Computers and Electronics in Agriculture 206 (2023) 107654

Contents lists available at ScienceDirect

Computers and Electronics in Agriculture

journal homepage: www.elsevier.com/locate/compag

Original papers

Computer vision-based smart monitoring and control system for food

drying: A study on carrot slices
Swathi Sirisha Nallan Chakravartula a, Andrea Bandiera a, Marco Nardella a, Giacomo Bedini a,
Pietro Ibba b, Riccardo Massantini a, Roberto Moscetti a
a
Department for Innovation in Biological, Agro-food and Forest System, University of Tuscia, Viterbo 01100, Italy
b
Faculty of Science and Technology, Free University of Bolzano-Bozen, Bolzano 39100, Italy

A R T I C L E I N F O A B S T R A C T

Keywords: Convective dryer embedded with computer vision (CV) system and load cell was used to continuously monitor
Process analytical technology carrot slices that are either unblanched or blanched (90 ◦ C for 2 min) during product drying (35 ◦ C, 35 % R.H., 3
Quality-by-design m s− 1 airflow). The CV system and load cell were selected as in-line Process Analytical Technology tools within a
Computer vision
proactive Quality-by-Design framework and embedded for, i) monitoring of product features (i.e., weight, colour,
Daucus Carota L.
Segmented-linear model
and size); and ii) developing shrinkage-dependent moisture prediction models using linear regression. The
evaluated shrinkage-dependent linear models showed superior performances (RMSE, 0.005–0.007) benchmarked
against selected thin-layer models of increasing complexity. The study tested a smart-enabled prototype dryer
with the potential for automation and integrating proactive quality strategies.

1. Introduction 12th sustainable development goal of ‘Responsible production and

consumption, SDG 12 (UN Report, 2019). As such, alternate strategies
A key element of food security is the uninterrupted availability of have been developed to improve drying systems that can be divided into
safe and quality products ensured by the agro-food sector through pre- two broad categories: (i) heat recovery, renewable energy-based and/or
and post-production processes like drying Consequently, the agro-food hybrid-drying technologies for improved energy efficiency; and (ii)
industry is critical to the global greenhouse gas emissions contributing implementation of sensors, and machine learning models for efficient
up to one-third of the emissions, more than half of which are related to product-process monitoring and control (Khan et al., 2020; Moses et al.,
drying operations (Crippa et al., 2021). Industrially, convection-based 2014; Zhang et al., 2017). Among these, the second strategy has been
drying is the most viable process for extending the shelf-life of foods gaining considerable interest both in academia and industry due to its
by minimising physicochemical changes and microbial spoilage (Mayor feasibility of implementation with minimal efforts compared to the
and Sereno, 2004; Moscetti et al., 2017). However, these drying systems scaling-up required in implementing new systems as in the first cate­
are criticised for their resource-intensiveness (energy, labour), high gory. However, these technologies are yet to have explosive industrial
environmental impact, and food quality issues that can be related to the acceptance and implementation due to major roadblocks like the un­
(i) dominance of fossil fuel-based dryers (~85 %) that contribute to certainty of application for a wide range of products-processes, inno­
approx. 12–25 % of the total energy consumption; (ii) heavy reliance on vation and financial lags, and long technological gestation periods
the operator’s experience in process control; and (iii) conventional among others.
reactive approaches of analyses for product testing which cause product In this context, it is of interest to investigate further the use of real-
quality failures i.e., under or over-drying (Aghbashlo et al., 2014; time optical tools like imaging systems with or without spectroscopy
Raponi et al., 2017) and thereby low energy efficiency (Menon et al., (Aghbashlo et al., 2014; Bedini et al., 2020; Liu et al., 2017; Raponi
2020; Raponi et al., 2021). et al., 2021) falling under the umbrella of Process Analytical Technology
To overcome the stated issues, the drying sector invests continuously (PAT) tools as monitoring and control systems (MCS). Specifically,
in innovative solutions that consider the complex, nonlinear and dy­ computer vision (CV) systems have been found to be reliable in the
namic nature of food drying while ensuring high-quality products, context of food processing offering cost-effective, non-invasive solutions
efficient processes, and sustainable production practices to meet the for monitoring product’s attributes like appearance, shape, and texture

E-mail addresses: [email protected] (R. Massantini), [email protected] (R. Moscetti).

https://doi.org/10.1016/j.compag.2023.107654
Received 7 September 2022; Received in revised form 20 December 2022; Accepted 11 January 2023
Available online 21 January 2023
0168-1699/© 2023 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-
nc-nd/4.0/).
S.S.N. Chakravartula et al. Computers and Electronics in Agriculture 206 (2023) 107654

quantifiable into relevant quality parameters (colour, shrinkage, 2. Materials and methods
texture) (Aghbashlo et al., 2014; Moscetti et al., 2021; Raponi et al.,
2021, 2017). Further, CV systems can be implemented as real-time data 2.1. Sample preparation
acquirement tools in-line for quality monitoring, analyses, multivariate
modelling, and feedback with the potential for proactively strength­ Carrots (Daucus carota L.) of the same maturity (~18-cm length)
ening and/or replacing conventional offline methods. This also offers were purchased from a local market (Ipercoop, Viterbo, Italy) and
the advantage of comprehensive and precise quality monitoring from immediately stored at 4 ± 1 ◦ C until further processing. Carrots free of
the start to the end of the process (Ghasemi-Varnamkhasti and Agh­ blemishes and defects were tempered to room temperature for 12 h,
bashlo, 2014). However, in practice, PAT tools alone might not washed, peeled, and cut into 5-mm thick slices. Samples were split into
completely render the process efficient due to the inherent variability of batches of 350 g of fresh weight before processing (~70 carrot slices).
food matrices and the dynamic nature of food drying, thereby necessi­ One batch of samples (BL) were blanched in water at 90 ◦ C for 2 min
tating strategic approaches that can allow for establishing pro-active and (Klarstein Slowcooker 550 W, Germany) to reduce the residual peroxi­
robust quality systems (Moscetti et al., 2019; Rathore and Kapoor, dase (POD) activity to less than 10 % based on our previous study
2016). (Moscetti et al., 2017). The control (CNT) batch was dipped in water at
Quality-by-Design (QbD) is one such approach that integrates room temperature for 2 min. Immediately after the treatment, the
product quality as a part of the process design by a systematic and dy­ samples were cooled using an ice bath, rested on a cotton cloth for 2 min
namic understanding of the product-process spaces and related changes to remove excess surface moisture, re-weighed, and subjected to drying
along the production line (Rathore and Kapoor, 2016; van den Berg tests.
et al., 2013). However, this integration is highly complex owing to the
simultaneous heat, mass, and momentum transport that occurs within 2.2. ‘Smart cabinet dryer’ set-up
drying at different scales, phases, and physical states that are preceded
and/or followed by physiological and biochemical changes (Defraeye, The prototype smart-enabled cabinet dryer set-up depicted in sup­
2014; Moscetti et al., 2019). These drying transitions although can be plementary Fig. 1 used in the drying tests consisted of (i) a prototype
generalised are considered highly product-process specific within a QbD temperature-controlled cabinet dryer developed by Innotech (Germany)
approach. Therefore, the desired product quality and the relevant with a drying chamber of 91-L holding a drying tray of 45x45-cm
quality characteristics, i.e., physical, chemical, or microbiological at­ dimension; (ii) a monitoring and control system (MCS) for tempera­
tributes within a given product design space should be identified, ture and relative humidity (mod. DICON touch, JUMO, Germany); (iii) a
defined, and monitored (Rathore and Kapoor, 2016; van den Berg et al., digital balance (mod. HT1500, NHU, Germany), placed on the bottom of
2013). the drying tray used during the tests connected with a RS232 cable
In fact, within the drying process, the product’s quality character­ pinout to an external personal computer; and (iv) a computer vision
istics and drying behaviour to ensure optimal process-end point as well (CV) system placed in a black box on the top of the drying chamber.
as the overall design and operations can be estimated with a wide range Specifically, the CV system consisted of (i) a CMOS camera (mod.
of macroscopic and microscopic measures. Among these, drying curve- DFK 33UX264, The Imaging Source Europe GmbH, Germany) equipped
based approaches, particularly thin-layer drying models, remain to be with a C1/1.2′ ’-8 mm-F/2.4 optical lens; (ii) an illumination source
the most popular owing to their simplicity of use and implementation for consisting of four 4200 K light-emitting diode strips or LEDs (mod.
model-based prediction (Buzrul, 2022; Defraeye, 2014; Erbay and Icier, LBRX-00–200-3-W-24 V, TMS Lite, Malaysia); and (iii) a DC 24 V power
2010). Most of these are time-dependent models with time as indepen­ supply controller (mod. ANG-4000-CH4-24 V-A1, TMS Lite, Malaysia).
dent variable and their robustness is affected by process parameters and The camera was mounted inside a black box to be insulated from
physical product characteristics such as shrinkage, bending, case hard­ external light and placed on the top of the dryer chamber. The four LED
ening, and surface cracking. Alternately, the use of product deformation strips were arranged on the bottom of the black box and along the top
as a model factor can allow for estimation of the drying behaviour side of the drying chamber at 90◦ angle to the product to improve the
considering product’s physical changes as evidenced in our recent intensity and uniformity of the light along the field of view, and the
findings (Raponi et al., 2021) as well as by other researchers by use of response of the camera.
imaging systems (Gulati and Datta, 2015; Purlis et al., 2021; Seyedabadi
et al., 2019). 2.3. Drying process and real-time data acquisition
In this context, the CV system was used within the dryer as a
component of the smart MCS (SMCS) that can allow for real-time data The hot-air drying experiments were performed at 35 ◦ C, 35 % R.H
measurements and subsequent model-based monitoring for process end- and 3 m s− 1 air flow based on preliminary trials. Each drying trial
point determination as well as product quality control. For testing the consisted of about 350 g of carrot slices dried for 36 h. Each treatment
set-up, convective drying of carrots (Daucus carota L.) was identified as a was replicated 4 times. The samples were subjected to offline analysis of
suitable product-process matrix owing to the wide industrial use of hot- moisture content at the end of the drying process according to AOAC
air dryers as well as the importance of carrot as a vegetable in the drying 934.06 official method.
sector. Thus, the objective of this work was to lay the foundation for a Weight loss and images as pre-selected product characteristics were
PAT-based QbD approach for carrot drying by (i) implementing a CV monitored and recorded in real-time i.e., when the drying process is in
system as a component of SMCS in a prototype drier for real-time progress. The in-line data was recorded in pre-set time intervals and
monitoring of product changes; (ii) to develop linear models utilising relayed to the connected computer with minimal time lag (5 min) be­
the in-line product changes (shrinkage) in predicting drying behaviour tween data recording and relaying. The in-line acquisition of weight loss
of samples subjected to common industrial pre-treatments like blanch­ and images of the product was performed every 5 min by using in
ing; and (iii) to evaluate the performance and robustness advantages of combination an ad-hoc Python3 script and the IC Capture software v.2.4
the CV-based shrinkage-dependent solutions benchmarked against (The Imaging Source Europe GmbH, Germany). The weight measure­
classical thin-layer methods for real-time drying monitoring and control. ments were stored as Comma-Separated-Values (CSV) files, while sam­
ple images were saved as Tagged Image File Format (TIFF). The
resolution of the digital image was 2192 × 1888 pixels with a colour
depth of 24 bits, corresponding to 8 bits per RGB channel. The schematic
for the process flow for monitoring the carrot drying with the use of
selected PAT tools is presented in Fig. 1.

2
S.S.N. Chakravartula et al. Computers and Electronics in Agriculture 206 (2023) 107654

Fig. 1. Schematic of smart dryer prototype embedded with in-line PAT tools - Load cell and digital camera for QbD-based real-time monitoring of carrot slice drying
(steps 1, 2 and 3).

2.4. Data handling and features extraction As for the in-line images acquired, they were utilised for features
extraction (i.e., colour and size information over time from samples).
The changes in dry-basis moisture content (MCdb) during drying Prior to feature extraction, the camera was (i) calibrated, to remove the
were computed by combining data from the in-line acquisitions (i.e., the lens distortion on the image; (ii) profiled, to correct colour distortion
fresh mass of the sample before drying; and mass changes during dry­ due to the camera sensor; and the images were (iii) flat-field corrected,
ing); and the off-line measurement of moisture content after drying. to remove shading distortion due to uneven illumination; (iv) and
Thus, the moisture ratio (MR) was computed by applying a min–max segmented, to separate the true image of each sample (i.e., foreground)
data normalisation to the MCdb, i.e., scaling values of moisture content from its background (i.e., non-sample data) (Fig. 2).
to a common 0–1 scale, without distorting differences among the Firstly, the camera calibration was performed using the ‘Camera
treatments (Raponi et al., 2021). Additionally, the drying rate at each Calibrator’ app in the ‘ComputerVision’ toolbox of Matlab 2019b
instance and for the overall process was computed and expressed as rate (MathWorks, CA, USA) by acquiring multiple images of a checkerboard
of moisture removed per gram of dry matter per hour (i.e., gH2O calibration pattern randomly placed on the top tray of the dryer. Sub­
gDW− 1h− 1). sequently, images were colour corrected using the ‘makecform()’ and

Fig. 2. Sequential steps in image correction and segmentation from raw image (sRGB) to the final corrected and segmented image.

3
S.S.N. Chakravartula et al. Computers and Electronics in Agriculture 206 (2023) 107654

‘applycform()’ functions of Matlab 2019b fed with the ICC camera product deformation as predictor instead of time were tested as an
profile computed through the Colorchecker Passport kit (X-Rite ltd., alternative approach. The segmented linear approach allows a predic­
UK). Subsequently, the flat-field correction was applied using the eq. (1): tive model to be optimised by identifying a breakpoint (BP) at an abrupt
change between the response (y) and the explanatory variable (x)
I = (I0 − D)/(F − D) (1)
(Lyndgaard et al., 2012). This was considered advantageous in the case
where, of area shrinkage (AS), as the partitioning of the influential factor into
‘I’ is the corrected image, ‘I0’ is the raw image, ‘D’ is the dark image intervals helps describe its relationship to MR during drying. In detail,
and ‘F’ is the flat-field image, respectively. The ‘D’ image corresponds to the BP was identified using the segmented() function of the Segmented R-
the ‘dark’ current acquired by covering the camera lens with a non- package to fit the standard linear model with segmented relationships
reflective opaque black cap. The ‘F’ image was acquired by scanning a between MR and AS; for the intended purpose, the ‘npsi’ attribute of
white-reference tile of 43x40-cm (H × W), used to capture unevenness in function (i.e., number of BR to be identified) was set equal to 1.
colour and brightness on the top tray of the dryer due to the illumination The model’s goodness of fit was evaluated in terms of Root Mean
system. Square Error (RMSE), Mean Absolute Error (MAE), systematic error
Further, the image segmentation was performed as a last stage of the (BIAS), reduced chi-square (red. χ2), and adjusted coefficient of deter­
image post-processing using an HSV filter (0.0 ≤ H ≤ 0.1; 0.3 ≤ S ≤ 1.0; mination (adj-R2).
0.5 ≤ V ≤ 1.0) written in Matlab. Finally, the features extraction (i.e.,
colour and spatial data extraction) was performed using a script written 2.6. Data handling and statistical analysis
in Python v3.6.9 coupled with the OpenCV 3.4.2 library. Specifically,
the Python script was used to extract from each carrot slice (i) the One-way analysis of variance (ANOVA) was performed to evaluate
average colour in terms of lightness (L*), redness (a*), yellowness (b*), and compare the main effect of the treatment factor (CNT and BL) on
hue angle (h) and chroma (C*) (Moscetti et al., 2017); (ii) the shape (i.e., parameters and performance metrics of each model, as well as to iden­
minor and major axis lengths, and eccentricity); (iii) the surface area (in tify the best model in terms of performances. Tukey’s pairwise com­
pixels); and (iv) the relative area shrinkage (AS), computed by dividing parison method was performed, and the Honestly Significant Difference
each surface area value over time by its initial value (Raponi et al., (HSD) was calculated for an appropriate level of interaction (P ≤ 0.05).
2021). Scripts for data handling, breakpoint identification for the development
of linear-segmented models, modelling and ANOVA were prepared using
the R software v3.6.9 in combination with ‘dplyr’, ‘segment’, and
2.5. Computer Vision-based moisture prediction models
‘agricolae’ R-packages.

Moisture ratio (MR) prediction models were computed for each

3. Results and discussion
treatment (i.e., CNT and BL), evaluating two different mathematical
approaches (Table 1): (i) time- dependent models i.e., thin-layer models
3.1. In-line quality parameters
using changes in MR as a function of drying time (t); and (ii) shrinkage-
dependent models i.e., the linear and segmented-linear models using
Moisture content and drying rates
changes in MR as a function of area shrinkage (AS). Specifically,
Fig. 3a shows the change in moisture content (dry basis, MCdb) along
considering diffusion as the dominant mechanism of moisture removal
the experimental drying times for control (CNT) and blanched (BL)
in carrots with drying predominantly taking place in the falling rate
carrot slices determined by the in-line weight changes.
period (Aghbashlo et al., 2009; Erenturk and Erenturk, 2007; Krokida
The initial MCdb for either BL or CNT treatments ranged from 8.90 to
et al., 2003; Planinić et al., 2005) the four most frequently used thin-
9.10 g g− 1 which decreased to a final equilibrium moisture content
layer drying models of increasing complexity levels were chosen for
(MCe) of 0.13 g g− 1 along 36 h of drying. The loss in moisture followed a
the time-dependent approach (Onwude et al., 2016). Although
pseudo first-order kinetic, irrespective of the pre-treatment. In detail,
frequently used to describe the moisture changes in food drying, thin
the moisture content decreased rapidly at the initial stages up to ~ 8 h to
layer drying models have boundary conditions which depend on the
MCdb values of 1.60 (BL) to 2.06 (CNT) g g− 1 followed by a slower and
drying method, process condition, and product dimensions and geom­
consistent moisture loss moving towards equilibrium. Interestingly, a
etry. Among them the critical ones are that the shrinkage effect during
significant difference in moisture loss was observed after 7–8 h between
drying should be negligible, and the product size should be homoge­
CNT and BL samples. This can be attributed to the blanching effects, as
neous and isotropic (Onwude et al., 2016). These requirements are
the samples were cut to the same thickness and dried under similar
difficult to fulfil in most food matrices, including carrots due to the
conditions. In fact, it is well known that blanching leads to damage of the
obvious product deformation and heterogeneity in drying.
cell membranes thereby influencing the cell wall permeability as well as
To circumvent the stated issues, dynamic shrinkage-dependent
the effective moisture diffusivity resulting in lower moisture content as
models specifically, linear, and segmented linear models which use
observed in carrots and other fruits and vegetables (Fellows and Fellows,
2017; Planinić et al., 2005).
Table 1 Further, the average drying rates as shown in Fig. 3b rose rapidly as
Time-dependent and time-independent models tested for the drying of carrot
expected to 1.35 and 1.39 g H2O g DW− 1h− 1 in CNT and BL, respec­
slices.
tively, within the first hour of drying due to the initial warm-up period
Model type Model Equations and consequent increase in the product’s surface temperature (Raponi
Time-dependent Newton-Lewis y= exp(− kt) et al., 2021; Wu et al., 2014). Subsequently, up to approx. 11 h of pro­
(Thin layer) Page y= exp(− (ktn)) cessing the drying rate decreased sharply registering only a falling rate
Handerson & Pabis y= a exp(− ktn) period and reached a plateau after about 24 h with the MCdb tending
Logarithmic y= a exp(− kt) + c
towards MCe. No constant rate drying period was observed, regardless of
Time-independent Simple linear y = mx + q treatment. This behaviour confirms that moisture diffusion was the
Linear segmented x > BP, y = m1x + q1 predominant mechanism for moisture removal which further limits the
x ≤ BP, y = m2x + q2 drying rate, as is the case for most porous, cellular structured agro-food
MR, Moisture Ratio; k, drying constant (h− 1); t, drying time (h); n, a and c model materials. These observations on drying behaviour were found to be
constant (dimensionless); m, m1 and m2, slope; q, q1 and q2, intercept; AS, consistent with other studies on convective as well as infrared or mi­
relative area shrinkage; BP, break point. crowave drying of carrots in form of cubes and/or slices (Aghbashlo

4
S.S.N. Chakravartula et al. Computers and Electronics in Agriculture 206 (2023) 107654

et al., 2009; Moscetti et al., 2017; Toǧrul, 2006; Wu et al., 2014) as well
as in other fruits and vegetables (Lahsasni et al., 2004; Raponi et al.,
2021). As for the different treatments, there were no significant differ­
ences between the drying rates of CNT and BL, even though blanching
treatment tended to have a slightly higher rate of moisture loss due to its
well-known effects on the cellular structure.
Shrinkage
Fig. 3c shows the general tendency of the relative area shrinkage
(AS) of carrot slices along the drying time with significant differences
between CNT and BL samples. The grey and red shaded areas along each
response line in graph represent the experimental dispersion of mea­
surements for CNT and BL samples respectively. Both samples exhibited
pseudo first-order kinetics similar to that of MCdb (Fig. 3a) with rapid
decrease in AS from an initial value of 1.00 to about 0.50 in CNT and
0.36 in BL samples within 8 h of drying time. Subsequently, the decrease
in AS values gradually slowed, reaching a plateau around 24 h, generally
proportional to the loss of MCdb. In fact, previous studies showed that
carrots similar to that of potato and apple tissues had the ideal shrinkage
behaviour with the shrinkage being linearly related to that of moisture
loss (Dhall and Datta, 2011; Raponi et al., 2021). This proportionality
has been attributed to the stress created due to loss of moisture and other
contributing factors like sample dimensions, porosity, and material state
(Madiouli et al., 2012; Mayor and Sereno, 2004; Nguyen et al., 2018).
However, it is necessary to note that the linearity of the shrinkage at
very low moisture contents might be disrupted due to the continuing
change in the tissue’s material state and mechanical characteristics. In
particular, the transition of material state from rubbery to glassy states
was evidenced to result in formation of relatively rigid structure which
was evidenced to limit the rate and degree of shrinkage (Mayor and
Sereno, 2004; Nguyen et al., 2018).
Furthermore, it is known that the shrinkage behaviour is also
affected by various factors like pre-treatments and drying conditions,
which eventually influence the moisture diffusivity and drying rates. As
previously noted, BL samples registered steeper shrinkage values
significantly different from the CNT after 2 h of drying (marked with
green vertical line and asterisk in Fig. 3c) up to the final drying of 36 h
with the final relative area shrinkage being 0.208 for BL compared to
0.260 for CNT samples. These differences observed in BL samples can be
attributed to the modifications caused by blanching in the cellular
construct and the gas expulsion from the carrot tissue. The cell mem­
brane and cell wall breakages because of blanching create pores (i.e.,
improved porosity) which improves moisture diffusivity and thereby the
drying rates resulting in a higher tissue contraction (González-Fésler
et al., 2008; Madiouli et al., 2012). Also, the gas expulsion decreases the
instances of intercellular spaces causing the matrix to collapse which
was more evident in BL than CNT samples even under mild drying
conditions (Aprajeeta et al., 2015). In fact, it was observed by Lemmens
et al., (2009) that high temperature blanching can result in weaker
cellular structure susceptible to damage during subsequent thermal
treatments due to further chemical demethoxylation compared to that of
unblanched or mildly blanched (60◦ C or less) samples.
Surface colour changes
Colorimetric parameters (CIELab), as an important visual aspect for
dried food products critically impacting the purchase decisions, were
monitored along the drying process by the CV system. The samples as
expected exhibited changes in the CIELab coordinates (L*, luminance;
Fig. 3. In-line monitoring of a) dry basis moisture content, MCdb; b) drying a*, redness; b*, yellowness) along the drying time represented by
rate, DR; and c) relative area shrinkage, AS along 36 h of drying of control luminance (L*) and hue angle (h) in Fig. 4a and 4b, respectively. In
(CNT) and blanched (BL) carrot slices dried at 35 ◦ C, 35 % R.H. and ~ 3 m s− 1 general, these colour changes in vegetal matrices can be attributed to the
air velocity. Results are reported as mean and standard deviation of the mean.
extensively studied multifaceted effects of drying as well as the pre-
The vertical green line with the * symbol at the top represents the initial
treatment (blanching) on the material state and properties like water
moment at which the CNT and BL values become statistically different (p ≤
0.05). (For interpretation of the references to colour in this figure legend, the activity and phase transition, cellular shrinkage and/ or disruption,
reader is referred to the web version of this article.) enzymatic and/or non-enzymatic reactions as well as pigment degra­
dation and/or concentration.
In terms of luminance (L*), the values were observed to initially
increase and then continuously decrease, concurrent to the loss of MCdb

5
S.S.N. Chakravartula et al. Computers and Electronics in Agriculture 206 (2023) 107654

general decrease in the L* value, it can be considered as a direct

consequence of moisture loss during drying (Fig. 4b) which resulted in
lower refractive indices similar to dried apple and potato slices (Moscetti
et al., 2018; Onwude et al., 2018; Raponi et al., 2021).
Regarding the hue angle (h), a continuous decrease in values was
observed (Fig. 4b) for CNT from 53.35 to 42.35 and BL samples from
52.52 to 37.47 during 36 h of drying. The initial values reflected orange-
red hues for both evaluated samples, whereas those of dried carrots
tended towards intense red–orange hues with significant differences
between the CNT and BL samples. This decrease in hue values can be
attributed to the moisture loss and carotenoid degradation during drying
(Liu et al., 2016; Moscetti et al., 2017). Interestingly after 6 h of drying
when the relative MCdb reached approx. value of 4.57 g g− 1, the BL
samples registered a drastic drop in the hue values and continued to be
significantly lower than CNT samples all through the drying. This
observation was in concurrence with the trends of MCdb (Fig. 3a) and AS
(Fig. 3c), with the critical drop being closer to the time frame at which
the BL samples registered significantly smaller shrinkage values. This is
indicative of the effect of physical changes i.e., shrinkage on the
reflective characteristics and consequently the hue values which is
known to be impacted by the differential light absorbance at different
wavelengths. The observations are similar to those of Liu et al. (2016) in
carrot and papaya evaluated using multi-spectra and laser imaging
systems who noted that increase in shrinkage caused higher density of
cells due to their closely packed structure leading to lower light reflec­
tion and diffusion.

3.2. Computer vision-based moisture prediction models

As explained in the methodology section 2.5, two types of moisture

prediction models, time-dependent models using drying time (t) and
shrinkage-dependent models using relative area shrinkage (AS) as pre­
dictors were fitted. The selected time-dependent models were used as
benchmarks despite their boundary conditions like product homogene­
ity and negligible shrinkage which are challenges for continually
evolving product-process spaces like food drying.
Among the tested non-linear thin-layer models, Fig. 5a represents the
logarithmic model for an individual replica of CNT samples visually
showing a precise fit which was also the case in BL samples (figure not
shown). This can be attributed to the use of multiple model parameters
(3 or more), which contrarily can also result in insignificant model pa­
Fig. 4. In-line monitoring of colour changes a) luminance (L*) and b) hue angle rameters (principle of parsimony) and data heteroscedasticity due to
(h) with respect to dry basis moisture content (MCdb) along 36 h of drying for logarithmic transformation. In consideration of the stated concerns,
control (CNT) and blanched (BL) carrot slices dried at 35 ◦ C, 35 % R.H. and ~ 3 simple non-linear models like Newton Lewis and Henderson-Pabis have
m s− 1 air velocity. The arrow (in black, from right to left) denotes the time flow also been evaluated. However, these models (figures not shown) have
for both L* and h. Results are reported as mean and standard deviation of the exhibited visibly evident under and/or over-estimation, which signifi­
mean. The vertical green line with the * symbol at the top represents the initial
cantly affected the model performances. Moreover, like the logarithmic
moment at which the CNT and BL values become statistically different (p ≤
model, one and two-parameter thin layer models use time as an input
0.05). (For interpretation of the references to colour in this figure legend, the
reader is referred to the web version of this article.)
model factor, and do not consider spatial changes in the product and its
subsequent effects on the drying rate eventually resulting in less resilient
and robust models (Aghbashlo et al., 2009; Buzrul, 2022; Onwude et al.,
along the drying time (t, shown in Fig. 4a) for both CNT and BL samples.
2016; Raponi et al., 2021). In this context, linear regressions can offer a
The observed initial increase in L* values can be attributed mainly to the
simplistic solution considering the spatial changes of the product to
‘white blush/ bloom’ i.e., surface whitening due to surface dehydration,
provide models that are more robust and not solely time dependent. This
which when intense might be irreversible due to the collapsing of su­
hypothesis was based on our recent study on modelling apple drying for
perficial cells changing their structure (Simões et al., 2010). Some au­
which the AS changed proportionally to the MCdb loss (Raponi et al.,
thors also attributed the white blush to coinciding enzyme activity post-
2021). However, in the case of carrot slices, this linearity might have
cut, particularly those of phenylalanine ammonia-lyase (PAL) which can
deviated considering the observation that a sharp decrease along the
cause phenol accumulation as precursor for lignin biosynthesis (Rocha
drying time was observed at 8 h for AS values, whereas at 11 h for the
et al., 2007a), although other authors evidenced that this effect is un­
drying rate. In other words, it can be speculated that the AS was pro­
related (Simões et al., 2010). Within the tested samples, BL evidenced
portional to the MCdb loss for most of the falling-rate period and regis­
significantly lower L* values than CNT samples from time ‘0′ as indi­
tered deviation after the MCdb reached approx. 1.40 (CNT) to 0.80 (BL) g
cated by the green-dashed area in the figure. This can be attributed to
g− 1. This observation can be visualised in Fig. 5b which shows both the
blanching which could have limited the stated enzyme activity as well as
general linearity between AS and MCdb as well as the breakpoint cor­
improved the surface homogeneity by gas expulsion as observed in other
responding to the threshold value at which the linearity deviated. Sub­
studies on drying (Moscetti et al., 2017; Rocha et al., 2007b). As for the
sequently, the linear segmented model seemed to fit precisely to the

6
S.S.N. Chakravartula et al. Computers and Electronics in Agriculture 206 (2023) 107654

Fig. 5. Plots depicting a replication of a) MR prediction using the logarithmic model wit; b) model fit of the linear segmented model, MR vs AS; c) MR prediction
using the linear segmented model; d) drying rate curve as a function of the MR predicted based using the linear segmented model.

experimental values, with no observable under and/or over-estimation for CNT and BL samples. This can be attributed to the fact that drying
for CNT samples (Fig. 5c) as well as BL samples (figure not shown) for rates were higher in BL due to improved cell permeability by surface
which different regression equations were used. The MR thus predicted microcracks, as previously observed in drying rates (section 3.1)
using the linear-segmented model was further visually projected in the resulting in higher values of rate constant k (t− 1) for non-linear models
drying rate curve in Fig. 5d. This visualisation was tested as it can be of and slope (m) for linear models, respectively. An exception observed was
interest for developing user interfaces that can allow for multiple pro­ the linear segmented model, wherein the model parameters (m1, q1)
jections of the real-time data within the scope of a unified query thereby were statistically non-significant when AS > BP. Whereas, when AS ≤ BP
allowing for optimal monitoring and decision flow by the operator. after an AS value of approx. 0.47, the intercept (q2) values differed
As for the model behaviours, they can be further described by the significantly between the segments of the model indicating two non-
regression parameters presented in Tables 2 and 3. As stated, separate intersecting regression lines for CNT and BL samples. In other words,
regression equations were used for CNT and BL samples as the model an averaged model can be used for MCdb prediction irrespective of pre-
parameters like model constants (k, a, c, m and q) differed significantly treatment until AS > BP; after which i.e., AS ≤ BP the MCdb changes

Table 2
Model parameters and ANOVA for the fitted time-dependent models.
Model Name Treatment Model Parameters

k n a c

Newton-Lewis Control 0.163 b – – – – – –

Blanched 0.186 a – – – – – –
Page Control 0.158 ns 1.015 b – – – –
Blanched 0.186 ns 1.136 a – – – –
Handerson & Pabis Control 0.166 b – – 1.020 b – –
Blanched 0.196 a – – 1.057 a – –
Logarithmic Control 0.179 ns – – 1.018 b 0.020 a
Blanched 0.200 ns – – 1.056 a 0.006 b

*ns, no significant differences (P > 0.05).

7
S.S.N. Chakravartula et al. Computers and Electronics in Agriculture 206 (2023) 107654

Table 3
Model parameters and ANOVA for the fitted time-independent models.
Model name Treatment Model parameters

m1 q1 m2 q2 BP

Linear Control − 0.320 b 1.254 a – – – – – –

Blanched − 0.243 a 1.186 b – – – – – –
Linear segmented Control − 0.432 ns 1.410 ns − 0.257 ns 0.900 b 0.430 ns
Blanched − 0.401 ns 1.393 ns − 0.207 ns 1.041 a 0.510 ns

*ns, no significant differences (P > 0.05).

should be predicted based on individual equations considering the sig­ (segmented-linear) to 2.07 × 10− 02 (simple linear) for CNT samples and
nificant effect of pre-treatments on the AS behaviour. Also, due 1.35 × 10− 16 (segmented-linear) to 1.06 × 10− 01 (simple linear) for BL
consideration must be given to the fact that in the present study, the samples, respectively. The lowest values were noted for logarithmic and
utilised BP was observed beyond 9 h of drying in the carrot slices, which linear-segmented models, whereas other models were highly impacted
can be stated as prolonged drying at MCdb approx. 2.022 (CNT) to 1.22 by systematic error indicative of the consideration that the improvement
(BL) g g− 1 in CNT and BL samples respectively. However, the proposed of model performance will be limited even when replicates are increased
linear segmented model and in general the use of BP can be of practical (Bellon-Maurel et al., 2010). In fact, the significant higher bias of the
significance in cases of monitoring vegetal matrices wherein prolonged simple linear model in both CNT and BL samples indicative of an
drying times and variable pre-treatments that can affect the shrinkage overestimation of the moisture along drying was concurrent with the
and drying rates are utilised. poor fit observed from red. χ2 and RMSE values.
Further, in terms of comparison of prediction performances between Finally, evaluating the presented performance metrics it can be
the time-dependent and shrinkage-dependent models, the selected per­ highlighted that the linear-segmented model showed the best perfor­
formance metrics i.e., RMSE, MAE, BIAS, and red. χ2 are presented in mance metrics that can be benchmarked to the best performing thin
Table 4. In general, all the fitted models registered an adj. R2 of approx. layer model i.e., logarithmic model for moisture prediction.
0.99 and reduced χ2 approaching to zero indicating excellent prediction
capability for the drying behaviour of carrot slices. As for the other
3.3. Remarks on the SMCS implementation
performance metrics i.e., RMSE, MAE, and BIAS, the values for each
model were significantly affected by the model as well as the treatment i.
Thanks to the CV system and the load cells integrated in the dryer as
e., CNT or BL. The RMSE values ranged from 0.007 to 0.174 in CNT
SMCS, the attributes of product i.e., weight loss, colour, and shrinkage as
comparable to those of 0.005 to 0.115 in BL samples. Among the eval­
well as the pre-defined operation variables (Section 2.3 - temperature,
uated models, the linear model registered significantly higher RMSE
relative humidity, and airflow) were successfully recorded in real-time i.
values (CNT, 0.174; BL, 0.115), whereas other models had comparable
e., during the product drying for the 36 h of drying period. The use of
values and the lowest values were observed in logarithmic (CNT, 0.008;
reverse-engineering approach as part of quality strategy allowed to pre-
BL, 0.013) and linear-segmented models (CNT, 0.007; BL, 0.005). In
define and select the said product attributes within carrot drying.
fact, it was of interest to note that the error of the linear-segmented
Consequently, the acquired in-line spatial changes i.e., shrinkage of
model based on relative area shrinkage was superior to the 3-parameter
carrot slices were successful in prediction of the product moisture
logarithmic model, indicating a good fit and performance in moisture
changes through use of simplistic linear regression models as explained
prediction for the carrot slices. As the RMSE is sensitive to outliers, the
in section 3.2. The evaluated shrinkage dependent linear models when
MAE metric was also considered for evaluation. However, MAE showed
benchmarked against conventional thin-layer models, have shown
a similar pattern to that of RMSE with significantly higher error for the
promising prediction accuracy.
linear model and the lowest error noted for logarithmic and linear-
Supplementary video 1 (please see the link below)
segmented models in both CNT and BL samples. Further, the BIAS
https://youtu.be/3UYMkjV6On0 (Supplementary Video 1. Time-
values (systematic error) were found to be in the range of 2.23 × 10− 16
lapse video of 36 h of drying representing the variable shrinkage of the carrot

Table 4
Model performance metrics analysed by ANOVA for the time-dependent and time-independent models fitted to predict MR changes in Control and Blanched carrot
slices.
Model type Model name Sample RMSE MAE BIAS red. χ2

Newton-Lewis Control 0.016 abc 0.014 abc 7.61E− 03 a 2.53E− 04 bc

Time-dependent Blanched 0.018 ab 0.014 abc 2.67E− 03 bc 3.45E− 04 ab
(Thin layer) Page Control 0.015 abc 0.013 abc 8.03E− 03 a 2.35E− 04 bc
Blanched 0.010 cde 0.009 cde 5.58E− 03 ab 1.14E− 04 cde
Henderson-Pabis Control 0.015 abc 0.013 bc 7.56E− 03 a 2.31E− 04 bc
Blanched 0.015 abc 0.011 bc 2.25E− 03 bc 2.17E− 04 bcd

Logarithmic Control 0.008 de 0.006 de 1.97E− 11 c 5.87E− 05 de

Blanched 0.013 bcd 0.009 bcd 1.28E− 09 c 1.82E− 04 bcde

Time-independent Linear Control 0.174 a 0.170 a 2.07E− 02 c 3.09E− 02 a

Blanched 0.115 abc 0.106 bc − 1.06E− 01 c 1.33E− 02 bc
Linear Segmented Control 0.007 de 0.006 de 2.23E− 16 c 5.69E− 05 de
Blanched 0.005 e 0.003 e 1.35E− 16 c 2.86E− 05 e

RMSE – Root mean square error; MAE - Mean average error; BIAS - Systematic error; red. χ2 - reduced chi-square; Mean performance metrics with different letters
indicate significant differences (p < 0.05).

8
S.S.N. Chakravartula et al. Computers and Electronics in Agriculture 206 (2023) 107654

slices on the dryer tray (Panel A), the segmented images (Panel B) and the Giacomo Bedini: Writing – review & editing. Pietro Ibba: Writing –
colour-mapped slices using moisture prediction from results of the linear review & editing, Supervision, Project administration. Riccardo Mas­
segmented model (Panel C). The change in colour from purple to white santini: . Roberto Moscetti: Conceptualization, Methodology, Soft­
gradients in panel C shows the heterogenous change in the moisture content ware, Data curation, Writing – review & editing, Supervision, Project
proportional to the degree of shrinkage within sample set.) administration.
To visualise the model performance, the predicted moisture values
from the best performing linear-segmented model for pre-defined
thresholds of shrinkage were mapped onto the images of individual Declaration of Competing Interest
carrot slices using an interpolated colour palette (purple to white). The
resultant images for 36 h were time-lapsed into a video (suppl. Video 1) The authors declare that they have no known competing financial
which allows the observer to visualise the potential of the smart system interests or personal relationships that could have appeared to influence
for process–product monitoring as well as the inherent heterogeneity of the work reported in this paper.
the process on a hypothetical segment of dryer bed. In the video, the
changes in the original carrot slices on the dryer tray (Panel A), the Data availability
segmented image changes (Panel B) and the colour-mapped slices using
moisture prediction from linear segmented model (Panel C) are repre­ Data will be made available on request.
sented. When played, the video clearly shows the following: (i) the
change in the shape of the carrot slices measured as area shrinkage (AS) Acknowledgements
with respect to the drying time; (ii) the change in colour indicative of the
loss in MCdb with respect to the changes in AS; and (iii) the heteroge­ The authors gratefully acknowledge (1) CORE Organic Plus con­
neity among the samples within the same drying tray owing to the sortium (ERA-NET action) and MiPAAF (Ministero delle politiche agri­
natural variation in the diameter and drying behaviour. cole alimentari e forestali, Italy) for financial support through the
Based on the experimental results it can be said that the use of CV and SusOrgPlus project (D.M. 20/12/2017, n. 92350); (2) the “Departments
load cell system supplemented with validated shrinkage-dependent of Excellence-2018” Program (Dipartimenti di Eccellenza) of the Italian
model can enable continuous monitoring and risk assessment through Ministry of Education, University and Research, DIBAF-Department of
prediction and visualization of a specific process–product space. More­ University of Tuscia, Project “Landscape 4.0 – food, wellbeing and
over, adapting a quality strategy can allow for systematic pre-definition environment”; (3) MISE (Ministero dello Sviluppo Economico, Italy) for
of the specific process parameters and target product characteristics. To the financial support through the READYDRY 4.0 project.
conclude, the tools and approaches adapted in this work form the basis
for a QbD-ready smart dryer that can allow for proactive risk assessment Appendix A. Supplementary data
through predictive monitoring for improved operability and continuous
learning. Supplementary data to this article can be found online at https://doi.
org/10.1016/j.compag.2023.107654.
4. Conclusions
References
The computer vision (CV) system and load cells were successfully
implemented for in-line monitoring of carrot slices during drying within Aghbashlo, M., Kianmehr, M.H., Arabhosseini, A., 2009. Performance analysis of drying
a convective dryer. Further, shrinkage-dependent models were evalu­ of carrot slices in a semi-industrial continuous band dryer. J. Food Eng. 91, 99–108.
https://doi.org/10.1016/j.jfoodeng.2008.08.020.
ated to identify best models that include spatial data for predicting the Aghbashlo, M., Hosseinpour, S., Ghasemi-Varnamkhasti, M., 2014. Computer vision
process flow with scope for proactive risk assessment. The study lays technology for real-time food quality assurance during drying process. Trends Food
foundation for a practical instrument capable of real-time monitoring Sci. Technol. https://doi.org/10.1016/j.tifs.2014.06.003.
Aprajeeta, J., Gopirajah, R., Anandharamakrishnan, C., 2015. Shrinkage and porosity
and control based in QbD strategy. However, further studies are being
effects on heat and mass transfer during potato drying. J. Food Eng. 144, 119–128.
performed to integrate the SMCS, predictive models and process auto­ https://doi.org/10.1016/j.jfoodeng.2014.08.004.
mation by use of AI based algorithms (like neural networks) as well as Bedini, G., Nallan Chakravartula, S.S., Bastianelli, G., Caccia, R., Contarini, M., Morales-
multiple product-process spaces for real-time monitoring and prediction Rodríguez, C., Rossini, L., Speranza, S., Vannini, A., Moscetti, R., Massantini, R.,
2020. Feasibility of ft-nir spectroscopy and vis/nir hyperspectral imaging for sorting
of product changes and process-flow, which is beyond the scope of the unsound chestnuts. Italus Hortus 27, 3–18. https://doi.org/10.26353/j.itahort/
present study. 2020.1.0318.
The salient findings of this study can be outlined as below: Bellon-Maurel, V., Fernandez-Ahumada, E., Palagos, B., Roger, J.-M., McBratney, A.,
2010. Critical review of chemometric indicators commonly used for assessing the
quality of the prediction of soil attributes by NIR spectroscopy. TrAC Trends Anal.
(i) the CV system successfully tracked the shrinkage and colour Chem. 29, 1073–1081. https://doi.org/10.1016/j.trac.2010.05.006.
changes of carrot slices during drying irrespective of pre- Buzrul, S., 2022a. Processes Reassessment of Thin-Layer Drying Models for Foods: A
Critical Short Communication. https://doi.org/10.3390/pr10010118.
treatments; Crippa, M., Solazzo, E., Guizzardi, D., Monforti-Ferrario, F., Tubiello, F.N., Leip, A.,
(ii) the moisture evolution during drying can be effectively predicted 2021. Food systems are responsible for a third of global anthropogenic GHG
by linear-segmented model considering deviations of shrinkage emissions. Nat. Food 2, 198–209. https://doi.org/10.1038/s43016-021-00225-9.
Defraeye, T., 2014. Advanced computational modelling for drying processes - A review.
linearity along the drying period by use of the breakpoint (BP);
Appl. Energy. https://doi.org/10.1016/j.apenergy.2014.06.027.
(iii) the CV system enabled with shrinkage-based modelling and Dhall, A., Datta, A.K., 2011. Transport in deformable food materials: A poromechanics
proactive quality strategy can be an appropriate and feasible approach. Chem. Eng. Sci. 66, 6482–6497. https://doi.org/10.1016/j.
ces.2011.09.001.
approach for a ‘Smart dryer’ capable of real-time monitoring for
Erbay, Z., Icier, F., 2010. A review of thin layer drying of foods: Theory, modeling, and
better process and product control. experimental results. Crit. Rev. Food Sci. Nutr. 50, 441–464. https://doi.org/
10.1080/10408390802437063.
Erenturk, S., Erenturk, K., 2007. Comparison of genetic algorithm and neural network
approaches for the drying process of carrot. J. Food Eng. 78, 905–912. https://doi.
CRediT authorship contribution statement org/10.1016/j.jfoodeng.2005.11.031.
Fellows, P.J., Fellows, P.J., 2017. Blanching, in: Food Processing Technology. pp.
Swathi Sirisha Nallan Chakravartula: Data curation, Writing – 525–538. https://doi.org/10.1016/B978-0-08-100522-4.00009-2.
Ghasemi-Varnamkhasti, M.G., Aghbashlo, M., 2014. Electronic nose and electronic
original draft, Writing – review & editing. Andrea Bandiera: Method­ mucosa as innovative instruments for real-time monitoring of food dryers. Trends
ology, Software, Formal analysis, Data curation. Marco Nardella: . Food Sci. Technol. https://doi.org/10.1016/j.tifs.2014.05.004.

9
S.S.N. Chakravartula et al. Computers and Electronics in Agriculture 206 (2023) 107654

González-Fésler, M., Salvatori, D., Gómez, P., Alzamora, S.M., 2008. Convective air Moses, J.A., Alagusundaram, K., Tiwari, B.K., 2014. Novel Drying Techniques for the
drying of apples as affected by blanching and calcium impregnation. J. Food Eng. 87, Food Industry, pp. 43–55. https://doi.org/10.1007/s12393-014-9078-7.
323–332. https://doi.org/10.1016/j.jfoodeng.2007.12.007. Nguyen, T.K., Mondor, M., Ratti, C., 2018. Shrinkage of cellular food during air drying.
Gulati, T., Datta, A.K., 2015. Mechanistic understanding of case-hardening and texture J. Food Eng. 230, 8–17. https://doi.org/10.1016/j.jfoodeng.2018.02.017.
development during drying of food materials. J. Food Eng. 166, 119–138. https:// Onwude, D.I., Hashim, N., Janius, R.B., Nawi, N.M., Abdan, K., 2016. Modelling the
doi.org/10.1016/J.JFOODENG.2015.05.031. Thin-Layer Drying of Fruits and Vegetables: A Review. Compr. Rev. Food Sci. Food
Khan, M.I.H., Sablani, S.S., Joardder, M.U.H., Karim, M.A., 2020. Application of machine Saf. 15, 599–618. https://doi.org/10.1111/1541-4337.12196.
learning-based approach in food drying: opportunities and challenges. Dry. Technol. Onwude, D.I., Hashim, N., Abdan, K., Janius, R., Chen, G., 2018. The potential of
1–17. https://doi.org/10.1080/07373937.2020.1853152. computer vision, optical backscattering parameters and artificial neural network
Krokida, M.K., Karathanos, V.T., Maroulis, Z.B., Marinos-Kouris, D., 2003. Drying modelling in monitoring the shrinkage of sweet potato (Ipomoea batatas L.) during
kinetics of some vegetables. J. Food Eng. 59, 391–403. https://doi.org/10.1016/ drying. J. Sci. Food Agric. 98, 1310–1324. https://doi.org/10.1002/jsfa.8595.
S0260-8774(02)00498-3. Planinić, M., Velić, D., Tomas, S., Bilić, M., Bucić, A., 2005. Modelling of drying and
Lahsasni, S., Kouhila, M., Mahrouz, M., Jaouhari, J.T., 2004. Drying kinetics of prickly rehydration of carrots using Peleg’s model. Eur. Food Res. Technol. 221, 446–451.
pear fruit (Opuntia ficus indica). J. Food Eng. 61, 173–179. https://doi.org/ https://doi.org/10.1007/s00217-005-1200-x.
10.1016/S0260-8774(03)00084-0. Purlis, E., Cevoli, C., Fabbri, A., 2021. Modelling Volume Change and Deformation in
Lemmens, L., Tibäck, E., Svelander, C., Smout, C., Ahrné, L., Langton, M., Alminger, M., Food Products/Processes: An Overview. Foods 10, 778. https://doi.org/10.3390/
Van Loey, A., Hendrickx, M., 2009. Thermal pretreatments of carrot pieces using FOODS10040778.
different heating techniques: Effect on quality related aspects. Innov. Food Sci. Raponi, F., Moscetti, R., Nallan Chakravartula, S.S., Fidaleo, M., Massantini, R., 2021.
Emerg. Technol. 10, 522–529. https://doi.org/10.1016/j.ifset.2009.05.004. Monitoring the hot-air drying process of organically grown apples (cv. Gala) using
Liu, C., Liu, W., Lu, X., Chen, W., Yang, J., Zheng, L., 2016. Potential of multispectral computer vision. Biosyst. Eng. https://doi.org/10.1016/j.
imaging for real-time determination of colour change and moisture distribution in biosystemseng.2021.07.005.
carrot slices during hot air dehydration. Food Chem. 195, 110–116. https://doi.org/ Raponi, F., Moscetti, R., Monarca, D., Colantoni, A., Massantini, R., 2017. Monitoring
10.1016/j.foodchem.2015.04.145. and optimization of the process of drying fruits and vegetables using computer
Liu, Y., Pu, H., Sun, D.W., 2017. Hyperspectral imaging technique for evaluating food vision: A review. Sustain. 9 https://doi.org/10.3390/su9112009.
quality and safety during various processes: A review of recent applications. Trends Rathore, A.S., Kapoor, G., 2016. Implementation of Quality by Design for processing of
Food Sci. Technol. https://doi.org/10.1016/j.tifs.2017.08.013. food products and biotherapeutics. Food Bioprod. Process. https://doi.org/10.1016/
Lyndgaard, C.B., Engelsen, S.B., Van Den Berg, F.W.J., 2012. Real-time modeling of milk j.fbp.2016.05.009.
coagulation using in-line near infrared spectroscopy. J. Food Eng. 108, 345–352. Rocha, A.M., Ferreira, J.F., Silva, Â.M., Almeida, G.N., Morais, A.M., 2007a. Quality of
https://doi.org/10.1016/J.JFOODENG.2011.07.029. grated carrot (var. Nantes) packed under vacuum. J. Sci. Food Agric. 87, 447–451.
Madiouli, J., Sghaier, J., Lecomte, D., Sammouda, H., 2012. Determination of porosity https://doi.org/10.1002/jsfa.2723.
change from shrinkage curves during drying of food material. Food Bioprod. Process. Rocha, A.M.C.N., Mota, C.C.A.R., Morais, A.M.M.B., 2007b. Physico-chemical qualities of
90, 43–51. https://doi.org/10.1016/j.fbp.2010.12.002. minimally processed carrot stored under vacuum. J. Foodserv. 18, 23–30. https://
Mayor, L., Sereno, A.M., 2004. Modelling shrinkage during convective drying of food doi.org/10.1111/j.1745-4506.2007.00043.x.
materials: A review. J. Food Eng. 61, 373–386. https://doi.org/10.1016/S0260- Seyedabadi, E., Khojastehpour, M., Abbaspour-Fard, M.H., 2019. Online measuring of
8774(03)00144-4. quality changes of banana slabs during convective drying. Eng. Agric. Environ. Food
Menon, A., Stojceska, V., Tassou, S.A., 2020. A systematic review on the recent advances 12, 111–117. https://doi.org/10.1016/j.eaef.2018.10.004.
of the energy efficiency improvements in non-conventional food drying Simões, A. do N., Ventrella, M.C., Moretti, C.L., Carnelossi, M.A.G., Puschmann, R., 2010.
technologies. Trends Food Sci. Technol. https://doi.org/10.1016/j.tifs.2020.03.014. Anatomical and physiological evidence of white blush on baby carrot surfaces.
Moscetti, R., Haff, R.P., Ferri, S., Raponi, F., Monarca, D., Liang, P., Massantini, R., 2017. Postharvest Biol. Technol. 55, 45–52. https://doi.org/10.1016/J.
Real-Time Monitoring of Organic Carrot (var. Romance) During Hot-Air Drying POSTHARVBIO.2009.07.007.
Using Near-Infrared Spectroscopy. Food Bioprocess Technol. 10, 2046–2059. Toǧrul, H., 2006. Suitable drying model for infrared drying of carrot. J. Food Eng. 77,
https://doi.org/10.1007/s11947-017-1975-3. 610–619. https://doi.org/10.1016/j.jfoodeng.2005.07.020.
Moscetti, R., Sturm, B., Crichton, S.O.J.O., Amjad, W., Massantini, R., 2018. Postharvest van den Berg, F., Lyndgaard, C.B., Sørensen, K.M., Engelsen, S.B., 2013. Process
monitoring of organic potato (cv. Anuschka) during hot-air drying using visible–NIR Analytical Technology in the food industry. Trends Food Sci. Technol. https://doi.
hyperspectral imaging. J. Sci. Food Agric. 98, 2507–2517. https://doi.org/10.1002/ org/10.1016/j.tifs.2012.04.007.
jsfa.8737. Wu, B., Pan, Z., Qu, W., Wang, B., Wang, J., Ma, H., 2014. Effect of simultaneous infrared
Moscetti, R., Massantini, R., Fidaleo, M., 2019. Application on-line NIR spectroscopy and dry-blanching and dehydration on quality characteristics of carrot slices. LWT - Food
other process analytical technology tools to the characterization of soy sauce Sci. Technol. 57, 90–98. https://doi.org/10.1016/j.lwt.2013.11.035.
desalting by electrodialysis. J. Food Eng. 263, 243–252. https://doi.org/10.1016/j. Zhang, M., Chen, H., Mujumdar, A.S., Tang, J., Miao, S., Wang, Y., 2017. Recent
jfoodeng.2019.06.022. developments in high-quality drying of vegetables, fruits, and aquatic products. Crit.
Moscetti, R., Berhe, D.H., Agrimi, M., Haff, R.P., Liang, P., Ferri, S., Monarca, D., Rev. Food Sci. Nutr. 57, 1239–1255. https://doi.org/10.1080/
Massantini, R., 2021. Pine nut species recognition using NIR spectroscopy and image 10408398.2014.979280.
analysis. J. Food Eng. 292, 110357 https://doi.org/10.1016/j.
jfoodeng.2020.110357.

10