure 3. Architecture of the 1D-CNN used in this work. The yellow elements symbolize dense layers, the blue elements
symbolize convolutional layers, the green elements symbolyze max pooling layer.
1. convolutional layer with 6 filters each of size 40;
2. max pooling layer with a size of 10;
3. convolutional layer with 6 filters each of size 20 (with a drop-out layer with a drop-out rate of 0.5);
4. a dense layer obtained flattening and concatenating the output of the previous layer;
5. dense layer with 4 neurons;
6. dense layer with 4 neurons;
7. dense layer with 3 neurons (output layer).
Also for this architecture the ReLU activation function27 was chosen. A mini-batch size of 64 was used and the
network was trained for 500 epochs.
2.3.3 Conditional 1D-CNN Architecture
The new network architecture proposed to address the heterogeneity of the oils is depicted in Figure 4. It is
characterized by two inputs: the measured spectra and the information of the producer or harvest year of the
oil. Note that this second input, here called producer class, is categorical, thus in the Python implementation its
value has been one-hot encoded.27 In other words, effectively the second input is a tuple that is (1,0) for Conte
de Benal´ua oils of the 2019-2020 harvest and (0,1) for all others oils.
The network used in this paper is composed by the following sequence of layers for respectively the top and
bottom branches in Figure 4:
1. top branch in Fig. 4- convolutional layer with 6 filters each of size 20;
2. top branch in Fig. 4- max pooling layer with a size of 10;
3. top branch in Fig. 4- convolutional layer with 6 filters each of size 10;
4. top branch in Fig. 4- a dense layer obtained flattening and concatenating the output of the previous layer;
5. top branch in Fig. 4- dense layer with 32 neurons;
6. top branch in Fig. 4- dense layer with 16 neurons;
Convolution
Layer
Max Pooling
Layer
Convolution
Layer
Output
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Dense Layers
Dense Layers
Input 1:
Measured Spectra
Input 2:
Producer Class (1,0) or (0,1)
.
.
.
.
.
..
.
Figure 4. Architecture of the Cond-CNN used in this work. The yellow elements symbolize dense layers, the blue elements
symbolize convolutional layers, the green elements symbolyze max pooling layer.
7. bottom branch in Fig. 4- dense layer with 16 neurons;
8. bottom branch in Fig. 4- dense layer with 36 neurons;
9. dense layer with 3 neurons (output layer).
For all the layers the swish activation function31 was used. For the results presented in this paper a mini-
batch size of 8 was used. The models were all trained for 3000 epochs. In this case no drop-out was used as
early stopping was enough to avoid overfitting.
3. RESULTS AND DISCUSSION
The three ANN described in Section 2.3 were tested by predicting the quality of unseen olive oils performing a
classification into the three classes EVOO, VOO, and LOO. The results illustrated with the confusion matrix
are shown in Figure 5. The confusion matrices present the accuracy of each network averaged over ten different
splits of the dataset and reported as % of the number of oil in each class.
Figure 5shows that the performance of the Cond-CNN exceeds that of the two other networks. In particular,
the Con-CNN can distinguish exceptionally well between edible (EVOO and VOO) and non-edible (LOO) oils.
Only 1% of the LOO oils were misclassified as VOO and only 2% of EVOO and 2% of VOO were wrongly
classified as LOO. The result is reasonable since the quality of the EVOO and VOO may be very close and differ,
for example, in the organoleptic properties, which are difficult to extract from a single fluorescence spectra.
EVOO
31% 3% 5%
5%
16%8%
3% 2% 27%
LOO
VOO
Predicted label
FFNN Cond-CNN1D-CNN
True label
VOO
LOO
EVOO
EVOO
40% 2% 4%
6%
17%8%
1% 2% 20%
LOO
VOO
Predicted label
True label
VOO
LOO
EVOO
EVOO
37% 4% 2%
2%
19%9%
0% 1% 26%
LOO
VOO
Predicted label
True label
VOO
LOO
EVOO
Figure 5. Confusion matrix visualising the accuracy in the prediction of the quality of olive oil using the three types of
neural networks discussed in this work. The value reported are given as % of the oils in each class and are averaged over
10 splits.
The performance is further illustrated by the calculation of the accuracy adefined in Equation 3. The results
for each of the ANN architectures are reported in Table 2.⟨a⟩indicates the mean of the accuracy, and σ(a) its
standard deviation when evaluated over the validation dataset for 10 splits.
Network Architecture ⟨a⟩σ(a)
FFNN 0.74 0.10
1D-CNN 0.77 0.09
Cond-CNN 0.82 0.06
Table 2. Performance comparison the different neural network architectures. aindicates the accuracy, ⟨a⟩its mean, and
σ(a) its standard deviation evaluated over 10 different splits.
The inferior performance of the FFNN and of the 1D-CNN is not surprising and can be understood in terms of
the heterogeneity of the oils. These two architectures do not take into account the different years of harvest and
different producers and, therefore, assume that all the 45 olive oils have similar and comparable characteristics.
The results indicate that this is not the case. Adding the information that there are two groups of oils (one
would probably need more than just two producer classes), the use of the Cond-CNN architecture increases the.
Due to the lack of specific information (e.g. cultivar, producer) of the 21 oils of the 2020-2021 harvest, it is not
possible to obtain a much higher accuracy than 82% obtained with the Cond-CNN architecture.
The high standard deviation is also related to the small number of oils used for the validation (10 oils). Since
the accuracy is an averaging metric, the central limit theorem (CLT)32 states that the smallest the sample size,
the largest the variance (and thus the standard deviation) of an averaging metric. It is possible to estimate what
is the expected standard deviation of the accuracy by using statistical arguments. In fact, one can note that the
accuracy, as written in Equation 3, is the average of a Bernoulli variable (I(ˆyi, yi)) with a probability p≈aof
being one. Thus, the average over nsamples of the above-mentioned Bernoulli variables will follow a Binomial
distribution. For binomial distribution the central limit theorem starts to be a reasonable approximation for
values of np =na > 1033 . In our case na ≈7 for FFNN and na ≈8 for the Cond-CNN. In this case, the
condition np =na > 10 is not really satisfied and an estimation from the CLT can only be considered a rough
approximation, but one that can let us understand if the results that were obtained are reasonable. It is easy
to show that, defining Ii=I(ˆyi, yi), one can write Var(Ii) = σ2(Ii) = a−a2therefore making the calculation
of the Var(Ii) very easy. Due to the small number of samples the variance should be calculated by using the
Bessel’s correction33 by using n−1 in the denominator for the calculation of the variance instead of n. The CLT
tells∗that the average of a sample of size nof the Ii(the accuracy) will have a standard deviation that when
n→ ∞ will tend to σ/√n.33 With the sample size that has been used here n= 10, the CLT gives an estimate
∗The CLT validity condition will not be discussed here, although it can be easily verified that they are satisfied.
of the standard deviation of the accuracy of the order of 0.1 (the exact values for the three cases reported in
Table 2are 0.13, 0.12 and 0.11 respectively). This matches the order of magnitude of the obtained results and
the fact that the standard deviation diminishes with increasing accuracy. The differences between this estimate
and the observed results are due to three main factors: the small sample size (therefore the CLT is not a good
approximation), the fact that aand, therefore, pare only approximation evaluated with a sample size of 10,
and the fact that σis also a rough approximation evaluated with a sample size of 10 (although corrected with
the Bessel’s correction). In conclusion, such a large standard deviation is not a reflection of the model being
unstable, but more of the small sample size.
4. CONCLUSIONS
This work presents a new compact and low-cost sensor based on fluorescence spectroscopy and artificial neural
networks that can perform olive oil quality assessment. Thanks to an innovative low-cost sensor and a novel
neural network architecture, the conditional one-dimensional convolutional neural network (Cond-CNN), it is
possible to perform a classification of olive oil into the three quality classes: EVOO, VOO, and LOO. The
method uses a portable, easy-to-use, and low-cost device, which works with undiluted samples, and without any
pre-processing and is, therefore, fast, simple to use, and cost-effective.
The performance of the new Cond-CNN architecture exceeds that of the feed-forward neural network and of
the one-dimensional convolutional neural network since can take into account the heterogeneity of the spectral
and chemical characteristics of the oils in the dataset. The accuracy obtained with this sensor is 82%. Most
of the errors in the classification occur for the distinction of EVOO and VOO, which probably have very close
chemical characteristics and may differ for properties not significantly affecting the fluorescence spectrum. If the
sensor is used to distinguish edible (EVOO and VOO) from non-edible (LOO) oils, the accuracy reaches 95%.
These results show still a large potential of fluorescence spectroscopy when combined with artificial neural
networks for the assessment of the quality of food and food technology in general.
ACKNOWLEDGMENTS
This work was supported by the projects: Innosuisse - Swiss Innovation Agency, Grant No. 36761.1 INNO-LS;
“SUSTAINABLE” funded by the European Union’s Horizon 2020 H2020-MSCA-RISE-2020 program, Grant No.
101007702; ”PARENT” funded by the European Union’s Horizon 2020 H2020-MSCA-ITN-2020 program, Grant
No. 956394; “Project of Excellence” from Junta de Andalucia-FEDER-Fondo de Desarrollo Europeo 2018. Ref.
P18-H0-4700.
REFERENCES
[1] Willett, W. C., Sacks, F., Trichopoulou, A., Drescher, G., Ferro-Luzzi, A., Helsing, E., and Trichopoulos,
D., “Mediterranean diet pyramid: a cultural model for healthy eating,” The American journal of clinical
nutrition 61(6), 1402S–1406S (1995).
[2] Kris-Etherton, P. M., Hecker, K. D., Bonanome, A., Coval, S. M., Binkoski, A. E., Hilpert, K. F., Griel,
A. E., and Etherton, T. D., “Bioactive compounds in foods: their role in the prevention of cardiovascular
disease and cancer,” The American journal of medicine 113(9), 71–88 (2002).
[3] “Commission regulation (eec) no. 2568/91 of 11 july 1991 on the characteristics of olive oil and olive-residue
oil and on the relevant methods of analysis official journal l 248, 5 september 1991,” Offic. JL 248, 1–83
(1991).
[4] “Commission implementing regulation no 1348/2013 of december 17 2013,” Official Journal of the European
Union 338, 31–67 (2013).
[5] Escuderos, M. E., S´anchez, S., and Jim´enez, A., “Quartz crystal microbalance (qcm) sensor arrays selection
for olive oil sensory evaluation,” Food Chemistry 124(3), 857–862 (2011).
[6] Garc´ıa-Gonz´alez, D. L. and Aparicio, R., “Classification of different quality virgin olive oils by metal-oxide
sensors,” European Food Research and Technology 218(5), 484–487 (2004).
[7] Baltazar, P., Hern´andez-S´anchez, N., Diezma, B., and Lle´o, L., “Development of rapid extra virgin olive oil
quality assessment procedures based on spectroscopic techniques,” Agronomy 10(1), 41 (2020).
[8] Violino, S., Ortenzi, L., Antonucci, F., Pallottino, F., Benincasa, C., Figorilli, S., and Costa, C., “An
artificial intelligence approach for italian evoo origin traceability through an open source iot spectrometer,”
Foods 9(6), 834 (2020).
[9] Mart´ın-Tornero, E., Fern´andez, A., P´erez-Rodriguez, J. M., Dur´an-Mer´as, I., Prieto, M. H., and Mart´ın-
Vertedor, D., “Non-destructive fluorescence spectroscopy as a tool for discriminating between olive oils
according to agronomic practices and for assessing quality parameters,” Food Analytical Methods , 1–13
(2021).
[10] Ruiz-Sambl´as, C., Cuadros-Rodr´ıguez, L., Gonz´alez-Casado, A., de Paula Rodr´ıguez Garc´ıa, F., de la Mata-
Espinosa, P., and Bosque-Sendra, J. M., “Multivariate analysis of ht/gc-(it) ms chromatographic profiles of
triacylglycerol for classification of olive oil varieties,” Analytical and bioanalytical chemistry 399(6), 2093–
2103 (2011).
[11] Taiti, C., Marone, E., Fiorino, P., and Mancuso, S., “The olive oil dilemma: To be or not to be evoo?
chemometric analysis to grade virgin olive oils using 792 fingerprints from ptr-tof-ms,” Food Control ,
108817 (2022).
[12] Kongbonga, Y. G. M., Ghalila, H., Onana, M. B., Majdi, Y., Lakhdar, Z. B., Mezlini, H., and Sevestre-
Ghalila, S., “Characterization of vegetable oils by fluorescence spectroscopy,” Food and Nutrition Sci-
ences 2(7), 692–699 (2011).
[13] Sikorska, E., Khmelinskii, I., and Sikorski, M., “Analysis of olive oils by fluorescence spectroscopy: methods
and applications,” Olive oil-constituents, quality, health properties and bioconversions , 63–88 (2012).
[14] Skoog, D. A., Holler, F. J., and Crouch, S. R., [Principles of instrumental analysis], Cengage learning
(2017).
[15] Guimet, F., Boqu´e, R., and Ferr´e, J., “Cluster analysis applied to the exploratory analysis of commercial
spanish olive oils by means of excitation- emission fluorescence spectroscopy,” Journal of agricultural and
food chemistry 52(22), 6673–6679 (2004).
[16] Poulli, K. I., Mousdis, G. A., and Georgiou, C. A., “Classification of edible and lampante virgin olive oil
based on synchronous fluorescence and total luminescence spectroscopy,” Analytica Chimica Acta 542(2),
151–156 (2005).
[17] Sayago, A., Morales, M., and Aparicio, R., “Detection of hazelnut oil in virgin olive oil by a spectrofluori-
metric method,” European Food Research and Technology 218(5), 480–483 (2004).
[18] Poulli, K. I., Mousdis, G. A., and Georgiou, C. A., “Rapid synchronous fluorescence method for virgin olive
oil adulteration assessment,” Food chemistry 105(1), 369–375 (2007).
[19] Ali, H., Saleem, M., Anser, M. R., Khan, S., Ullah, R., and Bilal, M., “Validation of fluorescence spec-
troscopy to detect adulteration of edible oil in extra virgin olive oil (evoo) by applying chemometrics,”
Applied spectroscopy 72(9), 1371–1379 (2018).
[20] Hern´andez-S´anchez, N., Lle´o, L., Ammari, F., Cuadrado, T. R., and Roger, J. M., “Fast fluorescence
spectroscopy methodology to monitor the evolution of extra virgin olive oils under illumination,” Food and
Bioprocess Technology 10(5), 949–961 (2017).
[21] Mishra, P., Lle´o, L., Cuadrado, T., Ruiz-Altisent, M., and Hern´andez-S´anchez, N., “Monitoring oxidation
changes in commercial extra virgin olive oils with fluorescence spectroscopy-based prototype,” European
Food Research and Technology 244(3), 565–575 (2018).
[22] Lobo-Prieto, A., Tena, N., Aparicio-Ruiz, R., Garc´ıa-Gonz´alez, D. L., and Sikorska, E., “Monitoring virgin
olive oil shelf-life by fluorescence spectroscopy and sensory characteristics: A multidimensional study carried
out under simulated market conditions,” Foods 9(12), 1846 (2020).
[23] Dupuy, N., Le Dr´eau, Y., Ollivier, D., Artaud, J., Pinatel, C., and Kister, J., “Origin of french virgin olive
oil registered designation of origins predicted by chemometric analysis of synchronous excitation- emission
fluorescence spectra,” Journal of agricultural and food chemistry 53(24), 9361–9368 (2005).
[24] Jim´enez-Carvelo, A. M., Lozano, V. A., and Olivieri, A. C., “Comparative chemometric analysis of fluores-
cence and near infrared spectroscopies for authenticity confirmation and geographical origin of argentinean
extra virgin olive oils,” Food Control 96, 22–28 (2019).
[25] Al Riza, D. F., Kondo, N., Rotich, V. K., Perone, C., and Giametta, F., “Cultivar and geographical origin
authentication of italian extra virgin olive oil using front-face fluorescence spectroscopy and chemometrics,”
Food Control 121, 107604 (2021).
[26] Venturini, F., Sperti, M., Michelucci, U., Herzig, I., Baumgartner, M., Caballero, J. P., Jimenez, A., and
Deriu, M. A., “Exploration of spanish olive oil quality with a miniaturized low-cost fluorescence sensor and
machine learning techniques,” Foods 10(5), 1010 (2021).
[27] Michelucci, U., [Applied Deep Learning - A Case-Based Approach to Understanding Deep Neural Networks],
APRESS Media, LLC (2018).
[28] Wu, Z., Guo, Y., Lin, W., Yu, S., and Ji, Y., “A weighted deep representation learning model for imbalanced
fault diagnosis in cyber-physical systems,” Sensors 18(4), 1096 (2018).
[29] Aurelio, Y. S., de Almeida, G. M., de Castro, C. L., and Braga, A. P., “Learning from imbalanced data sets
with weighted cross-entropy function,” Neural processing letters 50(2), 1937–1949 (2019).
[30] Kingma, D. P. and Ba, J., “Adam: A method for stochastic optimization,” Proceedings of 3rd International
Conference on Learning Representations, ICLR 2015, San Diego, CA, USA , 1–15 (2015).
[31] Ramachandran, P., Zoph, B., and Le, Q. V., “Searching for activation functions,” arXiv preprint
arXiv:1710.05941 (2017).
[32] Michelucci, U. and Venturini, F., “Estimating neural network’s performance with bootstrap: A tutorial,”
Machine Learning and Knowledge Extraction 3(2), 357–373 (2021).
[33] Casella, G. and Berger, R. L., [Statistical inference], Cengage Learning (2021).
