remote sensing

Article
A Rapid, Low-Cost, and High-Precision Multifrequency
Electrical Impedance Tomography Data Acquisition System for
Plant Phenotyping
Rinku Basak * and Khan A. Wahid

Department of Electrical and Computer Engineering, University of Saskatchewan, Saskatoon, SK S7N 5A9,
Canada; [email protected]
* Correspondence: [email protected]; Tel.: +1-306-203-6124

Abstract: Plant phenotyping plays an important role for the thorough assessment of plant traits such
as growth, development, and physiological processes with the target of achieving higher crop yields
by the proper crop management. The assessment can be done by utilizing two- and three-dimensional
image reconstructions of the inhomogeneities. The quality of the reconstructed image is required to
maintain a high accuracy and a good resolution, and it is desirable to reconstruct the images with
the lowest possible noise. In this work, an electrical impedance tomography (EIT) data acquisition
system is developed for the reconstruction and evaluation of the inhomogeneities by utilizing a non-
destructive method. A high-precision EIT system is developed by designing an electrode array sensor
using a cylindrical domain for the measurements in different planes. Different edible plant slices along
with multiple plant roots are taken in the EIT domain to assess and calibrate the system, and their
reconstructed results are evaluated by utilizing an impedance imaging technique. A non-invasive
imaging is carried out in multiple frequencies by utilizing a difference method of reconstruction. The
performance and accuracy of the EIT system are evaluated by measuring impedances between 1 and
Citation: Basak, R.; Wahid, K.A. A 100 kHz using a low-cost and rapid electrical impedance spectroscopy (EIS) tool connected to the
Rapid, Low-Cost, and High-Precision sensor. A finite element method (FEM) modeling is utilized for image reconstruction, which is carried
Multifrequency Electrical Impedance out using electrical impedance and diffuse optical tomography reconstruction software (EIDORS).
Tomography Data Acquisition The reconstruction is made successfully with the optimized results obtained using Gauss–Newton
System for Plant Phenotyping. (GN) algorithms.
Remote Sens. 2022, 14, 3214.
https://doi.org/10.3390/ Keywords: electrical impedance tomography; multifrequency; image reconstruction; plant phenotyping
rs14133214

Academic Editor: Xia Yao

Received: 1 April 2022

1. Introduction
Accepted: 30 June 2022
Published: 4 July 2022
EIT is an imaging method that can estimate the electrical impedance distribution inside
an object. The EIT sensor system consists of an electrode array with a single or multiple
Publisher’s Note: MDPI stays neutral
layers, and the current density is distributed inside the medium based on the stimulated
with regard to jurisdictional claims in
current through the electrodes [1–3]. A small current is injected into the object’s surface
published maps and institutional affil-
through a pair of electrodes and then the impedance distribution and the corresponding
iations.
conductivity are estimated. Electrical resistivity or conductivity images of any closed
domain under test can be reconstructed using the boundary potentials [1,2]. The data
obtained by measuring all possible impedances are used to reconstruct an image, which
Copyright: © 2022 by the authors.
may provide qualitative and quantitative information.
Licensee MDPI, Basel, Switzerland. EIT could provide a cost-effective alternative to the established imaging methods
This article is an open access article across a wide range of applications. EIT has a high temporal resolution, but it has a poor
distributed under the terms and spatial resolution, so this technique is sensitive to noise. Although the spatial resolution
conditions of the Creative Commons of EIT is less than magnetic resonance imaging (MRI), EIT is easy to use and its cost is
Attribution (CC BY) license (https:// much less than the cost of MRI, which can make EIT a useful screening tool for imaging.
creativecommons.org/licenses/by/ The spatial resolution can be improved by increasing the number of electrodes in the EIT
4.0/). system and by choosing the correct drive pattern upon the selection of electrode pairs for

Remote Sens. 2022, 14, 3214. https://doi.org/10.3390/rs14133214 https://www.mdpi.com/journal/remotesensing

Remote Sens. 2022, 14, 3214 2 of 28

current and voltage stimulation [4,5]. Due to its unique advantages, EIT has enormous
applications in biomedical imaging, plant physiology, biotechnology, nanotechnology, and
material engineering. EIT is becoming popular in research, offering exceptionally good
benefits of non-invasive, non-ionizing, fast imaging speed, and low-cost monitoring. From
the images of EIT, the physiological information of the matter can be obtained that can be
used for real-time applications [5,6].
Multifrequency EIT (MFEIT) systems give more useful information about biological
matters because the electrical voltage appearing across the matter is frequency depen-
dent [7–10]. Significant information about the matters can be obtained by injecting currents
in multiple frequencies and measuring the impedances at different electrode positions. A
good conductivity distribution of an object can be mapped in the domain under test by
measuring the electrical impedances through the multiple electrodes at various frequen-
cies. The image of an object can be reconstructed by calculating the boundary voltages
in homogeneous and inhomogeneous conditions of the domain under test and with the
help of finite element method (FEM) modeling [2,8]. In EIT, the FEM technique is used to
derive the forward model from the governing equation, which is described by Laplace’s
equation [1,2,5]. In forward solve, the boundary potentials are calculated by the injected
current and known conductivity distribution inside the EIT domain. On the other hand,
the unknown conductivity changes are calculated in the inverse solve by knowing the
differences in boundary potentials of two different media for the given stimulation current
in the domain.
A multifrequency impedance imaging technique can be utilized considering a multi-
electrode array with eight or more electrodes in a domain for obtaining the more useful
information of an inhomogeneity. The EIS method was applied in the electrode array
system, and the reconstructed images were presented in multiple frequencies by the phan-
tom experiments [11]. Therefore, EIT is a non-invasive, and non-destructive impedance
imaging technique. It is a radiation-free, rapid, and cost-effective alternative to other
laboratory-based imaging methods such as MRI, computed tomography (CT), and positron
emission tomography (PET). The estimation performance can be potentially improved
by considering a multispectral impedance imaging technique using an EIT system [7,9].
Due to its unique advantages, EIT has enormous applications in cell imaging [6], brain
imaging [12], anomaly detection [13,14], and crop root imaging [15–18], respectively.
Previously, EIT image reconstruction was made considering different algorithms in
multiple works [8,19–22]. An FPGA-based data acquisition system at 1–190 kHz was devel-
oped considering 16 electrodes for two-dimensional (2D) brain imaging by Shi et al. [12].
Two-dimensional imaging, characterization, and monitoring of oilseed plant root systems
were made at frequencies of 0.46 Hz–45 kHz considering an array of 38 electrodes by
Weigand and Kemna [15,16]. A rapid estimation of wheat plant root biomass by measuring
capacitance up to 20 kHz using a handheld LCR meter was proposed by Postic et al. [17].
The development of oilseed plant root was visualized using a data acquisition system in
three-dimensions (3D) at frequencies of 5–10 kHz considering an array of 32 electrodes
by Corona-Lopez et al. [18]. The methods of all these works required expensive instru-
mentations and most of them were found as laboratory-based and not suitable for in situ
measurements. In addition, several data acquisition systems based on EIT were developed
for industrial and medical applications and those were utilized for different previous re-
search works on clinical imaging [10], health monitoring [23], and diagnosis of human
body diseases with anomaly detection [24]. The outputs of these EIT systems were limited
to 2D imaging. Hence, a portable and high-speed multifrequency 3D EIT data acquisition
system for in situ applications in plant phenotyping is still a constant requirement.
The characterization in phenotyping refers to qualitative and quantitative descriptions
of the plant’s biological characteristics, such as morphological, physiological, and biochem-
ical. Identifying the size, shape, and structure of the roots or other bodies of the plant,
growth, and development with the event of photosynthesis, and quantifying the chemical
properties such as water and nutrients are very important in biological study of the plant
Remote Sens. 2022, 14, 3214 3 of 28

characteristics. The root system of a plant plays an important role in photosynthesis by

which the water and nutrients are transferred to the plant stems and leaves by absorbing
from soil. The flowering plants are classified as vascular (reproduced by seeds/vegetative
parts), which contain a specialized xylem and phloem tissues for the transportation of water
and nutrients, while the non-flowering plants are classified as non-vascular (reproduced by
spores), which do not contain specialized vascular tissues for transport.
An investigation on growth, development, and biomass of the root is very important
for characterization in plant phenotyping. The investigation can be made by varying the
electrical parameters such as, capacitance, resistance, or impedance. In several experiments,
the dependency of electrical impedance on plant characteristics was found and evaluated
at different frequencies [11,15,16,18]. An imaging method was utilized to characterize the
root by measuring the electrical parameters. Previously, based on the electrical param-
eters, the root biomass was estimated by Postic et al. [17], analysis on root growth was
made by Ozier-Lafontaine et al. [25], and the root body of a plant in water was recovered
by Liao et al. [26]. The imaging of water distribution in the root zone was made in a
laboratory-scale rhizotron container in a soil media and recovered 2D information only
by Newill et al. [27]. More information on root structure, density, and distributions can
be obtained by utilizing a 3D imaging system, which is advantageous over 2D. Spectral
variations on root characteristics can be made using a 3D imaging system. The growth and
development of the root can be monitored precisely by knowing the electrical characteristics
in multiple frequencies obtained using a 3D EIT system. Presently, Weigand and Kemna
utilized a multifrequency EIT system in a laboratory for root characterization and imaging
in a water-filled rhizotron considering polarization effects and recovered 2D information
only [15,16], Corona-Lopez et al. visualized the developing root system in a compost-filled
container using multifrequency EIT and recovered some 3D information of the root at a low
frequency of 5 kHz; no information was found about the high spectral performance [18].
In addition, the changes on root characteristics during growth can be evaluated
by employing the measurements considering frequency-difference and time-difference
EIT [7,8,14,28]. Image reconstruction using difference data at two injecting frequencies
at a particular time, or the data obtained for single frequency measurements over a time
duration can be made successfully using a 3D imaging system. A new cost-effective and
high spectral range 3D EIT system along with the existing measurement methods is still a
constant requirement in the field of root study. Non-invasive imaging using 3D, monitoring
growth, and estimating biomass at the laboratory and field scale are still demanding. EIT
using a portable multiple electrode array with capability of 3D imaging seems to be a
promising method to fulfill the scope of further research on plant phenotyping.
In this work, a rapid, low-cost, and high-precision EIT data acquisition system is
developed with the target of applications in plant phenotyping for the non-destructive eval-
uation of the inhomogeneities considering a non-invasive impedance imaging technique in
multiple frequencies.

2. Materials and Methods

2.1. Design of EIT Electrode Array Sensor System
A cylindrical plastic domain of 4.5-inch diameter, 3-inch height, and 0.0625-inch
thickness was taken. An EIT electrode array of sixteen metal electrodes was designed
and the electrodes were configured in two layers (eight electrodes each) of the domain
as shown in Figure 1b. The electrodes were with one-half inch diameter, one inch per
arm length (L shaped electrodes with two arms), and 0.0625-inch thickness. The electrode
arrangement in the layers is aligned. The bottom layer consisted of eight electrodes that
were placed equally 1.25 inches apart and at a 1.25-inch top from the bottom position by
keeping 0.25-inch separation of the electrodes from the base of the domain as shown in
Figure 1a. Another eight electrodes of same size are placed similarly in another layer at a
0.75-inch top position from the bottom layer as shown in Figure 1b.
 distribution was found with the optimized dimensions of the electrodes and the electrode
array system was found suitable for 3D EIT imaging by the measurements considering a
planar-aligned electrode placement configuration [30]. The electrodes were placed
vertically in the same line of the domain and found an optimal design as shown in Figure
1b. A low-cost EIT sensor was able to be designed and the measurement method with this
Remote Sens. 2022, 14, 3214 4 of 28
arrangement was found robust in obtaining tomography of a sample with less noise and
less error.

(a) (b)

Figure
Figure 1.
1. An electrode array
array sensor
sensorwith
with(a)
(a)8 8electrodes
electrodes (bottom
(bottom layer)
layer) forfor
2D2D
EIT,EIT,
andand (b)elec-
(b) 16 16
electrodes
trodes (top(top
andand bottom
bottom layers)
layers) designed
designed in aincylindrical
a cylindrical domain
domain for for
3D 3D
EITEIT measurements.
measurements.

The size of the

2.2. Development domain,
of EIT number, and
Data Acquisition location of electrodes in a layer, and type and di-
System
mensions
An automated EIT data acquisition by
of the electrodes were optimized conducting
system several experiments
is developed considering
for the measurements
previous research works [10,29,30]. Steel electrodes were chosen, which were
considering 2D and 3D operations, as shown in Figure 2. An EIS tool EVAL-AD5933EBZ found com-
paratively less expensive than the other electrodes made by highly conductive
was interfaced with Arduino Uno (ATmega328P), EIS data storage (PuTTY), and electrode materials
such as silver, copper, or aluminium. The chosen electrodes were silver-colored electro-
plates made of zinc-plated steel material. The electrodes were found suitable with good
current carrying capability and with a good correlation between the measured impedances
and dimensions (length and diameter) of the electrodes. The electrodes were numbered
sequentially in the domain of EIT sensor system. After designing, the sensor was tested
and characterized by the injected current. The measurements were carried out considering
8 metal electrodes for 2D and 16 metal electrodes for 3D. A good conductivity distribution
was found with the optimized dimensions of the electrodes and the electrode array system
was found suitable for 3D EIT imaging by the measurements considering a planar-aligned
electrode placement configuration [30]. The electrodes were placed vertically in the same
line of the domain and found an optimal design as shown in Figure 1b. A low-cost EIT
sensor was able to be designed and the measurement method with this arrangement was
found robust in obtaining tomography of a sample with less noise and less error.

2.2. Development of EIT Data Acquisition System

An automated EIT data acquisition system is developed for the measurements con-
sidering 2D and 3D operations, as shown in Figure 2. An EIS tool EVAL-AD5933EBZ
was interfaced with Arduino Uno (ATmega328P), EIS data storage (PuTTY), and electrode
switching multiplexers (CD74HC4067). AD5933 is a high-precision impedance converter
system that combines an on-board frequency generator; it has programmable graphic user
interface with frequency sweep capability and serial I2 C interface [31–33]. The EIT sensor
is connected to the EIS tool through multiplexers. Two multiplexers were connected with
EIT electrode array system for 2D or 3D operations by injecting currents and capturing
boundary voltages. The impedance of an object in the domain was measured by the current
injected through the electrodes with an automated and appropriate EIS measurement
settings using Arduino Uno programming. The measured impedances for all the electrodes
in a 2D/3D system were stored by varying frequencies and output excitations controlled by
 interface with frequency sweep capability and serial I2C interface [31–33]. The EIT sensor
is connected to the EIS tool through multiplexers. Two multiplexers were connected with
EIT electrode array system for 2D or 3D operations by injecting currents and capturing
boundary voltages. The impedance of an object in the domain was measured by the
current injected through the electrodes with an automated and appropriate EIS
Remote Sens. 2022, 14, 3214 5 of 28
measurement settings using Arduino Uno programming. The measured impedances for
all the electrodes in a 2D/3D system were stored by varying frequencies and output
excitations controlled by the EIS tool (AD5933). In addition, a 15 MHz DDS signal
the EIS tool
generator (AD5933).
(JDS6600) wasIn used
addition, a 15 MHz
to obtain DDS signal
the current generator
carrying (JDS6600)
capability of thewas used to
electrodes.
obtain the current carrying capability of the electrodes. The current is injected in
The current is injected in different frequencies through the driving electrodes and the AC different
frequencies
sensing through
voltages the measured
were driving electrodes and the
for different AC sensing
current levels voltages were measured
by the sensing for
electrodes
different current levels by the sensing electrodes using a multimeter. The data
using a multimeter. The data were stored using the acquisition systems and analyzed for were stored
usingimage
EIT the acquisition systems in
reconstruction andMATLAB
analyzed for EIT image
using reconstruction
an open-source in MATLAB
software, using
EIDORS
an open-source
[2,10,11,23]. software, EIDORS [2,10,11,23].

Figure
Figure 2.
2. A
A developed
developed EIT
EIT data
data acquisition
acquisition system
system for
for 2D
2D and
and 3D
3D imaging.
imaging.

2.3. Calculating
2.3. Calculating Conductivity
Conductivity
The physical
The physicalrelationship
relationshipbetween
betweenthe the conductivity
conductivity andand
thethe boundary
boundary voltages
voltages in
in the
the domain
domain is governed
is governed by by a partial
a partial differentialequation
differential equationthat thatisisderived
derived from
from Maxwell’s
Maxwell’s
equations. To
equations. Toobtain
obtainthe theconductivity,
conductivity, thethe EIT
EIT inverse
inverse solution
solution is required
is required and,and, hence,
hence, the
the forward problem is solved at first. The boundary voltage is
forward problem is solved at first. The boundary voltage is calculated for the given calculated for the given
conductivity distribution
conductivity distributionand andthe injected
the current.
injected current. In this work,
In this a complete
work, electrode
a complete model
electrode
is used and EIT problems are solved numerically by utilizing
model is used and EIT problems are solved numerically by utilizing FEM modeling.FEM modeling.
In the
In the EIT
EIT experiment,
experiment, aa smallsmall mAmA current
current isis injected
injected through
through an an electrode
electrode array
array with
with
L = 8 or 16 electrodes at frequencies 1–100 kHz, and the boundary voltages
L = 8 or 16 electrodes at frequencies 1–100 kHz, and the boundary voltages are measured. are measured.
When current IIll is
When current is injected
injected through electrode eell on
through electrode on the
the surface
surface δΩ. and the
𝛿𝛺 and the conductivity
conductivity
distribution σ
distribution 𝜎 isisknown,
known,the the electric potential V
electric potential in the
V in domain Ω
the domain 𝛺 can
can be
be solved from the
solved from the
governing equation with the boundary conditions for the complete electrode model.
governing equation with the boundary conditions for the complete electrode model.
According to KCL, in the absence of independent electric charges, the summation of the
According to KCL, in the absence of independent electric charges, the summation of
outward and inward current at any point of a closed surface inside the domain Ω is zero,
the outward and inward current at any point of a closed surface inside the domain 𝛺 is
zero, →
∇. J = 0 (1)
∇. 𝐽⃗ = 0 (1)
→
According to
According
→
to Ohm’s
Ohm’s law,
law, the
the relation
relation between
between the the current density((𝐽J⃗)) and
currentdensity and the
the electric
electric
field (𝐸⃗ ) is
( E) is
→ →
J ⃗= σ E⃗ (2)
𝐽 = 𝜎𝐸 (2)
With quasi-static
With quasi-static assumption,
assumption, the
the electric
electric field
field can
can be written in
be written the form
in the form of
of aa gradient
gradient
of a scalar potential (V) as
of a scalar potential (V) as →
E = −∇V (3)
In the continuum electrode model, there are no electrodes facing the boundary of an
object. The model assumes that the current density is a continuous function on the entire
boundary of the object. Comparing the Equations (1)–(3), the sensing field (EIT governing
equation) can be described by Laplace’s equation (derived from Maxwell’s equation) [1,2,5]
as follows:
∇.(σ∇V ) = 0 in domain Ω (4)
The gap model assumes the discrete electrodes on the surface of an object, and the
shunting effect of the electrodes and contact impedance is ignored. Consider, the current is
Remote Sens. 2022, 14, 3214 6 of 28

injected in the boundary through discrete electrodes, thus, the current density Jl is injected
through the l-th electrode is given in [1,5] by
→ ∂V
Jl = − J .n̂ = σ ∇V.n̂ = σ (5)
∂n̂
→
This must be satisfied in the δΩ boundary, where n̂ = |nn| is considered as the outward
unit normal vector.
The shunt model refines the gap model by taking into account the shunting effect of
the electrode. The current is injected in the boundary through the electrodes and calculated
using Neumann boundary condition [1,5] as
Z
∂V
Il = σ ds on el , l = 1, 2, 3, . . . , L (6)
el ∂n̂

The complete electrode model is a refinement of the shunt electrode model in which
the shunting effect and contact impedance between the electrodes and the object in the
medium is taken into account. The measured voltage on electrode el is Vl , which is given
by the Dirichlet boundary condition [1,5] as

∂V
Vl = V + Zl σ on el , l = 1, 2, 3, . . . , L (7)
∂n̂
where Zl is the effective contact impedance between the electrode and the object in
the medium.
The injected current and the measured voltages are also satisfied as ∑lL=1 Il = 0,
and ∑lL=1 Vl = 0. The optimum conductivity distribution is calculated by the voltage
difference, ∆V = Vi − Vh , where Vh and Vi are the measured voltages in homogeneous and
inhomogeneous media. The conductivity map of the inhomogeneity in a single step can be
calculated using [1,5,12] as follows:
  −1
∆σ = J T W J + λ2 R J T W∆V (8)

where J is the Jacobian, which is a determinant for the measurement of voltage sensitivity,
W is the inverse of the covariance of measurements, R is an estimation of the inverse of the
noise covariance, and λ is the hyperparameter that controls the trade-off between resolution
and noise attenuation in the reconstructed image.

2.4. FEM Modeling

EIT problems were solved numerically with the help of finite element method (FEM) mod-
eling. An EIT ‘distmesh’ model of ‘d2d4c’ with 8 electrodes using the ‘mk_common_model’
was optimized for the designed electrode array system in the 2D domain as shown in
Figure 3a. The model consisted of 2507 numbers of nodes, 4757 numbers of elements,
and 255 numbers of boundaries. On the other hand, a ‘Netgen’ cylindrical model using
‘ng_mk_cyl_models’ was optimized for modeling in the 3D domain with 16 electrodes
distributed in two layers (8 electrodes in each layer), as shown in Figure 3b. The model
consisted of 3247 numbers of nodes, 13,963 numbers of elements, and 3264 numbers of
boundaries, respectively. The effects of variations of nodes, elements, and boundary num-
bers on the reconstruction results were examined thoroughly. An optimization was made
by employing different boundary conditions and a satisfactory result was obtained using
the selected models.
The FEM modeling was carried out using EIDORS, whose operation was made with
the help of MATLAB for the reconstruction using the EIT measurements. EIS data were
stored using an open-source software PuTTY (interfaced with Arduino Uno COM3 port)
and a statistical analysis was performed. The flowchart of the EIDORS operation is pre-
 Remote Sens. 2022, 14, x FOR PEER REVIEW 8 of 28

Remote Sens. 2022, 14, 3214 1 7 of 28

𝛿𝜎 = 𝑉 (𝑝 )(2𝐽(𝑝 ) − 1) (10)
𝐿
where L is the number of electrodes, Vm is the measured voltage, p is the position of
electrodes,
sented and 4.
in Figure J isThe
thesensing
currentmethods
density function.
were applied, and the conductivity of the inhomo-
Theobject
geneous back-projection
was mappedalgorithm is knownsteps:
with the following to blur image contrasts.
(i) 2D/3D FEM model Reconstruction
selection, (ii)
stimulation,
using a total(iii)
Remote Sens. 2022, 14, x FOR PEER REVIEW forward(TV)
variation solve: homogeneous
functional and
helps to inhomogeneous
preserve data loading,
discontinuities and
7 of (iv)
in reconstructed28
inverse solve:
profiles, which reconstruction
are smoothened of the inhomogeneity
by traditional by the conductivity
reconstruction algorithmscalculations using
such as Newton’s
one-step
algorithm.Gauss–Newton
Edge-preserving (GN) algorithms
algorithms such
such as
asprior
thoseNOSER, Gaussian
using TV HPF,complex
are more Laplacian, to
and Tikhonov
implement in regularization.
addition to the The
high optimized
computationalcontrolling
cost. parameters
The TV of a such as
conductivitythe size ofis
image
elements, and boundary numbers on the reconstruction results were examined
the domain,
defined the inhomogeneity position and size, the stimulation current/voltage (I/V),
as An
thoroughly. optimization was made by employing different boundary conditions and
the frequency (f ), and the hyperparameter value (λ) were set for the measurements and
a satisfactory result was obtained using
stimulation, and the overall reconstruction = |∇𝜃|models.
the selected
𝑇𝑉(𝜃) performancedΩ was made accordingly. (11)
where Ω is the region to be imaged. Since the conductivity is constant over each FEM
element, ∇𝜃 is non-zero only on the edges between elements. For the i-th edge, shared by
the FEM elements m(i) and n(i), the jump in conductivity is: (𝜃 ( ) − 𝜃 ( ) ) . The variation
over the complete image can be found by integrating the jump over all the edges of the
mesh as

𝑇𝑉(𝜃) = 𝑙 (𝜃 () −𝜃 ( )) (12)

where li is the length of i-th edge in the mesh, and the index i covers all the edges.
Several machine learning algorithms such as an artificial neural network (ANN), a
least angle regression (LARS), and an elastic net were also studied previously for EIT
inverse solution [35–37]. A classic deterministic Gauss–Newton with Laplacian
regularization method were compared with the machine learning algorithms [35]. A good
reconstruction was made using the modified ANN compared to the modified LARS and
elastic net. ANN suffers from long training and reconstruction times; LARS and elastic
(a) (b)
net seem to be less accurate for real time data but much faster than ANN [35]. Most of the
Figure 3.3.FEM
algorithms
Figure FEMModels
have for
some
Models for(a) d2d4c:
d2d4c:2D
2D(circular)
(a)limitations with
with88electrodes,
in obtaining
(circular) 3D images,
electrodes,and
and(b)
(b)Netgen:
among 3D
3D(cylindrical)
these
Netgen: the Gause–
(cylindrical)
with 16
Newton electrodes
method EIT.
was found faster in training and computation and suitable for 3D image
with 16 electrodes EIT.
reconstruction with high accuracy.
The FEM modeling was carried out using EIDORS, whose operation was made with
the help of MATLAB for the reconstruction using the EIT measurements. EIS data were
stored using an open-source software PuTTY (interfaced with Arduino Uno COM3 port)
and a statistical analysis was performed. The flowchart of the EIDORS operation is
presented in Figure 4. The sensing methods were applied, and the conductivity of the
inhomogeneous object was mapped with the following steps: (i) 2D/3D FEM model
selection, (ii) stimulation, (iii) forward solve: homogeneous and inhomogeneous data
loading, and (iv) inverse solve: reconstruction of the inhomogeneity by the conductivity
calculations using one-step Gauss–Newton (GN) algorithms such as prior NOSER,
Gaussian HPF, Laplacian, and Tikhonov regularization. The optimized controlling
parameters such as the size of the domain, the inhomogeneity position and size, the
stimulation current/voltage (I/V), the frequency (f), and the hyperparameter value (λ)
were set for the measurements and stimulation, and the overall reconstruction
performance was made accordingly.
The common EIT inverse algorithms for FEM modeling are Gauss–Newton (GN),
Shefield back-projection (BP), and total variation (TV), respectively [10,19,34]. The GN
method is an iterative algorithm to solve nonlinear least squares problems, BP is a linear
reconstruction algorithm, and TV is a regularization-based algorithm. Electrical
impedance imaging is a highly nonlinear and ill posed inverse problem in which a
minimization algorithm is used to obtain its approximate solution [10,19].
The objective function can be minimized by taking the difference between the
experimental
Figure measured
Figure4.4.Flowchart
Flowchart data and
ofofEIDORS
EIDORS the predicted data. If Vm is the measured voltage matrix
operation.
operation.
and Vc is the calculated voltage matrix, then the Gauss–Newton (GN) algorithm gives a
The common
least square solutionEIT inverse
of the algorithms
minimized objectfor FEM modeling
function are Gauss–Newton
s(σ) [2], which is defined as (GN),
Shefield back-projection (BP), 1and total variation 1 (TV), respectively [10,19,34]. The GN
method is an iterative algorithm𝑠(𝜎) = ‖(𝑉 −𝑉‖
to solve = (𝑉 least
nonlinear − 𝑉 )squares
(𝑉 − 𝑉problems,
) (9)
BP is a linear
2 2
Back-projection (BP) is capable of producing images of changes in conductivity. The
images produced by BP have some clear artifacts because of the inherent character of the
asteroid trace and the conductivity is given by
Remote Sens. 2022, 14, 3214 8 of 28

reconstruction algorithm, and TV is a regularization-based algorithm. Electrical impedance

imaging is a highly nonlinear and ill posed inverse problem in which a minimization
algorithm is used to obtain its approximate solution [10,19].
The objective function can be minimized by taking the difference between the experi-
mental measured data and the predicted data. If Vm is the measured voltage matrix and
Vc is the calculated voltage matrix, then the Gauss–Newton (GN) algorithm gives a least
square solution of the minimized object function s(σ) [2], which is defined as

1 1
s(σ) = k (Vm − Vc ) k 2 = (Vm − Vc )T (Vm − Vc ) (9)
2 2
Back-projection (BP) is capable of producing images of changes in conductivity. The
images produced by BP have some clear artifacts because of the inherent character of the
asteroid trace and the conductivity is given by

1 L
δσ =
L ∑i=1 Vm ( pi )(2J ( pi ) − 1) (10)

where L is the number of electrodes, Vm is the measured voltage, p is the position of

electrodes, and J is the current density function.
The back-projection algorithm is known to blur image contrasts. Reconstruction using
a total variation (TV) functional helps to preserve discontinuities in reconstructed profiles,
which are smoothened by traditional reconstruction algorithms such as Newton’s algorithm.
Edge-preserving algorithms such as those using TV are more complex to implement in
addition to the high computational cost. The TV of a conductivity image is defined as
Z
TV (θ ) = |∇θ | dΩ (11)

where Ω is the region to be imaged. Since the conductivity is constant over each FEM
element, ∇θ is non-zero only on the edges between elements.
 For the i-thedge, shared by
the FEM elements m(i) and n(i), the jump in conductivity is: θm(i) − θn(i) . The variation
over the complete image can be found by integrating the jump over all the edges of the
mesh as
TV (θ ) = ∑ li (θm(i) − θn(i) ) (12)
i

where li is the length of i-th edge in the mesh, and the index i covers all the edges.
Several machine learning algorithms such as an artificial neural network (ANN), a least
angle regression (LARS), and an elastic net were also studied previously for EIT inverse
solution [35–37]. A classic deterministic Gauss–Newton with Laplacian regularization
method were compared with the machine learning algorithms [35]. A good reconstruction
was made using the modified ANN compared to the modified LARS and elastic net. ANN
suffers from long training and reconstruction times; LARS and elastic net seem to be less
accurate for real time data but much faster than ANN [35]. Most of the algorithms have
some limitations in obtaining 3D images, among these the Gause–Newton method was
found faster in training and computation and suitable for 3D image reconstruction with
high accuracy.

2.5. Experimental Setup and Sensor Characterization

An EIT experimental setup for the data acquisition is presented as shown in Figure 5.
The sensing methods applied to the EIT sensor were optimized and utilized based on
the requirements considering homogeneous and inhomogeneous conditions. The results
were validated, and the methods were found suitable in obtaining good reconstruction
performance. An automated and high-speed data acquisition for multiple samples such
as edible plant slices, and plant roots in the domain was possible in a short duration by
utilizing the AD5933 EIS tool. Finally, a difference method was applied for the obtained
 An EIT experimental setup for the data acquisition is presented as shown in Figur
5. The sensing methods applied to the EIT sensor were optimized and utilized based on
the requirements considering homogeneous and inhomogeneous conditions. The result
were validated, and the methods were found suitable in obtaining good reconstruction
Remote Sens. 2022, 14, 3214 performance. An automated and high-speed data acquisition for multiple samples 9 of 28 such
as edible plant slices, and plant roots in the domain was possible in a short duration by
utilizing the AD5933 EIS tool. Finally, a difference method was applied for the obtained
datasetsand
datasets andthethe reconstructed
reconstructed results
results of the of the inhomogeneities
inhomogeneities wereby
were obtained obtained by varying
varying the
stimulation current
the stimulation in 2D and
current 3Dand
in 2D planes.
3D planes.

Figure5.5.An
Figure AnEITEIT experimental
experimental setup
setup for image
for image reconstruction
reconstruction of the inhomogeneities.
of the inhomogeneities. The data The
weredata wer
storedbyby
stored thethe measurements
measurements usingusing EISand
EIS tool tool
DDSand DDS
signal signal generator.
generator. For
For example, example,
a carrot a carrot slic
slice was
was placed
placed at the of
at the centre centre of the and
the domain domain andreconstructed
that was that was reconstructed
using EIDORS.using EIDORS.

Initially, thethe
Initially, EISEIS
sensor using
sensor AD5933
using AD5933was characterized, and theand
was characterized, functionality of the of th
the functionality
tool was tested by measuring impedances of a carrot slice of height 1/5-inch and
tool was tested by measuring impedances of a carrot slice of height 1/5-inch and diameter diameter,
D = 25.4 mm to utilize in the EIT data acquisition system. The EIS measurements of the
D = 25.4 mm to utilize in the EIT data acquisition system. The EIS measurements of th
sample were carried out using two-pole method by varying the output excitation (Vout ) and
samplebetween
spacing were carried out using
two electrodes. two-pole
A pair method by varying
of electrocardiograms the output
(ECG) electrodes excitation (Vout
connected
and
to spacing
AD5933 between by
was separated two electrodes.
d distance and theA reactance,
pair of electrocardiograms (ECG) as
Xc of a sample was calculated electrode
connected to AD5933 was separated by d distance and the reactance,
Xc = 1/2π f C, and sample capacitance, C = εA/d, where A is the cross-sectional area and X c of a sample wa
εcalculated
is the medium as 𝑋 = 1/2𝜋𝑓𝐶
constant. The ,electrical
and sample capacitance,
impedance 𝐶 = 𝜀𝐴/𝑑
(Z)√of a sample , where
measured by A theisEIS
the cross
tool (AD5933EBZ) is related of re 2 + im2 and gain factor [32,33]
sectional area and ε is thetomedium
the DFT magnitude
constant. The electrical impedance (Z) of a sampl
as follows: by the EIS tool (AD5933EBZ) is related to the DFT magnitude of √𝑟𝑒 + 𝑖𝑚
measured 1
and gain factor [32,33] Impedance, Z (Ohm) =
as follows: √ (13)
Gain Factor × re2 + im2
where the gain factor is calibrated by a known resistance of 7.5 kΩ. 1 The gain factor varies
𝐼𝑚𝑝𝑒𝑑𝑎𝑛𝑐𝑒, 𝑍(𝑂ℎ𝑚) = (13
with the variation of output excitation and physical𝐺𝑎𝑖𝑛 𝐹𝑎𝑐𝑡𝑜𝑟
frequency √𝑟𝑒
(f ) for + 𝑖𝑚
a given sample. Here,
re and im are the DFT real and imaginary outputs registered at different frequency codes
where
generated thephysical
by the gain factor is calibrated
frequency, byFrequency
f = f clk × a knownCode/2
resistance of 7.5
29 , where kΩ.
fclk is theThe gain facto
master
varies
clock with the
frequency variation
of 16.776 MHzofforoutput excitation
the internal and[32].
oscillation physical frequency (f) for a given
sample. Here, re and im (X)are the DFT real and imaginary outputs registered at differen
The sample reactive and resistive (R) components  as X = Z sin θ.
are calculated
frequency
and R = Z cos codes generated
θ, where phase ofbythe
theelectrical
physical frequency,
impedance, θ =𝑓 tan
=𝑓 −1 X 𝐹𝑟𝑒𝑞𝑢𝑒𝑛𝑐𝑦
R , and Z is the
𝐶𝑜𝑑𝑒/2
where f
magnitude of the impedance. The influence of impedance magnitude in determining the [32].
clk is the master clock frequency of 16.776 MHz for the internal oscillation
sampleThe sample reactive
characteristics (X) and
was found muchresistive
higher (R)
thancomponents are calculated
the phase. Hence, as 𝑋 = 𝑍 sin 𝜃
the impedance
and 𝑅 = 𝑍 cos 𝜃, where phase of the electrical impedance, 𝜃 = tan
magnitude obtained from AD5933 was taken for a different analysis of the experiments.
, and Z is th
The sample characteristics can be determined by obtaining the impedance spectroscopy
magnitude of the impedance. The influence of impedance magnitude in determining th
in a wide range of frequencies. The EIS characteristics of the samples were obtained by
samplefrequencies
varying characteristics
up to was found
100 kHz much by
controlled higher
the EISthan
tool.the phase. Hence, the impedanc
magnitude
To studyobtained from AD5933the
the EIS characteristics, was taken
output for a different
excitation was variedanalysis
from 0.4of the
to 2 experiments
Vpp
The sample characteristics can be determined by obtaining the impedance
and the spacing between two electrodes was varied from 0.3 to 1.5 cm, as shown in Figure spectroscopy
6.
in a result,
As wide range of frequencies.
the impedance The EISwith
was decreased characteristics
the increase of in the samples
excitation, onwere obtained by
the other
hand, the frequencies
varying impedance was up increased
to 100 kHz with the increase
controlled in the
by the EISseparation
tool. of the electrodes.
More than a 95% correlation was found, and the impedance profile indicated a good
conductivity distribution with the optimized output excitation of 2 Vpp and 1 cm spacing
of the electrodes.
 To study the EIS characteristics, the output excitation was varied from 0.4 to 2 Vpp
and the spacing between two electrodes was varied from 0.3 to 1.5 cm, as shown in Figure
6. As a result, the impedance was decreased with the increase in excitation, on the other
hand, the impedance was increased with the increase in the separation of the electrodes.
More than a 95% correlation was found, and the impedance profile indicated a good
Remote Sens. 2022, 14, 3214 10 of 28
conductivity distribution with the optimized output excitation of 2 Vpp and 1 cm spacing
of the electrodes.

50000 50000
2 Vpp_1 cm

Impedance (Z) in Ohm

Impedance (Z) in Ohm

40000 1 Vpp_1 cm 40000

0.4 Vpp_1 cm
30000 30000

20000 20000 y = -16318x + 49400

R² = 0.976
10000 10000

0 0
0 20000 40000 60000 80000 100000 0 1 2 3
Frequency (f) in Hz Output Excitation (Vout) in volt
(a) (b)
25000 25000
1.5 cm_2 Vpp

Impedance (Z) in Ohm

Impedance (Z) in Ohm

1 cm_2 Vpp y = 6675.7x + 9216.9

20000 20000
0.8 cm_2 Vpp R² = 0.944
15000 0.3 cm_2 Vpp 15000

10000 10000

5000 5000

0 0
0 20000 40000 60000 80000 100000 0 0.5 1 1.5 2
Frequency (f) in Hz Electrodes Separation (d) in cm
(c) (d)
Figure 6. EIS measurements of a carrot slice using two-electrode method (a) by varying output
Figure 6. EIS measurements of a carrot slice using two-electrode method (a) by varying output
excitation, and (b) the corresponding correlation. EIS measurements (c) by varying separation of the
excitation, and
electrodes, and(b)
(d)the
thecorresponding correlation.
corresponding correlation. EIS measurements
A good correlation with(c) by than
more varying
95% separation
coefficient of the
electrodes,
was found and
for(d)
the the corresponding
measurements usingcorrelation. A good
2 Vpp excitation and correlation
1 cm spacingwith more
of the than 95% coefficient
electrodes.
was found for the measurements using 2 Vpp excitation and 1 cm spacing of the electrodes.
In this work, two approaches were taken for the measurements in homogeneous and
In this work,media
inhomogeneous two approaches were taken
of the EIT domain. The EITfordomain
the measurements
was filled withinwater
homogeneous
to create and
inhomogeneous
the homogeneous media of A
media. thecontinuous
EIT domain. signalTheof 2EIT
Vppdomain waswas
excitation filled withtowater
applied to create
the EIT
thesensor system, and
homogeneous the measurements
media. A continuous were carried
signal of 2out.
Vpp At excitation
first, a two-pole method was
was applied to the EIT
applied,
sensor and EIS
system, andmeasurements
the measurementswere carried
wereout using an
carried EVAL-AD5933EBZ
out. At first, a two-pole impedance
method was
analyzer. The impedances for different layers of electrodes
applied, and EIS measurements were carried out using an EVAL-AD5933EBZ in the domain were obtained
impedance
for different
analyzer. frequencies offor
The impedances 1–100 kHz. A
different total of
layers of 64 measurements
electrodes in the (1-1,
domain1-2, 1-3,
were1-4, 1-5,
obtained for
1-6, 1-7, and 1-8 with respect to electrode 1; 2-1, 2-2, 2-3, 2-4, 2-5, 2-6, 2-7, and 2-8 with
different frequencies of 1–100 kHz. A total of 64 measurements (1-1, 1-2, 1-3, 1-4, 1-5, 1-6,
respect to electrode 2, and so on) were taken from the eight-electrode system in one layer
1-7, and 1-8 with respect to electrode 1; 2-1, 2-2, 2-3, 2-4, 2-5, 2-6, 2-7, and 2-8 with respect to
for 2D. In addition, the measurements were taken in other layers of the sensor electrode
electrode
array for2,3D.
and so on) were
A constant taken
current from
source of 1the
mAeight-electrode
sinusoidal current system in one frequency
with varying layer for 2D. In
addition,
was found suitable for EIT [10,11,23]. The current is injected through the electrodesarray
the measurements were taken in other layers of the sensor electrode of thefor 3D.
A constant current source of 1 mA sinusoidal current with varying
array and the obtained boundary voltages for a single layer of the EIT sensor are depictedfrequency was found
suitable for7.EIT [10,11,23]. The current is injected through the electrodes of the array and
in Figure
the obtained boundary voltages for a single layer of the EIT sensor are depicted in Figure 7.
In addition, a four-pole method was utilized for obtaining the reconstruction results at
multiple frequencies up to 100 kHz considering a 1 mA or above sinusoidal current. The
current was injected through the driving electrodes using a 15 MHz DDS signal generator at
different frequencies and the voltages were measured through the sensing electrodes. The
signal responses for 5 kHz and 1 mA are depicted as shown in Figure 8. In the 1st projection,
the current (I) was applied through neighboring electrodes 1-2 and five differential voltages
(V 1 , V 2 , V 3 , V 4 , and V 5 ) were measured between 3-4, 4-5, 5-6, 6-7, and 7-8 electrode pairs.
In the 2nd projection, the current was injected through 2-3 electrodes and the differential
voltages (V 1 –V 5 ) were collected from 4-5, 5-6, 6-7, 7-8, and 8-1 electrode pairs. Thus, for
the eight current projections a total of L(L-3) = 8 × 5 = 40 voltage measurements were
taken from a L = 8 electrode EIT system. The highest current density was found in the
driving electrodes and that decreases with the distance. Hence, the voltage was low at the
opposite sensing electrodes position from the current-driven electrodes. The variation of
the voltages was observed sinusoid as shown in Figure 8a.
 Number of Measurements
Figure 7. The boundary voltages for a homogeneous media (water) at 5 kHz, and 1 mA signal
considering two-pole measurements in one layer of EIT domain.

In addition, a four-pole method was utilized for obtaining the reconstruction results
Remote Sens.
Remote Sens. 2022, 14, x3214
2022, 14, FOR PEER REVIEW 1111ofof 28
28
at multiple frequencies up to 100 kHz considering a 1 mA or above sinusoidal current.
The current was injected through the driving electrodes using a 15 MHz DDS signal
generator at different frequencies and the voltages were measured through the sensing
7.8
electrodes. The signal responses for 5 kHz and 1 mA are depicted as shown in Figure 8. In
EIT_Two-pole

Boundary Voltages (V) in volt

the 1st projection, the current Water_5 (I) was applied
kHz_1 mA through neighboring electrodes 1-2 and five
7.6
differential voltages (V1, V2, V3, V4, and V5) were measured between 3-4, 4-5, 5-6, 6-7, and
7-8 electrode pairs. In the 2nd projection, the current was injected through 2-3 electrodes
7.4
and the differential voltages (V1–V5) were collected from 4-5, 5-6, 6-7, 7-8, and 8-1 electrode
pairs.
7.2 Thus, for the eight current projections a total of L(L-3) = 8 × 5 = 40 voltage
measurements were taken from a L = 8 electrode EIT system. The highest current density
was found
7 in the driving electrodes and that decreases with the distance. Hence, the
voltage was low at the opposite sensing electrodes position from the current-driven
6.8
electrodes. The variation of the voltages was observed sinusoid as shown in Figure 8a.
In another approach using the four-pole method, the current through neighboring
6.6
electrodes 1–2 and five differential voltages (V1-V5) were measured between 3-4, 3-5, 3-6,
3-7, and0 3-8 electrode
8 16 pairs.
24 32 Next,40the 48 current56 was 64 driven through 2-3 electrodes and the
Number of Measurements
voltages (V1-V5) were measured between 4-5, 4-6, 4-7, 4-8, and 4-1 electrode pairs. In the
similar7.
Figure
Figure 7.way
The40
The voltagevoltages
boundary
boundary measurements
voltages for were takenmedia
for aa homogeneous
homogeneous for the(water)
media eight electrode
(water) at
at 55 kHz, EIT
and 1system.
kHz, and 1 mA
mA signalThe
signal
measured
considering voltage
considering two-pole in the sensing
two-pole measurements
measurements in electrodes was increased
in one layer of EIT domain. with the increase in distance
when moving towards clockwise, and the results are presented in Figure 8b.
In addition, a four-pole method was utilized for obtaining the reconstruction results
0.6 at multiple frequencies up to 100 kHz2considering a 1 mA or above sinusoidal2current.
EIT_Four-pole_Approach 1 EIT_Four-pole_Approach
Sensing Voltages (V) in volt
Sensing voltages (V) in volt

0.5 The current was injected through the driving electrodes using a 15 kHz_1
Water_5 MHz DDS mA signal
Water_5 kHz_1 mA 1.6the voltages were measured
generator at different frequencies and through the sensing
0.4 electrodes. The signal responses for 5 kHz and 1 mA are depicted as shown in Figure 8. In
1.2
the 1st projection, the current (I) was applied through neighboring electrodes 1-2 and five
0.3
differential voltages (V1, V2, V3, V4, and V5) were measured between 3-4, 4-5, 5-6, 6-7, and
0.8
0.2 7-8 electrode pairs. In the 2nd projection, the current was injected through 2-3 electrodes
and the differential voltages (V1–V5) were collected from 4-5, 5-6, 6-7, 7-8, and 8-1 electrode
0.1 0.4
pairs. Thus, for the eight current projections a total of L(L-3) = 8 × 5 = 40 voltage
0 measurements were taken from a L = 8 0electrode EIT system. The highest current density
0 5 10 15 was 20 found25 30 in the35 driving
40 electrodes and0 that5 decreases 10 15 with 20 the25distance.
30 35Hence, 40 the
Number of Measurements
voltage was low at the opposite sensing electrodes Number of Measurements
position from the current-driven
electrodes. The variation of the voltages was observed sinusoid
(a) (b) as shown in Figure 8a.
In another approach using the four-pole method, the current through neighboring
Figure 8. The
electrodes 1–2 sensing voltages
and five for a homogeneous
differential voltages (V media
1-V5) (water) at 5 kHz and
were measured 1 mA signal
between consider-
3-4, 3-5, 3-6,
ing four-pole
3-7, measurements
and 3-8 electrode with
pairs. (a) 1st
Next, theapproach,
currentand was(b) 2nd approach
driven throughin2-3 oneelectrodes
layer of EITand
domain.
the
voltages (V1-V5) were measured between 4-5, 4-6, 4-7, 4-8, and 4-1 electrode pairs. In the
In another approach using the four-pole method, the current through neighboring
similar way 40 voltage measurements were taken for the eight electrode EIT system. The
electrodes 1–2 and five differential voltages (V 1 -V 5 ) were measured between 3-4, 3-5, 3-6,
measured voltage in the sensing electrodes was increased with the increase in distance
3-7, and 3-8 electrode pairs. Next, the current was driven through 2-3 electrodes and the
when moving towards clockwise, and the results are presented in Figure 8b.
voltages (V 1 -V 5 ) were measured between 4-5, 4-6, 4-7, 4-8, and 4-1 electrode pairs. In the
0.6 similar way 40 voltage measurements2were taken for the eight electrode EIT system. The
EIT_Four-pole_Approach
measured voltage 1in the sensing electrodes was increased EIT_Four-pole_Approach
with the increase in2 distance
Sensing Voltages (V) in volt
Sensing voltages (V) in volt

0.5 Water_5
when kHz_1towards
moving mA clockwise, and 1.6the results are presented Water_5
in Figure kHz_1
8b. mA
The carrot sample was then placed in the EIT domain filled with water. The impedances
0.4
of the inhomogeneous media (water1.2 + carrot) were measured by the designed electrode
0.3 system in the bottom layer of the array by varying the output excitation from 0.4 to 2 Vpp
at 5, and 80 kHz, respectively. A two-pole 0.8 method was applied for the measurements using
0.2 the designed multi-electrode sensor. The impedance was decreased by increasing frequency
0.1
and output excitation as well as shown 0.4 in Figure 9. The effect of output excitation may
vary based on the size of the electrodes, electrodes spacing, signal frequency, and type of
0 the samples placed in the domain. A large 0 oscillation with high impedance was observed
0 5 10 15 for 20 a low25 output
30 35excitation
40 of 0.4 Vpp, where 0 5the gain
10 and15 DFT 20 outputs
25 30oscillated
35 40highly
Number offor
Measurements
a given frequency. The results were found stable Number
at a of Measurements
2 Vpp excitation where a good
(a)conductivity of the sample was found, and all the experiments (b) were made accordingly
with this high excitation. The suitability of the electrode system was found by the sensor
 The boundary voltages were calculated for different stimulation currents by
measuring impedances in homogeneous (water) and inhomogeneous (water with sample)
media, and the difference of those was used to map the conductivity of the sample in the
domain. The voltages were normalized to obtain the conductivity of the sample. The
Remote Sens. 2022, 14, 3214 effects of impedances of electrodes and wires/clips were minimized accordingly, 12 ofand
28 a
good tomographic result of the sample was obtained. The voltage distributions using two-
pole sensing method was found more stable and uniform. The measured boundary data
played an important role in obtaining tomography and using the two-pole method the
characterization using AD5933. In a similar way, the other samples were taken in the
impedances of the
measurements. electrodes
Later, were found
the boundary verywere
potentials sensitive to theusing
calculated measurements for the
forward solve andtaken
a
objects in the EIT domain.
good reconstruction was made by the calculated conductivity.

80000 40000
Water+Carrot_5 kHz_2Vpp Water+Carrot_80 kHz_2Vpp
Impedance (Z) in Ohm

Impedance (Z) in Ohm

Water+Carrot_5 kHz_1Vpp Water+Carrot_80 kHz_1Vpp
60000 Water+Carrot_5 kHz_0.4Vpp 30000 Water+Carrot_80 kHz_0.4Vpp

40000 20000

20000 10000

0 0
0 8 16 24 32 40 48 56 64 0 8 16 24 32 40 48 56 64
Number of Measurements Number of Measurements

(a) (b)
Figure
Figure9.9.EIS
EISmeasurements for an
measurements for aninhomogeneous
inhomogeneousmedia media (water
(water + carrot)
+ carrot) in in
thethe domain
domain consisting
consisting
ofofeight electrodes by varying output excitation at (a) 5 kHz, and (b) 80 kHz, respectively.
eight electrodes by varying output excitation at (a) 5 kHz, and (b) 80 kHz, respectively.

3. Results
The boundary voltages were calculated for different stimulation currents by measuring
impedances in homogeneous
The designed (water)
electrode and inhomogeneous
array system was (water withconsidering
tested sample) media, and
multiple
the difference of those was used to map the conductivity of the sample in the domain.
inhomogeneities in the domain and the reconstruction was made in 2D and 3D planes in
The voltages were normalized to obtain the conductivity of the sample. The effects of
different experiments. The dimensional and frequency dependency of different plant
impedances of electrodes and wires/clips were minimized accordingly, and a good to-
roots were examined considering 3D image reconstruction.
mographic result of the sample was obtained. The voltage distributions using two-pole
sensing method was found more stable and uniform. The measured boundary data played
an important role in obtaining tomography and using the two-pole method the impedances
of the electrodes were found very sensitive to the measurements for the taken objects in the
EIT domain.

3. Results
The designed electrode array system was tested considering multiple inhomogeneities
in the domain and the reconstruction was made in 2D and 3D planes in different experi-
ments. The dimensional and frequency dependency of different plant roots were examined
considering 3D image reconstruction.

3.1. Image Reconstruction in 2D Plane

Experiment 1: In an experiment, the impedances were measured for a carrot slice
(height, h = 1/5-inch, and diameter, D = 25.4 mm) in a cylindrical domain using an EIT
electrode array system. The impedances were varied for different frequencies of 5–100 kHz
with the output excitation of 2 Vpp. A good conductivity distribution was found, and
the image reconstruction was carried out by placing the carrot slice at different positions
(pos: centre, 1, 4, and 7) in the domain. The results were reconstructed successfully using a
‘d2d4c’ FEM model for 80 kHz and a 1 mA stimulation current considering GN: Laplace
(λ = 0.57) in the inverse model as shown in Figure 10. The maximum change of conductivity
was found higher (0.725 S/m or above) at the side wall than the centre (0.61 S/m) of the
domain because the maximum current passes through the side wall.
Further, the stimulation current was varied for a domain with a carrot slice at position
1 and the new reconstructed results were obtained as shown in Figure 11. The conductivity
was increased from 0.77 to 0.85 S/m with the increase in current from 0.5 to 2 mA for a
constant hyperparameter value of λ = 0.57. The reconstructed results were also obtained by
varying hyperparameter values of 0.97, 0.57, and 0.17, respectively, at 1 mA for a domain
 Experiment 1: In an experiment, the impedances were measured for a carrot slice
(λ = 0.57) in the inverse model as shown in Figure 10. The maximum change of
(height, h = 1/5-inch, and diameter, D = 25.4 mm) in a cylindrical domain using an EIT
conductivity was found higher (0.725 S/m or above) at the side wall than the centre (0.61
electrode array system. The impedances were varied for different frequencies of 5–100
S/m) of the domain because the maximum current passes through the side wall.
kHz with the output excitation of 2 Vpp. A good conductivity distribution was found, and
the image reconstruction was carried out by placing the carrot slice at different positions
Remote Sens. 2022, 14, 3214 13 of 28
(pos: centre, 1, 4, and 7) in the domain. The results were reconstructed successfully using
a ‘d2d4c’ FEM model for 80 kHz and a 1 mA stimulation current considering GN: Laplace
(λ = 0.57) in the inverse model as shown in Figure 10. The maximum change of
with a carrot slice
conductivity was at position
found 7 as(0.725
higher shownS/min Figure 12. The
or above) conductivity
at the was the
side wall than increased
centre from
(0.61
0.71
S/m)toof0.9
theS/m withbecause
domain the decrease in the hyperparameter
the maximum current passesvalue.
through the side wall.

(a) (b) (c) (d)

(a) (b) (c) (d)

(e) (f) (g) (h)
Figure 10. A carrot slice at different positions of (a) pos: centre, (b) pos: 1, (c) pos: 4, and (d) pos: 7
in the EIT domain. (e–h) The corresponding reconstructed images for 80 kHz, 1 mA, and Laplace (λ
= 0.57). The obtained conductivity is high at the side wall, and it is low at the centre.

Further, the stimulation current was varied for a domain with a carrot slice at
position 1 and the new reconstructed results were obtained as shown in Figure 11. The
(e) (f) (g) (h)
conductivity was increased from 0.77 to 0.85 S/m with the increase in current from 0.5 to
Figure
2Figure
mA for 10. Aconstant
carrot slice at different positions
valueofofof(a) =pos: centre, (b) pos: 1, (c) pos: 4, andwere
(d) pos: 7
10.a A hyperparameter
carrot slice at different positions λ(a) 0.57.
pos: The reconstructed
centre, (b) pos: 1, (c)results
pos: 4, andalso
(d)
in the EIT by
obtained domain. (e–h) The corresponding reconstructed images for 0.17,
80 kHz, 1 mA, and Laplace (λ
pos: 7 in the EIT domain. (e–h) The corresponding reconstructed images for 80 kHz, 1 at
varying hyperparameter values of 0.97, 0.57, and respectively, 1 mA
mA, and
=
for0.57). The
a domain obtained
withTheconductivity
a carrot slice is high at the side
at positionis 7high wall,
as shown and it is
in Figurelow at the
12.itThecentre.
conductivity
Laplace (λ = 0.57). obtained conductivity at the side wall, and is low at the centre.was
increased from 0.71 to 0.9 S/m with the decrease in the hyperparameter value.
Further, the stimulation current was varied for a domain with a carrot slice at
position 1 and the new reconstructed results were obtained as shown in Figure 11. The
conductivity was increased from 0.77 to 0.85 S/m with the increase in current from 0.5 to
2 mA for a constant hyperparameter value of λ = 0.57. The reconstructed results were also
obtained by varying hyperparameter values of 0.97, 0.57, and 0.17, respectively, at 1 mA
for a domain with a carrot slice at position 7 as shown in Figure 12. The conductivity was
increased from 0.71 to 0.9 S/m with the decrease in the hyperparameter value.

(a) (b) (c) (d)

Figure
Figure 11.
11. A
Acarrot
carrotslice
sliceatat(a)
(a)pos:
pos:1,1,
and thethe
and reconstructed
reconstructed images forfor
images thethe
stimulation current
stimulation of (b)
current of
0.5
(b) 0.5 mA, (c) 1 mA, and (d) 2 mA at 80 kHz considering GN: Laplace (λ = 0.57). The shape of the
mA, (c) 1 mA, and (d) 2 mA at 80 kHz considering GN: Laplace (λ = 0.57). The shape of the
inhomogeneity
Remote Sens. 2022, 14, x FOR PEER REVIEW changes with
inhomogeneity changes withthetheincrease
increaseinincurrent
current(a(aslight
slightchange
changeisisobserved).
observed).The
The conductivity
14 of 28
conductivity is
is increased, and the obtained maximum conductivity values are 0.77, 0.79, and 0.85 S/m,
increased, and the obtained maximum conductivity values are 0.77, 0.79, and 0.85 S/m, respectively.
respectively.

(a) (b) (c) (d)

Figure 11. A carrot slice at (a) pos: 1, and the reconstructed images for the stimulation current of (b)
0.5 mA, (c) 1 mA, and (d) 2 mA at 80 kHz considering GN: Laplace (λ = 0.57). The shape of the
inhomogeneity changes with the increase in current (a slight change is observed). The conductivity
is increased, and the obtained maximum conductivity values are 0.77, 0.79, and 0.85 S/m,
respectively.
(a) (b) (c) (d)
Figure 12. A carrot slice at (a) pos: 7,7, and
and the
the reconstructed
reconstructed images for the hyperparameter value (λ)
of (b)
of (b) 0.97,
0.97,(c)
(c)0.57,
0.57,and
and(d)
(d)0.17
0.17atat8080 kHz
kHz and
and 1 mA
1 mA considering
considering Laplace
Laplace algorithm.
algorithm. TheThe shape
shape of
of the
the inhomogeneity changes with the decrease in hyperparameter value. The conductivity
inhomogeneity changes with the decrease in hyperparameter value. The conductivity is increased, is
increased, and the obtained maximum conductivity values are 0.71, 0.73, and 0.9 S/m, respectively.
and the obtained maximum conductivity values are 0.71, 0.73, and 0.9 S/m, respectively.

Experiment 2:
Experiment 2: In
In another
another experiment,
experiment, three
three different
different edible
edible plant
plant slices
slices of
of carrot,
carrot,
radish, and
radish, and potato
potatoof ofaa1/3-inch
1/3-inchheight
heightwith
withaadiameter,
diameter,DD= =25.4
25.4
mm mmofof each
each were
were taken
taken in
in the
the EITEIT domain
domain and their
and their reconstructed
reconstructed resultsresults were obtained
were obtained using theusing the impedance
impedance imaging
imaging technique
technique in the 2Dinplane
the 2D plane‘d2d4c’),
(using (using ‘d2d4c’),
as shownasinshown
Figurein13.
Figure 13.case,
In this In this
thecase, the
overall
overall size of all the plant slices were taken same. EIS measurements of the samples were
carried out from 5 to 100 kHz. A total of 64 measurements were taken for both
homogeneous and inhomogeneous conditions using 2 Vpp output excitation, and the
difference was taken for the reconstruction of each plant slice for 1 mA stimulation current
considering the GN: GHPF algorithm. Potato was found more conductive at a low
 imaging technique in the 2D plane (using ‘d2d4c’), as shown in Figure
overall size of all the plant slices were taken same. EIS measurements o
carried out from 5 to 100 kHz. A total of 64 measurements we
Remote Sens. 2022, 14, 3214 homogeneous and inhomogeneous conditions using 2 Vpp 14 ofoutput
28

difference was taken for the reconstruction of each plant slice for 1 mA
considering
size of all the plantthe
slicesGN: GHPF
were taken same.algorithm. Potato
EIS measurements was found
of the samples more co
were carried
frequency
out from 5 to 100of 5kHz.kHz andofa64good
A total conductivity
measurements distribution
were taken of carrot wa
for both homogeneous
and inhomogeneous conditions using 2 Vpp output excitation, and the difference was
frequencies. At 5 kHz, the conductivity of the samples was obtained
taken for the reconstruction of each plant slice for 1 mA stimulation current considering
S/m,
the GN: respectively,
GHPF algorithm.considering
Potato was found the hyperparameter
more value of
conductive at a low frequency of 5λkHz= 0.17. T
and a good conductivity distribution of carrot was found at the high frequencies. At
increases with the increase in frequency and dimension as well
5 kHz, the conductivity of the samples was obtained as 0.9, 1.1, and 3.6 S/m, respectively,
information can be obtained
considering the hyperparameter value offrom the
λ = 0.17. Thesample. At 80
sensing voltage kHz,with
increases thetheconduc
increase
was in frequency and
increased to dimension
2.8 S/m.as well An and more useful information
improved conductivity can be obtained
of carrot sl
from the sample. At 80 kHz, the conductivity of the carrot was increased to 2.8 S/m. An
increasing the height
improved conductivity from
of carrot slice1/5-inch
was found to 1/3-inch.theThe
by increasing conductivity
height from 1/5-inch of the
increased
to 1/3-inch. Thetoconductivity
9.1 S/m of bytheincreasing theincreased
carrot was further stimulation
to 9.1 S/mcurrent from 1 t
by increasing
the stimulation current from 1 to 2 mA. Thus, the frequency, dimensions, and stimulation
frequency, dimensions, and stimulation current played an important
current played an important role in achieving a good conductivity distribution and in
good
obtainingconductivity
clear reconstructeddistribution
images. and in obtaining clear reconstructed im

Figure 13. (a–c) Different edible plant slices (carrot, radish, and potato) placed in bottom layer of
Figure 13. (a–c) Different edible plant slices (carrot, radish, and potato) place
electrodes of a cylindrical domain, (d–f) the corresponding reconstructed images, and (g,h) their
electrodes of a cylindrical
impedance spectroscopy domain,
results. GHPF algorithm(d–f) theand
is utilized, corresponding reconstructed
potato is found more conductive im
impedance
at 5 kHz, 1 mA, spectroscopy results. In
and 2 Vpp output excitation. GHPF algorithm
addition, is utilized,
the carrot sample has good and potato is fou
conductivity
distribution at the high frequencies.

Experiment 3: Later, multi-target anomalies such as bio-, conductive-, and non-

conductive targets were considered in the designed EIT domain and simultaneous imaging
was carried out using the different measurement techniques. The boundary voltages for
multiple inhomogeneities such as metal, plastic, metal + plastic, and plastic + carrot in the
domain were calculated by measuring impedances at different frequencies up to 100 kHz. A
good reconstruction was made considering GN: GHPF with the optimized hyperparameter
values for 5 kHz and 1 mA current and the results are presented as shown in Figure 14.
 were identified successfully in the domain and the simultaneous two-dimensional
imaging with a good accuracy was possible using the designed electrode array system.
The metal and carrot are conductive, and the plastic is resistive. The metal has a higher
conductivity than the carrot. The overall conductivity of the carrot was increased by
decreasing the hyperparameter value (λ) and a clear reconstructed image was obtained.
Remote Sens. 2022, 14, 3214 15 of 28
The multiple inhomogeneities with different electrical properties in the domain were able
to distinguish using the developed EIT system.

(a) (b) (c) (d)

(e) (f) (g) (h)

Figure 14. (a) Metal, (b) Plastic, (c) Metal + Plastic, and (d) Plastic + Carrot inhomogeneities at the
Figure 14. (a) Metal, (b) Plastic, (c) Metal + Plastic, and (d) Plastic + Carrot inhomogeneities at the
centre of the EIT domain. The corresponding reconstructed images are presented at 5 kHz and 1
centre of the EIT domain. The corresponding reconstructed images are presented at 5 kHz and
mA considering GHPF and the hyperparameter value (λ) of (e) 0.17, (f) 0.17, (g) 0.17, and (h) 0.017,
1respectively.
mA consideringThe GHPF
type ofand the hyperparameter
inhomogeneity from the value (λ) ofis (e)
mixed 0.17, (f) 0.17,
identified (g) 0.17,with
successfully and (h)
the
0.017, respectively.
optimized The typevalue.
hyperparameter of inhomogeneity from the mixed is identified successfully with the
optimized hyperparameter value.
3.2. Image Reconstruction in 3D Plane
The conductivity level varies depending on the position and type of inhomogeneity
in theExperiment
domain, and 1: by
At varying
first, thethe
EISother
measurements
controllingfor a homogenous
parameters. A goodmedia considering
reconstruction
water
was in a using
made cylindrical domaingenerator
a ‘distmesh’ was carried out at
‘d2d4c’ fordifferent frequencies
a circular model. The of inhomogeneities
5–100 kHz using
the designed multi electrode sensor system. The measurements were
were identified successfully in the domain and the simultaneous two-dimensional also carried out for
imaging
an inhomogeneous media considering water and a carrot (length: L = 3 inch,
with a good accuracy was possible using the designed electrode array system. The metal and diameter:
D = carrot
and 31.75 mm) with a biomass
are conductive, and theweight
plasticofisWresistive.
= 54 g inThethemetal
givenhas
area of the conductivity
a higher domain. The
than the carrot. The overall conductivity of the carrot was increased by decreasingwere
carrot was considered as a sample of a tap root system. A total of 64 measurements the
taken from the eight-electrode
hyperparameter value (λ) and array system
a clear in the topimage
reconstructed layer.was
The obtained.
variation of impedances
The multiple
was observed sinusoid
inhomogeneities for the
with different measurements
electrical properties at different
in the domainelectrode
were ablepositions. The
to distinguish
impedance was decreased
using the developed EIT system. by increasing the frequency for a constant output voltage
excitation of 2 Vpp and the results are presented for a homogeneous media, as shown in
3.2. Image Reconstruction in 3D Plane
Experiment 1: At first, the EIS measurements for a homogenous media considering
water in a cylindrical domain was carried out at different frequencies of 5–100 kHz using
the designed multi electrode sensor system. The measurements were also carried out
for an inhomogeneous media considering water and a carrot (length: L = 3 inch, and
diameter: D = 31.75 mm) with a biomass weight of W = 54 g in the given area of the
domain. The carrot was considered as a sample of a tap root system. A total of 64 measure-
ments were taken from the eight-electrode array system in the top layer. The variation of
impedances was observed sinusoid for the measurements at different electrode positions.
The impedance was decreased by increasing the frequency for a constant output voltage
excitation of 2 Vpp and the results are presented for a homogeneous media, as shown in
Figure 15a. The impedance distribution in the top layer considering an inhomogeneous
media (water + carrot) is presented for 80 kHz and 2 Vpp excitation, as shown in Figure 15b.
In addition, the EIS measurements were taken in the top to bottom and bottom
layers of the electrode array in the domain by using an experimental set up as shown in
Figure 16. The boundary voltages were calculated with impedance data obtained from
homogeneous and inhomogeneous media of the EIT domain and those were normalized.
A 3D reconstruction of the carrot (L = 3-inch, W = 54 g) was made by calculating the
changes of conductivity from the difference of normalized voltages at an 80 kHz, 1 mA, and
2 Vpp output excitation as shown in Figure 16c. The prior NOSER algorithm performed
well compared to the other inverse algorithms for 3D imaging and considered with the
Remote Sens. 2022, 14, 3214 16 of 28

optimized hyperparameter value of λ = 2.17. The conductivity of the carrot was obtained
considering the layers of electrodes at z = 1, and z = 0.3, respectively, with a virtual16layer
Remote Sens. 2022, 14, x FOR PEER REVIEW of 28
at z = 0.65 in the cylindrical FEM mesh. It was found that day wise (day 1 to 3) the
biomass weight of the carrot sample was decreased from W = 54 g (L = 3 inch) to W = 34 g
(L = 2.5 inch) at room temperature (20 ◦ C) and the average impedance was increased at
Figure 15a.
different The impedance
frequencies, as showndistribution
in Figure in theThe
16b. topmoisture
layer considering an inhomogeneous
level was reduced along with
media (water + carrot) is presented for 80 kHz and 2 Vpp excitation, as shown in Figure
the dimension of the carrot and the corresponding tomography was distorted by decreasing
15b.
the conductivity of the sample with time as shown in Figure 16d.

10000 10000
Vout = 2Vpp Water+Carrot_80 kHz_2Vpp

Top Layer Impedance (Z) in Ohm

Top Layer Impedance (Z) in Ohm

8000 8000

6000 6000

4000 4000
Water_5 kHz Water_15 kHz
2000 Water_25 kHz Water_40 kHz 2000
Water_60 kHz Water_80 kHz
Water_100 kHz
0 0
0 8 16 24 32 40 48 56 64 0 8 16 24 32 40 48 56 64
Number of Measurements Number of Measurements

(a) (b)
Figure 15.
Figure 15. Impedance
Impedance spectroscopy for (a)
spectroscopy forhomogeneous (water),
(a) homogeneous and (b)and
(water), inhomogeneous (water +
(b) inhomogeneous
carrot) media in the top layer of the cylindrical domain at 2 Vpp output excitation. Impedance
(water + carrot) media in the top layer of the cylindrical domain at 2 Vpp output excitation. is
decreased in different electrodes position by increasing frequency.
Impedance is decreased in different electrodes position by increasing frequency.

In addition,
Later, the EIS measurements
four different carrot sampleswere taken in
of various the topweights
biomass to bottom (Wand bottom
= 54, 88, 52,layers
and
of the electrode array in the domain by using an experimental set up as
46 g) and dimensions (length, L = 3, 3.1, 2.75, and 2.8 inches with diameter of D = 31.75 mm shown in Figure
16. The
each) were boundary
taken, and voltages were calculated
the correlation with impedance
was established data obtained
with the measured impedances.from
homogeneous
The layer-wise and inhomogeneous
impedance distributionmedia
in a of the EIT domain
cylindrical domainand wasthose
takenwereand normalized.
the average
A 3D reconstruction
impedances of the fullofmodel
the carrot (L = 3-inch,
were correlated W fresh
with = 54 g) was made
biomass weights by calculating
and lengthsthe of
changes of conductivity from the difference of normalized voltages
the carrots. At 5 kHz, more than an 85% correlation was found with weight and length of at an 80 kHz, 1 mA,
andcarrots,
the 2 Vppasoutput
shown in excitation
Figure 17.asThe shown in impedance
average Figure 16c.was Thedecreased
prior NOSER with thealgorithm
increase
performed
in weight and well compared
length of theto the other
carrot, and inverse algorithms
the carrot sample was for 3D imaging
found more and considered
conductive. In
with the optimized
addition, an improved hyperparameter
correlation of value
more ofthanλ =a2.17.
90% The
coefficient (R2 = 0.822)
conductivity of thewas
carrot was
found
obtainedthe
between considering the layers and
average impedance of electrodes at z = 1,
biomass weight of and z = 0.3,
the carrot at respectively,
a high frequency withofa
virtual
60 kHz,layer
where atRz2=is0.65
the in the cylindrical
coefficient FEM mesh.and
of determination It was
R isfound that day wise
the correlation (day 1 to
coefficient. It
indicates that the biomass of the carrot is dependent on not only the
3) the biomass weight of the carrot sample was decreased from W = 54 g (L = 3 inch) to W dimensions of the
sample
= 34 g (Lbut
= 2.5also the at
inch) frequencies of the signal.
room temperature (20 °C) and the average impedance was increased
Experiment
at different 2: In an as
frequencies, experiment, a potato
shown in Figure plant
16b. Thewas grownlevel
moisture at roomwastemperature
reduced along as
shown
with thein Figure
dimension18, andof the root of the
carrot andplant species was taken
the corresponding as one of the
tomography wassamples for the
distorted by
analysis
decreasingof 3Dtheimage reconstruction
conductivity using the
of the sample withdeveloped EIT data
time as shown acquisition
in Figure 16d. system. The
root structure, dimensions, and variation of frequencies were examined on obtaining the
tomography result of the root by measuring the electrical impedances using the designed
EIT sensor system. The length of the root was measured as 10 inches when the shoot
length of the plant was 16 inches. In addition, the root system of the potato species is tap
root, which includes an old seed piece. The potato root was placed in water media of the
EIT domain for the measurements of impedances at room temperature (20 ◦ C) and the
reconstruction was made using the developed EIT system as shown in Figure 19a.
Remote Sens.Sens.
Remote 2022,2022,
14, x14,
FOR3214PEER REVIEW 17 of 2817 of 28

Figure 16. (a) A 3D EIT experimental setup for obtaining impedance tomography of a carrot sample
(tap root) in a cylindrical domain. (b) Average impedance profile of the sample by varying frequency
Figure 16. (a) A 3D EIT experimental setup for obtaining impedance tomography of a carrot sample
in different days (day 1 and day 3). The reconstructed images of the sample obtained in (c) day 1,
(tap root) in a cylindrical domain. (b) Average impedance profile of the sample by varying
and (d) day 3 at 80 kHz, 1 mA, and 2 Vpp excitation considering GN: NOSER (λ = 2.17). The average
frequency in different days (day 1 and day 3). The reconstructed images of the sample obtained in
impedance of the carrot is decreased with the increase in frequency and the impedance is increased
(c) day 1, and (d) day 3 at 80 kHz, 1 mA, and 2 Vpp excitation considering GN: NOSER (λ = 2.17).
at different frequencies with time. The maximum conductivity of the sample is decreased from 0.033
The average impedance of the carrot is decreased with the increase in frequency and the impedance
to 0.02 by reducing the biomass weight and dimension with time. It is found that biomass of a carrot
is increased at different frequencies with time. The maximum conductivity of the sample is
has dimensional dependency.
decreased from 0.033 to 0.02 by reducing the biomass weight and dimension with time. It is found
that biomass of a carrot has
The impedances dimensional
were measureddependency.
for homogeneous (water) and inhomogeneous
(water + potato root) conditions by varying frequencies from 5 to 100 kHz using an EIS
toolLater, fourconnected
(AD5933) different to
carrot samples
the EIT sensor.ofThe
various biomass
impedances wereweights (W =at54,
measured a 288,
Vpp52, and
46excitation
g) and dimensions (length,
in different layers suchLas
= top,
3, 3.1,
top2.75, and 2.8
to bottom, andinches
bottomwith diameter
layers of D = 31.75
of the electrode
mm each)
array. Thewere taken,
average and the correlation
impedances obtained forwastheestablished
inhomogeneouswithmedia
the measured impedances.
are presented in
Figure
The 19b andimpedance
layer-wise the impedances were found
distribution to decrease domain
in a cylindrical with the wasincrease
takenin and
frequency.
the average
Layers-wise, the measured impedances are presented in Figure 19c. The boundary
impedances of the full model were correlated with fresh biomass weights and lengths of voltages
were calculated for a 1 mA stimulation current and normalized. The differences of the
the carrots. At 5 kHz, more than an 85% correlation was found with weight and length of
voltages were utilized to calculate the changes of conductivity for the given root system in
the carrots, as shown in Figure 17. The average impedance was decreased with the
the EIT domain, which represents the tomography. A tomography of the potato root system
increase
was made inatweight and1 length
80 kHz and mA using of the
thedifference
carrot, method
and thebycarrot
applying sample was found
the one-step GN al- more
conductive. In addition,
gorithm: NOSER with the an improved
optimized correlationvalue
hyperparameter of more
of 2.17than a 90%
as shown in coefficient
Figure 19d. (R =
2

0.822) was found between the average impedance and biomass weight of the carrot at a
high frequency of 60 kHz, where R2 is the coefficient of determination and R is the
correlation coefficient. It indicates that the biomass of the carrot is dependent on not only
the dimensions of the sample but also the frequencies of the signal.
 Remote Sens. 2022, 14, 3214 18 of 28

A conductive behaviour of the root was found with the maximum changes of conductivity
of 0.003 for the selected electrode positions at z = 1 and 0.3 vertically (including a virtual
Remote Sens. 2022, 14, x FOR PEER REVIEW
layer at z = 0.65) in the electrode array system. A tomography of the potato root system18 of 28
was made successfully using FEM modeling.

Remote Sens. 2022, 14, x FOR PEER REVIEW 19 of 28

The fresh weight of the root was measured as 46 g when it was separated from the
shoot
Figureof17.
the plant. During
Correlation the measurements,
between the root sample
(a) average impedance did not
and biomass sufferand
weight, any(b)physical
average
damage
impedanceandandwaslength
activeofin
theoperation.
carrot at 5Later,
kHz. the root was
A negative kept in the
correlation free space
is found between at the
roomcar-
Figure 17. Correlation between (a) average impedance and biomass weight, and (b) average
temperature,
rot size (suchthe root was
as weight anddried by and
length) removing
averagethe moisture,An
impedance. andimproved
day-wisecorrelation
the weight was
is found
impedance and length of the carrot at 5 kHz. A negative correlation is found between the carrot size
measured.
between (c)The biomass
average weightand
impedance wasbiomass
reduced to less
weight of than
the 20 g inataaweek, where excluding
(such as weight and length) and average impedance. An sample
improved high frequency
correlation of 60 kHz.
is found between (c)
the old
Overall,seed,
the the actual
average root
impedance was
is found
decreasedas 2
by g only in
increasing dry condition.
frequency and The
the severalisplants
correlation of
improved
average impedance and biomass weight of the sample at a high frequency of 60 kHz. Overall, the
the
for same can be
estimating the taken
carrot in a biomass.
root further experiment for modeling the measured impedances
average impedance is decreased by increasing frequency and the correlation is improved for
with fresh or dry biomass weights of the roots.
estimating the carrot root biomass.

Experiment 2: In an experiment, a potato plant was grown at room temperature as

shown in Figure 18, and the root of the plant species was taken as one of the samples for
the analysis of 3D image reconstruction using the developed EIT data acquisition system.
The root structure, dimensions, and variation of frequencies were examined on obtaining
the tomography result of the root by measuring the electrical impedances using the
designed EIT sensor system. The length of the root was measured as 10 inches when the
shoot length of the plant was 16 inches. In addition, the root system of the potato species
is tap root, which includes an old seed piece. The potato root was placed in water media
of the EIT domain for the measurements of impedances at room temperature (20 °C) and
the reconstruction was made using the developed EIT system as shown in Figure 19a.
The impedances were measured for homogeneous (water) and inhomogeneous
(water + potato root) conditions by varying frequencies from 5 to 100 kHz using an EIS
tool (AD5933) connected to the EIT sensor. The impedances were measured at a 2 Vpp
excitation
Figure
Figure18.
in different
18.(a)
(a)AAgrown
layers
grownpotato
such
potatoplant,
plant,and
as(b)
and(b)
top, top to
a apotato
potatoroot
bottom,
rootsystem
and bottom layers of the electrode
systemininplant.
plant.
array. The average impedances obtained for the inhomogeneous media are presented in
FigureThe19bfresh
andweight
the impedances
of the root waswere found to
measured as decrease
46 g whenwith the
it was increasefrom
separated in frequency.
the
shoot of the plant.
Layers-wise, During the measurements,
the measured impedances are the root sample did
presented not suffer
in Figure anyThe
19c. physical
boundary
voltages were calculated for a 1 mA stimulation current and normalized. The differences
of the voltages were utilized to calculate the changes of conductivity for the given root
system in the EIT domain, which represents the tomography. A tomography of the potato
root system was made at 80 kHz and 1 mA using the difference method by applying the
Remote Sens. 2022, 14, 3214 19 of 28

damage and was active in operation. Later, the root was kept in the free space at room
temperature, the root was dried by removing the moisture, and day-wise the weight was
measured. The biomass weight was reduced to less than 20 g in a week, where excluding
the old seed, the actual root was found as 2 g only in dry condition. The several plants of
the same can be taken in a further experiment for modeling the measured impedances with
fresh or 18.
Figure dry(a)
biomass
A grownweights of theand
potato plant, roots.
(b) a potato root system in plant.

(a) (b)

(c) (d)
Figure
Figure 19.19.(a)(a)An
AnEITEITexperimental
experimentalsetup
setupfor
forthe
themeasurements
measurementsofofa apotato
potatoroot
root system,
system, (b)
(b) average
average
measured impedances for 5–100 kHz of an inhomogeneous media (water + root), (c) layer-wise
measured impedances for 5–100 kHz of an inhomogeneous media (water + root), (c) layer-wise
measured impedances at 80 kHz, and (d) reconstructed potato root (tap root) using 3D EIT
measured impedances at 80 kHz, and (d) reconstructed potato root (tap root) using 3D EIT measure-
measurements in water at 80 kHz, 2 Vpp, and 1 mA stimulation current considering GN: NOSER (λ
ments in water
= 2.17). at 80 kHz,
The obtained 2 Vpp,
result anda1 conductive
shows mA stimulation current
behaviour ofconsidering
the root withGN: NOSER (λ
maximum = 2.17).of
changes
The obtained result
conductivity shows a conductive behaviour of the root with maximum changes of conductivity
of 0.003.
of 0.003.

Experiment 3: In another experiment, the root impedance of a Chinese money plant

(Pilea) species was measured by placing the sample in the designed cylindrical EIT domain
with the target of obtaining 3D reconstruction, as shown in Figure 20. The homogeneous
media was created by filling the water in the domain and the measurements were taken
without and with the root accordingly. The difference was used to calculate the root
conductivity. The impedances for two different media were measured by the electrode
layers of the domain using AD5933 from 5 to 100 kHz at room temperature (20 ◦ C) and the
corresponding boundary potentials were calculated at 1 mA. The changes of conductivity
of the root were calculated by the differences of normalized boundary voltages.
The obtained tomography results are presented at 5 kHz (Vout = 2 Vpp), and 80 kHz
(Vout = 0.4 Vpp), respectively. After optimizing, good results were found with the selected
frequencies and excitations considering GN: NOSER (λ = 2.17) in the inverse model. The
maximum changes of conductivity of 0.018 was obtained at 5 kHz and 2 Vpp, as shown
in Figure 20d. The impedance with a low excitation of 0.4 Vpp was minimized by in-
creasing the frequency to 80 kHz and the maximum changes of conductivity of the root
was increased to 0.08, as shown in Figure 20e. As a result, the tomography of the root
was made successfully using FEM modeling. The conductivity was varied at different
 frequencies and excitations considering GN: NOSER (λ = 2.17) in the inverse model. The
maximum changes of conductivity of 0.018 was obtained at 5 kHz and 2 Vpp, as shown
in Figure 20d. The impedance with a low excitation of 0.4 Vpp was minimized by
increasing the frequency to 80 kHz and the maximum changes of conductivity of the root
was2022,
Remote Sens. increased
14, 3214 to 0.08, as shown in Figure 20e. As a result, the tomography of the root was20 of 28
made successfully using FEM modeling. The conductivity was varied at different vertical
electrode positions of z = 1, and 0.3 including a virtual layer of z = 0.65. A higher
conductivity is found vertical electrode
at the bottompositions z = 1,
layer ofofthe and 0.3 including
electrodes’ a virtual
position where layer z = 0.65. A higher
theofimpedance
conductivity is found at the bottom layer of the electrodes’ position where the impedance is
is low because of distributed roots. It is concluded that the impedance varies by the
low because of distributed roots. It is concluded that the impedance varies by the variation
variation of root density
of root and distribution
density of the
and distribution roots,
of the roots,and
andthe variations
the variations in in tomography
tomography results are
results are obtainedobtained
accordingly.
accordingly.

(a) (b) (c)

(d) (e)
Figure 20. 3D imaging of 20.
Figure a 3DPilea rootof in
imaging a cylindrical
a Pilea domain
root in a cylindrical using
domain usingEIT measurements
EIT measurements and
and considering
considering GN: NOSER inverseinverse
GN: NOSER algorithm (λ =(λ2.17).
algorithm = 2.17).(a)(a)Measuring root
Measuring root impedance,
impedance, (b) Chinese
(b) Chinese money plant
money plant (Pilea), (c) Pilea
(Pilea), (c)root,
Pilea and the the
root, and reconstructed
reconstructed root images
root images at 1atmA
1 mA for5 kHz
for (d) (d) 5(2kHz
Vpp),(2and
Vpp),
(e) 80 kHz
and (e) 80 kHz (0.4 (0.4 Vpp),respectively.
Vpp), respectively. AnAnimproved conductivity
improved of the root of
conductivity wasthefound by increasing
root was found the frequency.
by
increasing the frequency.
The fresh weight of the root was measured as 90 g when it was separated from the
plant. During the measurements, the root sample did not suffer any physical damage and
The fresh weight
wasof the in
active root was measured
operation. as 90
Later, the EIS g when it was
measurements of theseparated fromcarried
Pilea root were the out
plant. During the measurements, the root sample
using two ECG electrodes connecteddid notEIS
to an suffer any physical
tool (AD5933) from 5damage
kHz to 100 andkHz by
varying the output excitation from 0.4 to 2 Vpp, as shown in Figure
was active in operation. Later, the EIS measurements of the Pilea root were carried out 21. The root impedance
was increased with the increase in electrodes separation and a good positive correlation
using two ECG electrodes connected to an EIS tool (AD5933) from 5 kHz to 100 kHz by
was found for 3 cm spacing. On the other hand, the root impedance was decreased with
the increase in frequency and the output excitation as well. A small variation of the root
impedances was found by varying the output excitation and the correlation was negative,
as shown in Figure 21b. After optimizing, a good match was found with the selected
spacing of the electrodes considering the output excitation of 2 Vpp, and more than a 94%
correlation was found at 5 kHz.
The EIS measurements of the Pilea root were taken on different days and the results
were correlated with the biomass weights. The impedances were found to increase by
decreasing the biomass of the root. The moisture level of the root was reduced with time
and, hence, the conductivity of the root was decreased. The biomass weight of the root
was reduced from 90 to 30 g in three weeks at room temperature (20 ◦ C) and those were
predicted by the measured impedances. The regression analysis was performed using the
measured root biomass and impedances obtained in 20 days when it was kept at free space.
 𝑌 = 𝜔 + 𝜔 𝑍 + 𝜔 𝑍 + ⋯+ 𝜔 𝑍 (14)
where Zf1, Zf2, …., Zfk are the average impedances for k number of features of f1 to fk. The
intercept is 𝜔 , and 𝜔 , 𝜔 , … , 𝜔 are the coefficients.
Remote Sens. 2022, 14, 3214 A dataset suffers from the multicollinearity problem: (i) if the correlation coefficient 21 of 28
(R) between the explanatory variables is close to one, (ii) if there is no change in the
coefficient of determination (R2) after adding an independent variable, and (iii) if the
tolerance value (tolerance =21−R2) is less than 0.1 and the variance inflation factor (VIF =
A good correlation with R = 0.983 and RMSE of 3.75 was found at 5–7 kHz considering
1/tolerance) is greater than 10. At first, the highly correlated features of more than a 95%
n = 8 samples and selected features of k = 4 in the dataset, as shown in Figure 21d. A
correlation coefficient were removed. Later, the Wrapper backward elimination method
statistical analysis was performed with the obtained data using PrimaXL Data Analysis
was applied considering the probability of rejection of the null hypothesis p ≤ 0.05 using
ToolPak [31,33]. A multiple linear regression analysis was performed for the obtained
an individual t-test. After few iterations, the training and validation was performed
dataset considering the least square method [31,33] with the help of Equation (14).
considering the overall F-test (p ≤ 0.05) and the features (5, 5.5, 6, and 7 kHz) were selected
accordingly. Finally, a model Ŷ =was
ω0 +extracted
ω1 Z f 1 +for
ω2 the
Z f 2 estimation
+ . . . + ωk Zoff kbiomass of a Pilea root,
(14)
as shown in Equation (15).
where Zf1 , Zf2 , . . . ., Zfk are the average impedances for k number of features of f1 to fk. The
𝐵𝑖𝑜𝑚𝑎𝑠𝑠 = 70.9851
intercept is ω− 0.8384𝑍 + 0.14795𝑍 . coefficients.
+ 1.49971𝑍 − 0.8017𝑍 (15)
0 , and ω , ω2 , . . . , ω are the
1 k

15500
Signal frequency,f=5 kHz

Impedance (Z) in Ohm

Electrodes spacing, d=3 cm
15000

14500 y = -543.36x + 15557

R² = 0.887

14000
Remote Sens. 2022, 14, x FOR PEER REVIEW 0 1 2 322 of 28
Output Excitation (Vout) in volt
(a) (b)
100 100
90 R² = 0.983
Predicted Biomass (gram)

80 80 Adjusted R2=0.96,
Root Biomass (gram)

RMSE=3.75
70
p-value=0.0057
60 60
(n=8, k=4)
50
40 40
30
20 20
10
0 0
0 2 4 6 8 10 0 20 40 60 80 100
Number of Measurements Actual Biomass (gram)
(c) (d)
Figure21.
Figure 21.(a)(a)
EISEIS measurement
measurement of aof a Pilea
Pilea root using
root using AD5933EBZ,
AD5933EBZ, and (b)and (b) the corresponding
the corresponding correla-
correlation with output excitation. (c) Day-wise the biomass of the root was decreased for different
tion with output excitation. (c) Day-wise the biomass of the root was decreased for different measure-
measurements, and the corresponding impedance was increased. (d) A good correlation with R2 =
ments, and the corresponding impedance was increased. (d) A good correlation with R2 = 0.983 and
0.983 and RMSE of 3.75 was obtained by the predicted model at 5–7 kHz considering 2 Vpp
RMSE of 3.75 was obtained by the predicted model at 5–7 kHz considering 2 Vpp excitation.
excitation.

A dataset suffers from the multicollinearity problem: (i) if the correlation coefficient (R)
4. Discussion
between the explanatory variables is close to one, (ii) if there is no change in the coefficient
The reconstruction
of determination (R2 ) afterperformance using the developed
adding an independent EIT(iii)
variable, and data acquisition
if the tolerancesystem
value
was evaluated
(tolerance = 1 −asRfollows:
2 ) is less than 0.1 and the variance inflation factor (VIF = 1/tolerance)
Variation
is greater of inverse
than 10. At first,solvers: Thecorrelated
the highly reconstruction performance
features of more than using 2D EIT
a 95% domain
correlation
was evaluated by varying inverse solvers. Three different solvers such
coefficient were removed. Later, the Wrapper backward elimination method was applied as Gauss–Newton
(GN): NOSER,
considering the back-projection (BP): naive,
probability of rejection of theand
nulltotal variancep(TV)
hypothesis ≤ 0.05described
using aninindividual
Equations
(9)–(12) were utilized for solving the EIT inverse problem. A comparative
t-test. After few iterations, the training and validation was performed considering study among
the
the solvers
overall was
F-test made
(p ≤ 0.05) using
and athe
‘c2c2’ FEM (5,
features model
5.5, (nodes:
6, and 7313,
kHz)elements: 576, and
were selected boundary:
accordingly.
48). Thea changes
Finally, model was areextracted
more observable for a lower
for the estimation of vertex
biomassdensity (such
of a Pilea as c)
root, asof an FEM
shown in
model. A(15).
Equation carrot slice (h = 1/5-inch, and D = 25.4 mm) was placed at the centre of the domain
and the impedances were measured using the designed EIT sensor for 80 kHz and 2 Vpp
excitation. The reconstruction was made at 1 mA, as shown in Figure 22. The obtained
changes of conductivity results are presented in Table 1. GN (λ = 0.17) performed well
compared to the other solvers, whereas the reconstructed image using unfiltered BP was
found noisy due to the fluctuation of amplitudes, and a lower sized image was
reconstructed using TV. Filtered BP can give linear amplitude scale with less noise. By
 Remote Sens. 2022, 14, 3214 22 of 28

Biomass Pilea Root = 70.9851 − 0.8384Z5kHz + 0.14795Z5.5kHz + 1.49971Z6kHz − 0.8017Z7kHz (15)

4. Discussion
The reconstruction performance using the developed EIT data acquisition system was
evaluated as follows:
Variation of inverse solvers: The reconstruction performance using 2D EIT domain was
evaluated by varying inverse solvers. Three different solvers such as Gauss–Newton (GN):
NOSER, back-projection (BP): naive, and total variance (TV) described in Equations (9)–(12)
were utilized for solving the EIT inverse problem. A comparative study among the solvers
was made using a ‘c2c2’ FEM model (nodes: 313, elements: 576, and boundary: 48). The
changes are more observable for a lower vertex density (such as c) of an FEM model. A
carrot slice (h = 1/5-inch, and D = 25.4 mm) was placed at the centre of the domain and the
impedances were measured using the designed EIT sensor for 80 kHz and 2 Vpp excitation.
The reconstruction was made at 1 mA, as shown in Figure 22. The obtained changes of
conductivity results are presented in Table 1. GN (λ = 0.17) performed well compared to
the other solvers, whereas the reconstructed image using unfiltered BP was found noisy
due to the fluctuation of amplitudes, and a lower sized image was reconstructed using TV.
14, x FOR PEER REVIEW Filtered BP can give linear amplitude scale with less noise. By increasing the iteration,
23 of 28the
reconstruction performance was found to improve for TV. Finally, GN algorithms were
chosen and utilized in different experiments to fulfill the objectives.

(b) (c)

(a)

(d) (e)
Figure 22. (a) A carrot slice22.at (a)
Figure theAcentre
carrot of theatEIT
slice the domain. TheEIT
centre of the corresponding reconstructed
domain. The corresponding images
reconstructed
for different inverseimages
solvers of (b) GN
for different (NOSER),
inverse (c)(b)BP
solvers of GN(Naive),
(NOSER), (d)
(c)TV (Iteration
BP (Naive), = 1),
(d) TV and (e)= TV
(Iteration 1), and
(Iterations = 2) at 80 kHz, 1 mA, and 2 Vpp excitation. Overall, GN: NOSER performed wellwell
(e) TV (Iterations = 2) at 80 kHz, 1 mA, and 2 Vpp excitation. Overall, GN: NOSER performed
considering λ = 0.17, a better shape
considering and
λ = 0.17, sizeshape
a better of theand
carrot
size ofwere reconstructed
the carrot with less
were reconstructed withnoise.
less noise.

Table 1. CalculatedTable
mean1. Calculated mean and
and standard standard deviation
deviation (SD) of(SD) of obtained
obtained changesof
changes of conductivity
conductivityby varying
by
inverse solvers considering a carrot slice in the EIT domain.
varying inverse solvers considering a carrot slice in the EIT domain.
GN (NOSER) BP (Naïve) TV (Iteration = 1)
GN (NOSER) BP (Naïve) TV (Iteration = 1)
Mean (S/m) 0.482597 0.374112 0.676118
Mean (S/m) 0.482597 0.374112 0.676118
SD (S/m) 0.643679 0.092603 1.065166
SD (S/m) 0.643679 0.092603 1.065166

Variation of FEM models: The carrot slice (h = 1/5-inch, and D = 25.4 mm) at the centre
of the EIT domain was reconstructed onto elements and nodes in the frame by varying
FEM models. The results were observable using an ‘a2c0’ (nodes: 41, elements: 64,
Remote Sens. 2022, 14, 3214 23 of 28

Variation of FEM models: The carrot slice (h = 1/5-inch, and D = 25.4 mm) at the
centre of the EIT domain was reconstructed onto elements and nodes in the frame by
varying FEM models. The results were observable using an ‘a2c0’ (nodes: 41, elements:
64, boundary: 16) or a ‘b2c0’ (nodes: 145, elements: 256, boundary: 32) model and a
comparative study was made, as shown in Figure 23. The reconstruction onto elements
contributed to obtaining clear images. On the other hand, the reconstructed image onto
nodes was found blurred because of the lower number of calculated conductivity data than
using elements of the model as shown in Figure 23c,d. The model with a higher vertex
density (such as d) represented the finer mesh that can provide a more accurate result in
reconstruction than the coarse mesh obtained using a lower vertex density (such as a or
b). A good reconstruction of the carrot slice was observed by increasing the vertex density,
although the overall conductivity was reduced. The number of conductivity data were
increased with the increased elements of the ‘b2c0’ model. Finally, the ‘distmesh’ model
using ‘d2d4c’ (nodes: 2507, elements: 4757, boundary: 255) performed well compared to
the other FEM models and a good reconstruction was made with more confined results
using the optimized GN algorithm: GHPF (λ = 0.17). A good conductivity distribution
Remote Sens. 2022, 14, x FOR PEER REVIEW 24 of 28
was found by increasing the number of elements in the frame. The obtained changes of
conductivity results are presented in Table 2.

(a) (b) (c) (d)

(e) (f) (g)

Figure 23. Reconstructed images of a carrot slice at the centre of an EIT domain for different FEM
Figure 23. Reconstructed images of a carrot slice at the centre of an EIT domain for different FEM
models at 80 kHz, 1 mA, and 2 Vpp excitation. (a) FEM (a2c0, 8)-elements, (b) corresponding
models at 80 kHz, 1 mA, and 2 Vpp excitation. (a) FEM (a2c0, 8)-elements, (b) corresponding
reconstructed slices, (c) FEM (a2c0, 8)-nodes, (d) corresponding reconstructed slices, (e) FEM (b2c0,
reconstructed(f)slices,
8)-elements, FEM(c) FEM (a2c0,
(d2d4c, 8)-nodes,
8)-elements, (d)(g)
and corresponding reconstructed
corresponding slices,
reconstructed (e) FEM
slices. (b2c0,
A blurred
8)-elements, (f) FEM (d2d4c, 8)-elements, and (g) corresponding reconstructed slices.
reconstructed image was observed for the reconstruction onto nodes, on the other hand, a good A blurred
reconstructed image
reconstruction was observed
was made for theAreconstruction
onto elements. onto nodes,
clear reconstructed on the
image of the carrot
other hand, a good
slice with an
accurate shape was
reconstruction was found
made using a ‘distmesh’
onto elements. A generator FEM model
clear reconstructed such of
image as the
‘d2d4c’ considering
carrot slice withthe
an
reconstruction ontofound
accurate shape was elements.
usingOverall, the generator
a ‘distmesh’ error wasFEMdecreased by increasing
model such as ‘d2d4c’ the number the
considering of
elements.
reconstruction onto elements. Overall, the error was decreased by increasing the number of elements.

Variation of number of measurements in the frame: In addition, the number of

measurements was reduced from 64 to 40 from the dataset with the target of minimizing
the overlapping. In this case, the selected ports (3-4, 3-5, 3-6, 3-7, and 3-8 electrode pairs
with respect to electrode 3; 4-5, 4-6, 4-7, 4-8, and 4-1 electrode pairs with respect to
electrode 4; and so on) are exactly the same as considered in the four-pole method. Finally,
the overlapping was reduced fully considering 28 measurements using the selected ports
(1-2, 1-3, 1-4, 1-5, 1-6, 1-7, and 1-8 electrode pairs with respect to electrode 1; 2-3, 2-4, 2-5,
2-6, 2-7, and 2-8 electrode pairs with respect to electrode 2; and so on). The shape, size,
and conductivity of the inhomogeneity obtained using the NOSER, Laplace, and
Tikhonov algorithms were affected highly for a lower number of boundary voltages data.
On the other hand, the distributed conductivity, shape, and size of the sample obtained
 Remote Sens. 2022, 14, 3214 24 of 28

Table 2. Calculated mean and standard deviation (SD) of obtained changes of conductivity by varying
FEM models considering a carrot slice in the EIT domain.

FEM: a2c0 FEM: b2c0 FEM: d2d4c

(Elements: 64) (Elements: 256) (Elements: 4757)
Mean (S/m) 0.464563 0.253775 0.011742
SD (S/m) 0.37216 0.148718 0.024548

Variation of number of measurements in the frame: In addition, the number of mea-

surements was reduced from 64 to 40 from the dataset with the target of minimizing the
overlapping. In this case, the selected ports (3-4, 3-5, 3-6, 3-7, and 3-8 electrode pairs with
respect to electrode 3; 4-5, 4-6, 4-7, 4-8, and 4-1 electrode pairs with respect to electrode
4; and so on) are exactly the same as considered in the four-pole method. Finally, the
overlapping was reduced fully considering 28 measurements using the selected ports (1-2,
1-3, 1-4, 1-5, 1-6, 1-7, and 1-8 electrode pairs with respect to electrode 1; 2-3, 2-4, 2-5, 2-6,
2-7, and 2-8 electrode pairs with respect to electrode 2; and so on). The shape, size, and
conductivity of the inhomogeneity obtained using the NOSER, Laplace, and Tikhonov
algorithms were affected highly for a lower number of boundary voltages data. On the
other hand, the distributed conductivity, shape, and size of the sample obtained by the
corresponding reconstructed images were not affected highly using GHPF. A comparative
study was made between NOSER and GHPF considering a metal inhomogeneity in the
Remote Sens. 2022, 14, x FOR PEER REVIEW
EIT domain (shown in Figure 14a), and the obtained results for different measurements 25 at of 28
5 kHz and 1 mA are presented in Figure 24. Overall, GHPF performed well considering the
number of measurements of 64/40/28 in the frame compared to the other GN algorithm.

(a) (b) (c)

(d) (e) (f)

Figure 24.24.
Figure Reconstructed
Reconstructedimages
images of
of aa metal
metalinhomogeneity
inhomogeneity at at
thethe centre
centre of EIT
of EIT domain
domain by varying
by varying
the the
number
number of of
measurements
measurementsofof(a)
(a)64,
64,(b)
(b)40,
40, (c)
(c) 28
28 considering NOSER,and
considering NOSER, and(d)
(d)64,64,
(e)(e)
40,40,
(f)(f) 28
considering GHPP
28 considering withwith
GHPP λ = λ0.17 at at
= 0.17 5 kHz,
5 kHz,11mA,
mA, and
and 22 Vpp
Vppexcitation.
excitation.
TheThe maximum
maximum conductivity
conductivity
changes was
changes wasdecreased
decreasedby
bydecreasing the number
decreasing the numberofofmeasurements
measurements andand a distorted
a distorted reconstructed
reconstructed
image was found for lower number of measurements using NOSER. The
image was found for lower number of measurements using NOSER. The obtained maximum changes obtained maximum
changes of conductivity using GHPF are 2.78, 2.57, and 1.72 S/m, respectively, whereas
of conductivity using GHPF are 2.78, 2.57, and 1.72 S/m, respectively, whereas the changes are very the changes
are high
veryusing
high NOSER.
using NOSER.

The designed EIT sensor system is found suitable to reconstruct and differentiate the
conductivity levels of bio-, conductive-, and non-conductive targets with less attenuation.
Single or mixed inhomogeneities in the domain and the corresponding positions can be
identified successfully considering 2D or 3D EIT and the useful information can be
obtained by measuring impedances at different frequencies. A high-speed data
Remote Sens. 2022, 14, 3214 25 of 28

The designed EIT sensor system is found suitable to reconstruct and differentiate the
conductivity levels of bio-, conductive-, and non-conductive targets with less attenuation.
Single or mixed inhomogeneities in the domain and the corresponding positions can be
identified successfully considering 2D or 3D EIT and the useful information can be obtained
by measuring impedances at different frequencies. A high-speed data acquisition using
AD5933 is possible by the in situ measurements precisely up to 100 kHz. A successful
evaluation on anomaly detection is made using the designed EIT sensor system as shown
in Figure 14, which is very useful in the diagnosis of any abnormalities. The sensor system
is found repeatable, and less error is found in multiple measurements for a short duration.
The method is found robust and rapid in measurements to analyze the plant traits.
In this work, different samples of multiple plant species were taken for the investi-
gation using 2D and 3D EIT measurements. The samples of different edible plant slices
such as two carrots, one radish, and one potato species were taken for 2D operations. In
addition, a total of six root samples of different plant species such as four carrot roots (tap
roots), one potato root (tap root), and one Pilea root were taken for 3D operations. The root
systems of different plants were examined by measuring impedances using the designed
EIT sensor system in a controlled environment. The changes in dimensions of a tap root
(such as a carrot) were identified and the reconstructions were made successfully by the
calculated conductivity from the measured impedances as shown in Figure 16. The biomass
was varied with the variation of dimensions of the roots (such as carrot samples) and a
good correlation was found with the measured impedances at high frequencies as shown
in Figure 17. The root size was found highly correlated with the measured impedances of
the root samples.
The root distributions and density of the potato and Pilea roots were identified in
different layers of the electrode array as shown in Figures 19 and 20, respectively. The
root structures were reconstructed by calculating the conductivity distributed in different
layers of the electrode array. The variations in the root systems were able to be detected
with the help of 3D imaging. Finally, the biomass of a Pilea root was estimated using
EIS measurements with the help of AD5933 and a good correlation with R2 = 0.983 was
found considering eight samples in the dataset by the day-wise measurements as shown
in Figure 21. A model for root biomass estimation was predicted by selecting features
in different frequencies, as shown in Equation (15). It is evident that the tomographic
results can be utilized considering multi-electrodes measurements in different layers of the
array for monitoring root growth and biomass estimation of the roots in future. The root
growth can be monitored in hydroponics or soil media in further experiment. In addition,
more useful information of the roots can be obtained by increasing the number of layers
of electrodes.
Overall, the output excitation and frequency played an important role in reconstruction
analysis. The output of AD5933 was found stable at a 2 Vpp excitation and the samples
were found more conductive at the high frequencies. The selection of frequency and
the hyperparameter value depended on the type of inhomogeneity in the domain. The
conductivity was improved with a lower value of the hyperparameter. The designed
electrode array system has a good current carrying capability and is suitable for measuring
impedances in a large range of frequencies. The data measured using the sensor system
were found suitable for the estimation performance. The developed EIT data acquisition
system was tested by the experiments and found suitable for real-time high-precision
multifrequency measurements and monitoring in plant phenotyping.
The EIS tools such as the Agilent 4284A LCR meter in addition with Agilent34970A
digital multimeter for measuring boundary voltage/current data [1], and the QuadTech
7600 impedance analyzer for measuring bioimpedance [11] utilized in different experiments
are very expensive and heavier in weight. On the other hand, a portable, lightweight,
and low-cost EIS tool: EVAL-AD5933EBZ, is proposed in this work for developing an
automated, reliable, high-precision with good accuracy, and rapid EIT data acquisition
Remote Sens. 2022, 14, 3214 26 of 28

system applicable for plant phenotyping. A comparative study was made with the previous
research works on the development of the EIT data acquisition system, as shown in Table 3.

Table 3. A comparison of the features of proposed EIT data acquisition system with the previous
research works.

Features Zamora-Arellano et al. [23] Aris et al. [24] Singh et al. [10] Proposed EIT System
2D EIT 2D EIT 2D EIT 2D, and 3D EIT
Imaging system
(16 electrodes) (16 electrodes) (16 electrodes) (16 electrodes)
Source frequency 4–80 kHz 20 kHz 1–1000 kHz 1–100 kHz
Data acquisition and Arduino Uno, Raspberry Arduino Uno,
Arduino Mega, Raspberry Pi4 MCP3008, Raspberry Pi 2
CPU Pi3 PC
Sample rate (kSPS) 30 (24 bit ADC) 0.86 (16 bit ADC) - (10 bit ADC) 1000 (12 bit ADC)
Multiplexers ADG1406 (4) AD506AKNZ (4) CD4067BE (1) CD74HC4067 (2)
Software for image
EIDORS (MATLAB) Python EIDORS (MATLAB) EIDORS (MATLAB)
Reconstruction
Applications Health monitoring Anomaly detection Clinical imaging Plant root imaging

5. Conclusions
In this work, a multifrequency EIT data acquisition system is developed for applica-
tions in plant phenotyping with the target of evaluation and 2D/3D reconstruction of the
inhomogeneities by measuring impedances in a non-destructive manner. The developed
EIT system is portable, low-cost, and radiation-free. The reconstruction performance is
evaluated by several algorithms and a comparative study is made. The tomography results
of multiple plant roots using the EIT sensor system are obtained successfully in the 3D
plane considering the GN: NOSER algorithm. The designed electrode array sensor system
is found suitable for in situ measurements and the developed EIT system can be utilized in
the field scale for future study.

Author Contributions: R.B. performed the experiments and analyzed the data. R.B. and K.A.W. wrote
the draft of the manuscript. K.A.W. suggested for the experiments and data analysis. K.A.W. secured
funding for this work. All authors have read and agreed to the published version of the manuscript.
Funding: This work is supported by the Canada First Research Excellence Fund (CFREF) through
the Global Institute for Food Security (GIFS), University of Saskatchewan, Canada (422018).
Data Availability Statement: Data is contained within the article.
Acknowledgments: The authors would like to extend their appreciation to the Department of Elec-
trical and Computer Engineering, University of Saskatchewan, for the administrative and technical
support for the experiment.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Kim, B.S.; Khambampati, A.K.; Jang, Y.J.; Kim, K.Y.; Kim, S. Image reconstruction using voltage–current system in electrical
impedance tomography. Nuclear Eng. Des. 2014, 278, 134–140. [CrossRef]
2. Bera, T.K.; Nagaraju, J. A MATLAB-Based Boundary Data Simulator for Studying the Resistivity Reconstruction Using Neigh-
bouring Current Pattern. J. Med. Eng. 2013, 2013, 1–15. [CrossRef] [PubMed]
3. Wang, H.; Liu, K.; Wu, Y.; Wang, S.; Zhang, Z.; Li, F.; Yao, J. Image Reconstruction for Electrical Impedance Tomography
Using Radial Basis Function Neural Network Based on Hybrid Particle Swarm Optimization Algorithm. IEEE Sens. J. 2021, 21,
1926–1934. [CrossRef]
4. Russo, S.; Nefti-Meziani, S.; Carbonaro, N.; Tognetti, A. A Quantitative Evaluation of Drive Pattern Selection for Optimizing
EIT-Based Stretchable Sensors. Sensors 2017, 17, 1999. [CrossRef] [PubMed]
5. Loyola, B.R.; Saponara, V.L.; Loh, K.J.; Briggs, T.M.; O’Bryan, G.; Skinner, J.L. Spatial Sensing Using Electrical Impedance
Tomography. IEEE Sens. J. 2013, 13, 2357–2367. [CrossRef]
6. Yang, Y.; Jia, J.; Smith, S.; Jamil, N.; Gamal, W.; Bagnaninchi, P. A Miniature Electrical Impedance Tomography Sensor and 3D
Image Reconstruction for Cell Imaging. IEEE Sens. J. 2017, 17, 514–523. [CrossRef]
Remote Sens. 2022, 14, 3214 27 of 28

7. Malone, E.; dos Santos, G.S.; Holder, D.; Arridge, S. Multifrequency Electrical Impedance Tomography Using Spectral Constraints.
IEEE Trans. Med. Imaging 2014, 33, 340–350. [CrossRef]
8. Liu, S.; Huang, Y.; Wu, H.; Tan, C.; Jia, J. Efficient Multi-Task Structure-Aware Sparse Bayesian Learning for Frequency-Difference
Electrical Impedance Tomography. IEEE Trans. Ind. Inform. 2021, 17, 463–472. [CrossRef]
9. Malone, E.; dos Santos, G.S.; Holder, D.; Arridge, S. A Reconstruction-Classification Method for Multifrequency Electrical
Impedance Tomography. IEEE Trans. Med. Imaging 2015, 34, 1486–1497. [CrossRef]
10. Singh, G.; Anand, S.; Lall, B.; Srivastava, A.; Singh, V. A Low-Cost Portable Wireless Multi-frequency Electrical Impedance
Tomography System. Arab. J. Sci. Eng. 2019, 44, 2305–2320. [CrossRef]
11. Bera, T.K.; Nagaraju, J.; Lubineau, G. Electrical impedance spectroscopy (EIS)-based evaluation of biological tissue phantoms to
study multifrequency electrical impedance tomography (Mf-EIT) systems. J. Vis. 2016, 19, 691–713. [CrossRef]
12. Shi, X.; Li, W.; You, F.; Huo, X.; Xu, C.; Ji, Z.; Liu, R.; Liu, B.; Li, Y.; Fu, F.; et al. High-Precision Electrical Impedance Tomography
Data Acquisition System for Brain Imaging. IEEE Sens. J. 2018, 18, 5974–5984. [CrossRef]
13. Sapuan, I.; Yasin, M.; Ain, K.; Apsari, R. Anomaly Detection Using Electric Impedance Tomography Based on Real and Imaginary
Images. Sensors 2020, 20, 1907. [CrossRef] [PubMed]
14. Bai, X.; Liu, D.; Wei, J.; Bai, X.; Sun, S.; Tian, W. Simultaneous Imaging of Bio- and Non-Conductive Targets by Combining
Frequency and Time Difference Imaging Methods in Electrical Impedance Tomography. Biosensors 2021, 11, 176. [CrossRef]
15. Weigand, M.; Kemna, A. Multi-frequency electrical impedance tomography as a non-invasive tool to characterize and monitor
crop root systems. Biogeosciences 2017, 14, 921–939. [CrossRef]
16. Weigand, M.; Kemna, A. Imaging and functional characterization of crop root systems using spectroscopic electrical impedance
measurements. Plant Soil 2019, 435, 201–224. [CrossRef]
17. Postic, F.; Doussan, C. Benchmarking electrical methods for rapid estimation of root biomass. Plant Methods 2016, 12, 33. [CrossRef]
18. Corona-Lopez, D.D.J.; Sommer, S.; Rolfe, S.A.; Podd, F.; Grieve, B.D. Electrical impedance tomography as a tool for phenotyping
plant roots. Plant Methods 2019, 15, 49. [CrossRef]
19. Liu, S.; Jia, J.; Zhang, Y.D.; Yang, Y. Image Reconstruction in Electrical Impedance Tomography Based on Structure-Aware Sparse
Bayesian Learning. IEEE Trans. Med. Imaging 2018, 37, 2090–2102. [CrossRef]
20. Liu, D.; Khambampati, A.K.; Du, J. A Parametric Level Set Method for Electrical Impedance Tomography. IEEE Trans. Med.
Imaging 2018, 37, 451–460. [CrossRef]
21. Yang, Y.; Jia, J. An Image Reconstruction Algorithm for Electrical Impedance Tomography Using Adaptive Group Sparsity
Constraint. IEEE Trans. Inst. Meas. 2017, 66, 2295–2305. [CrossRef]
22. Ren, S.; Wang, Y.; Liang, G.; Dong, F. A Robust Inclusion Boundary Reconstructor for Electrical Impedance Tomography with
Geometric Constraints. IEEE Trans. Inst. Meas. 2019, 68, 762–773. [CrossRef]
23. Zamora-Arellano, F.; López-Bonilla, O.R.; García-Guerrero, E.E.; Olguín-Tiznado, J.E.; Inzunza-González, E.; López-Mancilla,
D.; Tlelo-Cuautle, E. Development of a Portable, Reliable and Low-Cost Electrical Impedance Tomography System Using an
Embedded System. Electronics 2021, 10, 15. [CrossRef]
24. Aris, W.; Endarko. Design of low-cost and high-speed portable two-dimensional electrical impedance tomography (EIT). Int. J.
Eng. Technol. 2018, 7, 6458–6463.
25. Ozier-Lafontaine, H.; Bajazet, T. Analysis of root growth by impedance spectroscopy (EIS). Plant Soil 2005, 277, 299–313. [CrossRef]
26. Liao, A.; Zhou, Q.; Zhang, Y. Application of 3D electrical capacitance tomography in probing anomalous blocks in water. J. Appl.
Geophys. 2015, 117, 91–103. [CrossRef]
27. Newill, P.; Karadaglic, D.; Podd, F.; Grieve, B.D.; York, T.A. Electrical impedance imaging of water distribution in the root zone.
Meas. Sci. Technol. 2014, 25, 055110. [CrossRef]
28. Tan, C.; Liu, S.; Jia, J.; Dong, F. A Wideband Electrical Impedance Tomography System based on Sensitive Bioimpedance Spectrum
Bandwidth. IEEE Trans. Inst. Meas. 2020, 69, 144–154. [CrossRef]
29. Chowdhury, R.I.; Basak, R.; Wahid, K.A.; Nugent, K.; Baulch, H. A Rapid Approach to Measure Extracted Chlorophyll-a from
Lettuce Leaves using Electrical Impedance Spectroscopy. Water Air Soil Pollut. 2021, 232, 73. [CrossRef]
30. Graham, B.M.; Adler, A. Electrode placement configurations for 3D EIT. Physiol. Meas. 2007, 28, 29–44. [CrossRef]
31. Basak, R.; Wahid, K.; Dinh, A. Determination of Leaf Nitrogen Concentrations Using Electrical Impedance Spectroscopy in
Multiple Crops. Remote Sens. 2020, 12, 566. [CrossRef]
32. Matsiev, L. Improving Performance and Versatility of Systems Based on Single-Frequency DFT Detectors Such as AD5933.
Electronics 2015, 4, 1–34. [CrossRef]
33. Basak, R.; Wahid, K.A.; Dinh, A. Estimation of the Chlorophyll-A Concentration of Algae Species Using Electrical Impedance
Spectroscopy. Water 2021, 13, 1223. [CrossRef]
34. Putensen, C.; Hentze, B.; Muenster, S.; Muders, T. Electrical Impedance Tomography for Cardio-Pulmonary Monitoring. J. Clin.
Med. 2019, 8, 1176. [CrossRef]
35. Rymarczyk, T.; Kłosowski, G.; Kozłowski, E.; Tchórzewski, P. Comparison of Selected Machine Learning Algorithms for Industrial
Electrical Tomography. Sensors 2019, 19, 1521. [CrossRef]
Remote Sens. 2022, 14, 3214 28 of 28

36. Fernández-Fuentes, X.; Mera, D.; Gómez, A.; Vidal-Franco, I. Towards a Fast and Accurate EIT Inverse Problem Solver: A Machine
Learning Approach. Electronics 2018, 7, 422. [CrossRef]
37. Kłosowski, G.; Rymarczyk, T.; Niderla, K.; Rzemieniak, M.; Dmowski, A.; Maj, M. Comparison of Machine Learning Methods for
Image Reconstruction Using the LSTM Classifier in Industrial Electrical Tomography. Energies 2021, 14, 7269. [CrossRef]