See discussions, stats, and author profiles for this publication at: [Link]

net/publication/383623130

Supervised classification and circuit parameter analysis of electrical

bioimpedance spectroscopy data of water stress in tomato plants

Article in Computers and Electronics in Agriculture · August 2024

DOI: 10.1016/[Link].2024.109347

CITATION READS

1 78

5 authors, including:

Saleh Hamed Antonio Altana

10 PUBLICATIONS 62 CITATIONS
Free University of Bozen-Bolzano
35 PUBLICATIONS 47 CITATIONS
SEE PROFILE
SEE PROFILE

Paolo Lugli Luisa Petti

Free University of Bozen-Bolzano Free University of Bozen-Bolzano
879 PUBLICATIONS 19,451 CITATIONS 199 PUBLICATIONS 4,101 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Saleh Hamed on 04 September 2024.

The user has requested enhancement of the downloaded file.

 Computers and Electronics in Agriculture 226 (2024) 109347

Contents lists available at ScienceDirect

Computers and Electronics in Agriculture

journal homepage: [Link]/locate/compag

Original papers

Supervised classification and circuit parameter analysis of electrical

bioimpedance spectroscopy data of water stress in tomato plants
Saleh Hamed a ,∗, Antonio Altana a,b , Paolo Lugli a,b , Luisa Petti a,c , Pietro Ibba a ,∗
a
Faculty of Engineering, Free University of Bozen-Bolzano, Piazza Domenicani 3, Bozen-Bolzano, 39100, Italy
b
Competence Centre for Mountain Innovation Ecosystems, Free University of Bozen-Bolzano, Piazza Universitá 1, Bozen-Bolzano, 39100, Italy
c
Competence Centre for Plant Health, Free University of Bozen-Bolzano, Piazza Universitá 1, Bozen-Bolzano, 39100, Italy

ARTICLE INFO ABSTRACT

Keywords: Drought poses a significant challenge by inducing water stress in crops, urgently calling for effective monitoring
Bioimpedance and early intervention. In this work, we conducted a comprehensive 38-day investigation of the bioimpedance
Water stress of water stressed tomato stems under controlled environmental conditions. A total of 8000 measurements
Machine learning
were meticulously collected across a frequency spectrum spanning from 100 Hz to 10 MHz. The collected
Equivalent circuit models
data was categorized into control, early stress, and late stress groups, corresponding to distinct phases of
Early detection
Solanum lycopersicum
water stress treatment. To analyze the data, we employed established equivalent circuit models, including
Cole, Randles, and double Cole. Subsequently, we evaluated the performance of eight machine learning
algorithms widely-used in predicting the water stress stages of the plants. Remarkably, the Cole model, in
conjunction with a multi-layer perceptron (MLP) algorithm, demonstrated robust performance, achieving an
impressive F1 score of 0.89. In fact, in the control group, 378 out of 388 instances were accurately identified.
However, misclassifications were observed in the early and late stress groups, with 31 out of 179 and 14
out of 208 instances mislabeled, respectively. Furthermore, we scrutinized the Cole model’s circuit parameters
over time, providing insights aligned with the plant physiological behavior documented in the literature. Our
findings suggest that MLP models trained on stem bioimpedance data hold promise as a valuable technique
for estimating water stress in tomato plants, offering an essential tool for proactive management by farmers.

1. Introduction Drought’s impact on tomato plants involves reduced height, fewer

leaves, decreased leaf area, withering of stems and leaves, lowered
Plants are sessile organisms that react to various stress factors fresh shoot weight, and significant decreases in both stomatal and pore
influenced by environmental elements like drought, extreme tempera- lengths (Zhou et al., 2017). Given these factors, precise characterization
tures, and pollution (Eckardt et al., 2023). When facing water scarcity,
and monitoring of water stress in tomato plants are crucial. Of especial
primarily due to inadequate rainfall or insufficient soil moisture, plants
importance are in particular early detection methods, allowing timely
initiate stress reactions that extend from individual cells and organs
intervention, allowing a reduced resource use and a safeguarding of
to the entire plant structure. These reactions significantly impede crop
growth (Francini and Sebastiani, 2019). In this context, one of the overall yields.
plants most affected by drought is tomato (Solanum lycopersicum), an In this context, Agriculture 4.0 introduces new approaches, utilizing
especially important crop that yields a global production exceeding sophisticated technological tools for optimal operation (Araújo et al.,
170 million tons. This is expected to worse with climate change- 2021). Within this context, proximal sensors play a crucial role, with
induced drought, which is expected to reduce the global processing various techniques have been outlined in the existing literature (Presti
yields of tomato by 6% by 2050 (Cammarano et al., 2022). Beyond et al., 2022).
its agricultural importance, the tomato plant is a favored subject for Among the various state-of-the-art methods for proximal plant mon-
scientific research due to its manyfold attractive features. These in- itoring, it is worth mentioning those that focus on characterizing the
clude a fully sequenced genome, a brief life cycle that facilitates easy
plant’s electrical properties, utilizing transducing platforms based on
cultivation, horticultural manipulability, the availability of numerous
e.g., field-effect transistors (Elli et al., 2022; Hamed et al., 2021),
mutant lines and genomics tools, as well as of comprehensive exper-
resistance-based (Ehosioke et al., 2020), electrophysiological (Baluška
iment guidelines (Costa and Heuvelink, 2018; Schwarz et al., 2014).

∗ Corresponding authors.
E-mail addresses: [Link]@[Link] (S. Hamed), [Link]@[Link] (P. Ibba).

[Link]
Received 12 January 2024; Received in revised form 15 June 2024; Accepted 9 August 2024
Available online 27 August 2024
0168-1699/© 2024 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY license ([Link]
S. Hamed et al. Computers and Electronics in Agriculture 226 (2024) 109347

et al., 2021; Najdenovska et al., 2021), and bioimpedance-based (Liu 2. Materials and methods
et al., 2021) methods. These approaches offer easy, cost-effective, and
proximal plant monitoring techniques, able to provide valuable insights 2.1. Plant materials
into plant well-being. Due to their cost-effectiveness, these electrical-
based methods can be seamlessly integrated with traditional plant mon- Eight seedlings of Solanum lycopersicum cv. Pomodoro Tondo were
itoring approaches, including optical methods (e.g., Red, Green, and cultivated in an inert Perlite substrate (Karl Bachl Kunststoffverar-
Blue RGB, hyperspectral imaging, or Unmanned Aerial Vehicle (UAV) beitung GmbH & Co. KG), chosen to control the plant nutrient supply.
The plants were grown in pots with a diameter of 21 cm and a volume
and spectroscopic methods (e.g., infrared and Raman spectroscopy).
of 4 liters for a period of 7–8 weeks within a controlled environmental
This integration allows for a more thorough and well-rounded col-
conditions. The greenhouse conditions were set with a 12-h day (26 ◦ C,
lection of data in agricultural settings, as highlighted in Roper et al.
55% relative humidity), 328 μmol photons m−2 s−1 light intensity and
(2021).
a 12-h night (20 ◦ C, 88% relative humidity). To meet their nutritional
Within the available electrical methods, bioimpedance provides an requirements, the plants were watered daily with tap water, and a
extremely interesting proximal and non-invasive method for analyzing Hoagland solution (Hoagland et al., 1950), containing essential macro
biological materials, including plant tissue. When an alternating current and micro-nutrients, was provided daily. The Hoagland solution was
(AC) is applied, the flow of electrical current varies with the frequency prepared according to Maver et al. (2022).
of the AC stimulus (Liu et al., 2021). At low frequencies, the current The eight tomato plants were subjected to two experimental con-
predominantly moves through the extracellular fluid, while at high ditions. The control group comprised four plants that received a daily
frequencies, it mainly travels through the cells themselves (Liu et al., supply of the complete Hoagland nutrient solution throughout the ex-
2021). This frequency-dependent current can be utilized as indicators periment, in addition to two watering events per day. The
for plant health and for plant response to environmental conditions. bioimpedance measurements began at the late vegetative growth stage.
In fact, stress in a plant induces noticeable changes in the electrical Before this point, the plants were grown in controlled and optimal
conditions until an even growth level was achieved, to minimize
resistance of its stem (Kuromori et al., 2022), with changes resulting
possible influences of different maturity levels on the measurements.
from a combination of molecular and hormonal alterations within the
The initial phase of the bioimpedance experiment, from Day 0 to Day
plant and subsequently leading to variations in protein levels (Ogden
6, maintained optimal water and nutrient levels. During this time,
et al., 2020), ion concentrations (Janni et al., 2019), and physical
the plants transitioned through the flowering stage. The subsequent
attributes such as stem size and sap flow (Vermeulen et al., 2007). phase, from Day 7 to Day 30, induced water stress by intentionally
Indeed, numerous studies have shown that bioimpedance in plant stems withholding water, spanning the early fruiting stage and continuing
is effective for detecting water stress (Bar-On and Shacham-Diamand, through the development and maturation of fruit. This phase included
2021; Garlando et al., 2022). Additionally, recent research has utilized early stress from Day 7 to Day 18 and late stress from Day 19 to
bioimpedance monitoring to evaluate plant water stress, revealing that Day 30. The final phase, from Day 31 to Day 39, involved a recovery
extracellular resistance can increase by 25% within one hour under period where water was reintroduced to the plants, corresponding
simulated drought conditions (Reynolds et al., 2023). Recent related to the ripening and late fruiting stage, until the conclusion of the
research has focused on using plant stem bioimpedance to detect water experiment. While it is indeed possible that the different stages of the
stress in plants, exploring cost-effective approaches (Calvo et al., 2023) plant contributed in a different response to water stress, a clear impact
and integrating machine learning. Neural networks were employed was not observed.
in Barezzi et al. (2022), while another study (Cum et al., 2023) tested
various machine learning algorithms, finding that the random forest 2.2. Bioimpedance setup
classifier excelled with a 98.48% accuracy rate in assessing tobacco
Fig. 1 provides an illustration of the experimental setup utilized in
plant health status. Even though bioimpedance has already been used
this study. The bioimpedance data of the tomato plants were collected
for plant monitoring (Hamed et al., 2023), there is still limited ex-
using a Digilent Analog Discovery 2 semi-portable impedance analyzer
ploration into its consistent evaluation under controlled conditions,
(Digilent Inc.), covering a frequency range from 100 Hz to 10 MHz
particularly those referring to water deficiency that induces drought with 200 logarithmically spaced frequency points. The measurements
stress. were conducted in a two-electrode setup, employing stainless steel
This study aims at thoughtfully examining the influence of wa- subcutaneous needle electrodes (13 mm in length, 0.4 mm in diameter,
ter stress on tomato plants by continuously monitoring bioimpedance 27 g, Technomed Medical Accessories). These electrodes were inserted
changes in their stems for over a month in a controlled environment at a 90-degree angle with the stem, positioned in parallel, one above the
under 3 physiological plant status (Control, Early stress and Late stress). other, into the plant stem. The insertion point was 5 cm from the sub-
We chose to use two stress labels because plants react differently to strate, with a fixed separation of 0.5 cm between the needles (Fig. 2a).
mild and extreme deficiencies. This arbitrary naming helps us more To streamline the process, all eight plants were simultaneously con-
precisely characterize stress through bioimpedance data, distinguishing nected to the impedance analyzer using a custom-made multiplexing
between initial resistance and responses to prolonged stress. Using setup. The switching between channels, and thus plants, was coordi-
nated with data acquisition through. an Arduino Uno microcontroller
a semi-portable bioimpedance analyzer, bioimpedance data were col-
that is managed by a Python script to oversee various operations
lected across frequencies ranging from 100 Hz to 10 MHz. Initially,
within our measurement system. This setup incorporates libraries such
changes in bioimpedance magnitude were studied over time. Subse-
as pyfirmata for communication with the Arduino and numpy for
quently, these changes were used to derive circuit parameters by fitting
data management, with measurements taken hourly. The Python script
various equivalent models, namely the Cole model (Cole, 1940), the is responsible for selecting the active channel in the bi-directional
Randles model (Randles, 1947), and the double Cole model (Zhang and Multiplexer CD74HC4067, setting the selection inputs (𝑆0 -𝑆1 -𝑆2 -𝑆3 )
Willison, 1991). The performance of these models was then compared connected to the Arduino’s digital outputs. This system uses the Stan-
using eight different classification machine learning (ML) algorithms on [Link] firmware to facilitate communication through
a test dataset. The study concluded by evaluating the effectiveness of the Firmata protocol, enabling control via the pyFirmata interface.
the best-performing model selected through these analyses with (MLP) Additionally, the terminals of the portable impedance analyzer, Analog
classification model performing accuracy F1 score of 0.89. Discovery 2 (equipped with an impedance analyzer add-on), are linked

2
S. Hamed et al. Computers and Electronics in Agriculture 226 (2024) 109347

Fig. 1. Schematic of the bioimpedance monitoring setup for tomato plants under water stress. The system includes a Multiplexer (MUX) for connecting multiple electrodes to the
Arduino for data collection, and an Analog Discovery unit functioning as an impedance analyzer. A PC is integrated for logging and processing the data.

Fig. 2. (a) Illustration of the needle electrode insertion site (NEIS) into the tomato stem, with a microscopic cross section of the stem (b), highlighting the NEIS, alongside various
anatomical layers: epidermis (E), cortex (C), phloem (PH), xylem (X), and pith (PI).

to the multiplexer’s signal outputs. The synchronization of channel then immersed in a Safranin solution for 45 min, followed by a sequen-
selection with bioimpedance measurements is governed by a shared tial rinsing process with distilled water, 95% ethanol, and 99% ethanol,
flag file, which signals whether a measurement is in progress or com- each step being repeated 3–4 times. Subsequently, the stained sections
pleted. This control ensures that channel switching occurs only after were examined using a compound microscope. This staining procedure
a measurement is finished, allowing for precise and systematic data allowed for a detailed examination and identification of different stem
collection. Continuous measurements were taken every hour, extending tissues, contributing to a more comprehensive understanding of the
over a total duration of 38 days. This setup allowed for a detailed and bioimpedance measurements and their correlation with specific plant
extended monitoring of the bioimpedance changes in the tomato plants structures.
under various water stress conditions.
2.4. Data analysis

2.3. Tissue analysis

The data was analyzed using Matlab R2022b (The MathWorks Inc.,
Natick, MA, United States) and its Zfit function that is designed to
To precisely identify the specific layers of stem tissues that the handle impedance data by plotting, simulating, and fitting. It visualizes
needle electrodes came into contact with, including the epidermis, the data for easy analysis, generates data based on models for experimental
cortex, the phloem, the xylem, and the pith (as shown in Fig. 2), a purposes, and applies curve-fitting techniques to determine the best
staining procedure involving Safranin was implemented, following the model parameters. Dellis (2023). We focused on analyzing only the
methodology outlined in Von Arx et al. (2016). This staining process data collected during the periods when the plants were actually under
facilitated the differentiation of various tissues for subsequent observa- stress. This selective analysis is important, not only to differentiate
tion under a microscope. In essence, stem cross-sections were obtained between the various stages of stress, such as early and late stress,
from the needle electrode insertion site (NEIS). These sections were but also because the period prior to the stress involved the plants

3
S. Hamed et al. Computers and Electronics in Agriculture 226 (2024) 109347

stabilizing in their new environment and adjusting to the effects of phase angle trends, a frequency-dependent shift was evident in all
electrode insertion and polarization. By concentrating on the data conditions. For instance, the late stress condition exhibited the smallest
from the stress periods, the study aims to more accurately assess phase angle up to frequencies less than 105 Hz, followed by a change
how plants respond to stress at different times, avoiding confounding in direction at higher frequencies. This phase angle behavior typically
factors introduced immediately after experimental setup. To simplify indicates properties such as cell membrane capacitance and the balance
the dataset and make results easier to understand, the bioimpedance of water inside and outside cells (Ward and Brantlov, 2023).
data was modeled using three types of circuits: Cole, Randles, and
Based on our review and the trends shown in Fig. 3a and b, no
double Cole, considering both real and imaginary components of the
clear patterns were observed among the three conditions examined. The
bioimpedance. To streamline computational complexity, a subset of
randomly chosen data was employed to identify an optimal starting figures show overlapping standard deviations in the average spectra,
point for circuit fitting. The initial conditions resulting in the lowest indicating that simply analyzing average bioimpedance data is not
root mean square error (RMSE) were then utilized for fitting the entire effective in distinguishing between stressed and healthy plants. This
dataset. This subset, constituting 1/6 of the data for each condition, was issue is typical with bioimpedance data, which, although rich in in-
tested on the three circuits during the 20:80 test-train bootstrap phase. formation, does not always allow for easy interpretation. To manage
The balanced subsets were essential, ensuring an equal distribution this complexity and make the most of the extensive spectral data, we
for each class. These subsets were employed to train various machine applied three standard equivalent circuits to the raw bioimpedance
learning classification models, including discriminant analysis (DA), data. This method helps simplify the data while preserving most of the
kernel methods (KM), k-Nearest neighbors (KNN), linear regression information. The parameters derived from these circuits can be used
models (LRM), naive Bayes classifier (NBC), multi-layer perceptron to link the electrical characteristics of the plants to their responses to
(MLP), support vector machines (SVM), and decision trees (DT). It is water stress (Grimnes and Martinsen, 2006).
worth highlighting that only the period when the plant was under stress
was considered for classification. This approach was chosen to focus
on the most relevant period for analysis and to prevent class skewing. 3.1. Equivalent circuit models
Subsequently, the trained models’ performance was evaluated using
unseen test datasets. Finally, the Cole model, because of its simplicity,
was employed, and its components (Ibba et al., 2019) such as: To simplify the complexity of the data. Model fitting results were
initially assessed using RMSE, comparing the predicted data with the
𝑅𝑠 (Series Resistance) - that represents the resistance of the extracellu-
measured data. This evaluation served as a reliable indicator of fitting
lar medium, which includes the spaces outside of the plant cells, such
as the intercellular fluids within leaves, stems, and roots. quality to assess the performance of the considered models. Supple-
mentary Fig S1 and Table S1 presents a comparison of the obtained
𝑅𝑝 (Parallel Resistance) - Reflects the resistance of the intracellular RMSE values, alongside the results of a one-way ANOVA conducted to
medium, capturing the resistance offered by the fluids and components
evaluate statistical differences in the average error of each circuit.
inside the plant cells themselves.
On average, all three tested circuits demonstrated effective per-
𝐶𝑃 𝐸 − 𝑇 (Constant Phase Element - T) - Represents the pseudo- formance in the fitting task, yielding RMSE values of approximately
capacitance of the cell membrane. This pseudo-capacitance varies de- 331 ± 262 Ω, 433 ± 268 Ω, and 431 ± 261 Ω for the Randles, Cole, and
pending on factors like membrane integrity and ionic content, which double Cole models, respectively. Notably, the Randles model exhibited
influence how the membrane stores and releases electrical charge.
a significantly lower RMSE compared to the other two models. This
𝐶𝑃 𝐸 − 𝑃 (Constant Phase Element - P) - Indicates the heterogeneity in distinction is likely attributed to the inclusion of a Warburg element
electrical properties caused by variations in the size and shape of cells in the Randles model, a component associated with the mass diffusion
within the plant. This parameter adjusts the phase angle to account process at the interface between an electrode and an electrolyte (War-
for these irregularities, providing a more nuanced view of the plant’s burg, 1899). This characteristic aligns well with the interface between
structural complexity. These components were extracted at a frequency the needle electrodes and the plant’s sap rich in ions, as evidenced by
of 10 kHz and analyzed to understand how they changed over time, its effectiveness in representing biological tissues in apple and banana
providing insights into the dynamic response of the tomato plants under
fruits (Ibba et al., 2019, 2021b, 2020, 2018, 2021a; Ibba, 2021; Ibba
varying water stress conditions.
et al., 2022).
3. Results and discussion Despite the minor differences in circuit components and a compa-
rable level of complexity among the three options, selecting a model
An overview of the impact of water stress on the bioimpedance based solely on the RMSE metric proves challenging. Consequently,
magnitude and phase angle output of 8 tomato plants is presented in the extracted circuit parameters undergo additional testing for their
Fig. 3a and b, respectively, showcasing the average of 8000 individual efficacy in a more intricate task: discriminating between stressed and
measurements. The presented data focus on the time-frame considered healthy plants. This evaluation aims indeed at assessing the utility
for classification, specifically for the 4 control plants (depicted by the of bioimpedance for the early detection of water stress in plants and
blue line in the plot) and the 4 stressed plants under both early and at determining if the use of fitted circuit parameters as discriminant
late water stress conditions (depicted by yellow and red lines in the features is a suitable approach for training classification models.
plot, respectively). The data revealed that at lower frequencies, where
We tested the three considered models both in terms of RMSE
the bioimpedance is predominantly influenced by the current flowing
of the fitting and in terms of performance of the extracted circuit
through the fluid around the cells, the magnitude is higher. Notably,
parameters in the discrimination between stressed and healthy plants.
in the bioimpedance magnitude data for the three conditions (control,
early stress, and late stress), differences were most pronounced at This allowed us to select, by the end of the paper, the Cole model as
low frequencies below 105 Hz. Lower magnitudes were consistently a best performing equivalent circuit for the plant stress discrimination,
observed in the early stress condition compared to the other trends. through a combined evaluation of each model RMSE in fitting, simplic-
Conversely, higher magnitudes were noted in the late stress condition, ity (i.e., number of components) and performance in the classification
particularly up to frequencies less than 104 Hz, at which point the task. The study thus will focus exclusively on this model to explore its
control condition surpassed the others in magnitude. Concerning the parameters’ correlation with plant water stress responses.

4
S. Hamed et al. Computers and Electronics in Agriculture 226 (2024) 109347

Fig. 3. Summary of bioimpedance data across a frequency range of 100 Hz to 1 MHz (average of 4 plants for each curve) represented through Bode plots: Bioimpedance magnitude
(a) and phase angle (b). Control plants are represented by the blue curve, early water stress by the yellow curve, and late water stress by the red curve, with the shaded area
representing the standard deviation and the line representing the average.

3.2. Water stress classification models good discrimination power and generalization capabilities across the
three feature sets, showcasing its strong performance in predicting
To gain a deeper understanding of how the parameters extracted labels for unseen data. The utility of the KNN model in classifying
from these circuit models perform in a complex discrimination task, bioimpedance measurements has been demonstrated in other stud-
various supervised machine learning classification algorithms were ies (Tiitta et al., 2020), accurately classifying the heartwood content
trained, validated, and tested. The objective was to discriminate among of pine chips and bark content of birch chips with a precision of 0.91.
three distinct classes representing the control, early stress, and late In contrast, DTs exhibit a considerable difference of 9.4%, 45.11%,
stress conditions of the tomato plants. In the initial training phase, each and 33.6% between their train and test F1 scores, indicating overfitting
algorithm’s hyperparameters were automatically optimized to find the on the training data. Specifically for the Randles and double Cole circuit
best combination for achieving optimal training performance. models, the KNN model outperforms other models with a test F1 score
Subsequently, a 1000-round bootstrapping validation phase, cou- of 0.72 and 0.71, respectively. Conversely, the MLP model presents
pled with a 10 fold cross-validation, was conducted. This phase aimed a significant disparity (23.3% and 51.82%, respectively) between its
to assess the consistency in the models’ performance when slightly train and test F1 scores, likely due to overfitting. Regarding the specific
changing the training dataset and to identify the model exhibiting case of the Cole model’s parameters, while most tested algorithms
superior performance across diverse circuit configurations. Supplemen- showed good quality, the MLP demonstrates the best test performance,
tary Fig. S2 illustrates the results of the bootstrap validation in terms achieving an F1 score of 0.89. The utility of the MLP model has
of average F-score. From the histograms, it is evident that the Cole been observed in other studies using bioimpedance data for strawberry
model excels in DT, DS, KM, KNN, and MLP methods, showcasing ripeness classification, where MLP achieved a performance score of
consistent performance across different resamples. This consistency
0.77 (Ibba et al., 2021b).
indicates the Cole model’s reliability when tested on unseen data in
Fig. 4 displays the confusion matrix of the better performing model
the test set. Conversely, the Cole model demonstrates comparatively
on the test set, i.e., the MLP model, yielding from the use of the Cole
lower performance in the NBC, while displaying similar performance
model’s features, for both training and testing.
to the double Cole model in LRM and SVM, where the double Cole
It is clear that for both the training and test datasets, several key
model excels. In contrast, the Randles model consistently exhibits lower
observations emerge, elucidating the model’s performance character-
performance across all models, except in the case of NBC, where it
istics. First and foremost, the model excels in classifying the control
positions itself between the Cole model and the double Cole model in
samples in both sets. In the training set, 1549 out of 1555 control
terms of computed F-scores. In summary, it is evident that the Cole
samples were correctly classified, yielding a high true positive rate.
circuit model consistently outperforms other models across a range
This performance is mirrored in the test set, where 378 out of 388
of machine learning techniques, while the Randles model consistently
demonstrates lower performance levels, highlighting the reliability and control instances are correctly identified. The consistently low number
robustness of the Cole circuit model. of false positives and negatives in both sets emphasizes the model’s
These trends are further confirmed in the third phase, where a robustness in identifying control conditions. However, discrepancies
comparison of the results of both the three models and the eight begin to appear when examining the early Stress and late stress classes.
algorithms was performed, evaluating accuracy (i.e., F1-score) in train In the training set, the model misclassified early stress as late stress 23
and test, and stability between the two phases to assess the model’s gen- times out of 715 instances, whereas in the test set, this error occurred
eralization capabilities. Table 1 presents an overview of the obtained 31 times out of 179 instances. This suggests an expected decline in
results. performance when the model encounters new data. The late stress class
Overall, the Cole model consistently outperforms the Randles and revealed similar trends. In the training set, 30 out of 836 late stress
double Cole models, demonstrating good performance on both the samples were incorrectly labeled as early stress, compared to 14 out of
training and test datasets. The Randles model exhibits the poorest 208 in the test set.
overall performance, with subpar results on both the training and test Overall, while the model demonstrated good proficiency in clas-
datasets. On the other hand, the double Cole model performs well on sifying the control class, its performance decreased in distinguishing
the training dataset but shows poor performance on the test dataset, between early stress and late stress. The uptick in misclassifications
indicating limited generalization to unseen data. from the training set to the test set, particularly for the stress classes,
Taking a closer look at specific supervised machine learning classi- implies a lack of generalization that should be addressed in future
fication models, it is observed that KNN maintains the most stable per- iterations of the model. Addressing this generalization challenge is
formance across the training and test sets. It consistently demonstrates crucial for enhancing the model’s performance in real-world scenarios

5
S. Hamed et al. Computers and Electronics in Agriculture 226 (2024) 109347

Table 1
Comparison of Different Statistical Models Across Cole, Randles, and Double Cole models: This table provides a detailed comparison of several ML models such as Decision Tree
(DT), Support Vector Machine (SVM), and Multi-Layer Perceptron (MLP), among others. The models are evaluated based on optimized hyperparameters and their performance
metrics including precision (P), recall (R), and F1 score across training and testing datasets. Differences in F1 scores between training and testing phases are also highlighted to
assess model generalizability.
Cole model
Model Optimized hyperparameters P train P test R train Recall test F1 train F1 test F1% diff
DA 𝛥 = 7.2787𝑒−6 , 𝛾 = 7.1992𝑒4 0.65 0.67 0.63 0.64 0.64 0.66 2.10
KM KS = 0.647, 𝜆 = 4.1848𝑒−04 0.88 0.87 0.87 0.86 0.88 0.87 −1.05
KNN Distance = Mahalanobis, NN = 8 0.90 0.87 0.90 0.86 0.90 0.87 −3.26
LM 𝜆 = 4.9024𝑒−6 , Learner = Logistic 0.70 0.71 0.67 0.68 0.69 0.69 0.58
NBC Distribution names = Kernel, Width = 0.0102 0.62 0.61 0.63 0.61 0.62 0.61 −1.00
MLP Activation = Tanh, Layer size = [74,27,2] 0.97 0.89 0.97 0.89 0.97 0.89 −7.81
SVM Function = Linear, Scale = 0.9970 0.71 0.71 0.68 0.68 0.70 0.70 0.13
DT Number of nodes = 245 0.91 0.82 0.91 0.81 0.91 0.82 −9.40
Randles model
Model Optimized hyperparameters P train P test R train Recall test F1 train F1 test F1% diff
DA 𝛥 = 0.0966, 𝛾 = 0.0182 0.55 0.54 0.55 0.54 0.55 0.54 −1.47
KM KS = 0.6587, 𝜆 = 3.7381𝑒−07 0.78 0.61 0.79 0.60 0.79 0.58 −21.09
KNN Distance = cityblock, NN = 9 0.76 0.71 0.75 0.72 0.76 0.72 −4.21
LM 𝜆 = 2.0612𝑒−6 , Learner = SVM 0.77 n.a. 0.57 n.a. 0.49 n.a. n.a.
NBC Distribution names = Kernel, Width = 0.0060 0.74 0.60 0.74 0.61 0.74 0.61 −12.95
MLP Activation = tanH, Layer size = [28,10] 0.82 0.62 0.81 0.61 0.82 0.58 −23.33
SVM KS = 0.4792, Binary loss = Hinge 0.77 n.a. 0.56 n.a. 0.49 n.a. n.a.
DT Number of nodes = 165 0.86 0.52 0.85 0.46 0.86 0.41 −45.11
Double Cole model
Model Optimized hyperparameters P train P test R train Recall test F1 train F1 test F1% diff
DA 𝛥 = 2.0688𝑒−6 , 𝛾 = 0.0568 0.62 0.55 0.59 0.54 0.60 0.54 −5.68
KM KS = 1.0196, 𝜆 = 2.1249𝑒−5 0.86 0.57 0.86 0.60 0.86 0.57 −29.56
KNN Distance = Mahalanobis, NN = 15 0.82 0.71 0.81 0.70 0.82 0.71 −11.05
LM 𝜆 = 4.5281𝑒−5 , Learner = SVM 0.72 0.68 0.70 0.65 0.71 0.66 −5.09
NBC Distribution names = Kernel, Width = 0.0549 0.77 0.62 0.76 0.63 0.77 0.61 −15.33
MLP Activation = Sigmoid, Layer size = [50, 3, 7] 0.92 0.49 0.92 0.43 0.92 0.41 −51.82
SVM KS = 11.5592, Binary loss = Hinge 0.72 0.68 0.70 0.64 0.71 0.66 −5.48
DT Number of nodes = 215 0.88 0.60 0.87 0.56 0.88 0.54 −33.61

P = Precision, R = Recall, S = Kernel Scale, NN = NUmber of neighbors, n.a. = not available (i.e., model could not discriminate on of the three classes).

in the training set and dropping to 91.35% in the test set. Precision
measures the model’s effectiveness in correctly predicting specific cat-
egories while avoiding false positives. For the control category, the
model showed high precision in both sets, with scores of 99.17% and
96.36%. However, for early stress and late stress categories, precision
dropped when exposed to new test data, falling from 95.85% to 91.36%
and from 96.18% to 81.08%, respectively. Recall focuses on the model’s
ability to identify all relevant instances in specific categories, aiming to
reduce false negatives. In the control class, recall was high in both sets,
approximately 99.61% and 96.91%. However, it significantly dropped
for early stress and late stress in the test set, going from 96.78%
to 82.68% and from 94.86% to 87.50%. Such results allow us to
consider the combination of the MLP algorithms, together with feature
reduction through equivalent model fitting, as a promising tool for the
characterization of water stress in plants.
Furthermore, the effectiveness of the cole model parameters across
various machine learning techniques influenced their selection for char-
acterizing water stress in tomato plants in this study.
Henceforth, the analysis was conducted solely on this model, fo-
Fig. 4. Illustration of the Cole model’s classification performance using a confusion cusing on extracting meaningful insights from the Cole circuit model
matrix across three classes: control, early Stress, and late Stress, for the training Dataset parameters and their relationship with the plant’s response to water
(a) and test dataset (b). Rows represent the true class, while columns depict the stress. This streamlined approach not only enhances interpretability but
predicted class. Darker shades of blue indicate higher true positive occurrences.
also facilitates the practical application of the findings for agricultural
monitoring and management.

where unseen variations may exist. Fine-tuning the model parameters, 3.3. Dynamics of circuit parameters over time
incorporating additional relevant features, or exploring more advanced
machine learning techniques are potential avenues for improving the Using bioimpedance for continuous monitoring of plant stem phys-
model’s ability to generalize across diverse stress conditions. iology under stress conditions requires correlating the bioimpedance
In evaluating the machine learning model, metrics like accuracy, data with the underlying physiological changes occurring in the plant
precision, and recall were considered for both training and test sets. during water stress. Literature has shown that the circuit parameters
Accuracy indicates the model’s overall performance, scoring 97.29% of the equivalent model are correlated with physiological changes in

6
S. Hamed et al. Computers and Electronics in Agriculture 226 (2024) 109347

Fig. 5. Temporal changes in normalized circuit parameters of the Cole model. Evolution of series resistor 𝑅𝑠 (a), parallel resistor 𝑅𝑝 (b), magnitude component of the constant
phase element 𝐶𝑃 𝐸𝑡 (c), and phase angle component of the constant phase element 𝐶𝑃 𝐸𝑝 (d). The red line denotes plants subjected to water stress, while the blue line corresponds
to control plants.

the plant (González-Araiza et al., 2017; Ibba et al., 2019). Therefore, The traversal of electric current within fluidic media arises from the
in addition to examining the overall bioimpedance variations at dif- existence of charged entities, specifically ions, that exist in the phloem
ferent frequencies and conditions, as shown in Fig. 3, it is crucial to and xylem sap inside the plant stem. During plant stress, the primary
investigate changes in parameters which can be directly correlated with determinant influencing alterations in tomato stem bioimpedance is
specific plant features, such as the equivalent circuit elements. From attributable to a complex interplay of molecular and hormonal mod-
our previous analysis, we found that the equivalent circuit parameters ulations orchestrated by the plant to adapt to novel environmental
of the Cole model, typically used to represent the flow of current circumstances (Kuromori et al., 2022). This intricate process gives rise
in a biological tissue. The reason for such choice was twofold: (i) it to consequential transformations in the proteomic composition (Ogden
allowed to reduced our data from 200 single frequency points for both et al., 2020), ionic concentrations (Janni et al., 2019), and pheno-
Magnitude and phase angle spectra to 4 circuit component values. Such typic attributes of the stem, encompassing variations in diameter and
step had the advantage to be highly representative for the relevant sap flow dynamics (Vermeulen et al., 2007). Consequently, these al-
information contained in the main spectra while excluding the majority terations can be regarded as the principal determinant influencing
of the redundant one; (ii) The Cole model is widely used to represent modifications in the parameters of the equivalent circuit during stress.
the flow of current in the biological tissues, to which behavior are In the Cole model, bioimpedance measurements corresponding to 𝑅𝑠
linked the single circuit components. This allowed us to not only reduce and 𝑅𝑝 are utilized for the characterization of the extracellular and in-
our feature space, but also to select few parameters that could be tracellular resistive environments, respectively (González-Araiza et al.,
directly linked to a specific plant status. This is why, we are focusing 2017; Ibba et al., 2019). This phenomenon was explained from the
on this model only to understand its circuit parameters better. graphical time series in Fig. 5a, with the trend of the parameter 𝑅𝑠
Fig. 5 depicts the time-based trends of various circuit elements from remaining stable for both water-stressed and control plants during
the start of the stress application to the end of the recovery phase, for the experiment. This suggests that the resistance of the extracellular
both water-stressed and control tomato plants. Overall, seasonal pat- medium was consistent over time. In contrast, the trend for parameter
terns were clearly observed throughout the entirety of our experiment 𝑅𝑝 began to diverge between water-stressed and control plants starting
(Fig. 5). The presence of these circadian rhythms in plants has been in the late stress phase, as shown in Fig. 5b. This change continued
documented in prior research (Bukhamsin et al., 2021; Garlando et al., and became more pronounced after the recovery period, indicating
2022; Reynolds et al., 2020). This phenomenon can be elucidated by that water stress affected the resistance of the intracellular medium.
plant transpiration, which entails the emission of water vapor through Regarding the 𝐶𝑃 𝐸 parameters, seen in Fig. 5c and d, the trends were
stomatal openings. The rate of transpiration in tomato plants exhibits generally stable for water-stressed plants throughout the experiment.
fluctuations between day and night due to the closure or partial closure On the other hand, control plants showed some variations in 𝐶𝑃 𝐸𝑡 ,
of stomata during dark conditions when photosynthesis is inactive. In possibly due to changes in ion concentrations at different growth
these dark conditions, water and nutrients move from the shoots to the phases (Alfosea-Simón et al., 2022). The stable 𝐶𝑃 𝐸 trends in water-
roots through phloem tissue (Caird et al., 2007), which was proven to stressed plants could be due to the plants acclimation by adjusting their
be in contact with the electrode (see Fig. 2b). water uptake and loss mechanisms (Verslues et al., 2006).

7
S. Hamed et al. Computers and Electronics in Agriculture 226 (2024) 109347

4. Conclusion Declaration of competing interest

The authors declare the following financial interests/personal rela-

In summary, this research showcases the effectiveness of bioim-
tionships which may be considered as potential competing interests:
pedance techniques in detecting water stress in tomato plants, marking
Saleh Hamed reports financial support was provided by Italian Institute
a significant advancement in precision agriculture. With a compre-
of Technology. If there are other authors, they declare that they have
hensive dataset over 38 days and 16 000 measurements, and the use
no known competing financial interests or personal relationships that
of advanced circuit models and machine learning, the study achieves
could have appeared to influence the work reported in this paper.
impressive predictive accuracy. This work not only proves the practical
use of bioimpedance for identifying plant stress but also its potential Data availability
in monitoring stress progression, which is vital for timely interven-
tion and yield optimization. While the study identifies areas needing Data will be made available on request.
improvement, particularly in early and late stress detection, it paves
the way for future enhancements. Expanding the dataset, considering Declaration of generative AI in scientific writing
different plant varieties and environmental factors, and extending the
model to other species will further refine and broaden the applicability During the preparation of this work the author(s) used the free-to-
of this technology. Ultimately, this research opens exciting possibili- use AI system ChatGPT ([Link] in order to improve
ties for developing more robust, versatile decision support systems in language and readability of the manuscript. After using this tool, the
author(s) reviewed and edited the content as needed and take(s) full
agriculture, contributing to sustainable farming practices and resource
responsibility for the content of the publication.
management.
While the study is performed on a small number of plants, the
Acknowledgments
amount of tested bioimpedance datapoints can be considered enough
for the training and testing of a good quality classification model. We gratefully acknowledge the co-funding of this work by the Free
Nevertheless, we agree that in order to increase their robustness, such University of Bozen-Bolzano and the Istituto Italiano di Tecnologia
kind of studies need to be carried out on (i) a larger amount of samples (IIT). Special thanks are extended to Prof. Tanja Mimmo for her gen-
(in this case, more plants) and (ii) over several growing seasons. erous provision of the nutrient solutions, to Prof. Camilla Wellstein for
Regarding potential damage caused by needle insertion in plants, kindly providing the Safranin reagentm and to Dr. Maurizio Ventura
our methodology employed a sterilized stainless steel electrode, recog- for his technical support in the greenhouse. This study was carried
nized for its durability and resistance to corrosion. This material choice out within the Agritech National Research Center and received funding
is supported by its established use in bioimpedance measurements in from the European Union Next-GenerationEU (PIANO NAZIONALE
various studies (Muñoz-Huerta et al., 2014; Mizukami et al., 2006; DI RIPRESA E RESILIENZA (PNRR) – MISSIONE 4 COMPONENTE 2,
Meiqing et al., 2016; Jinyang et al., 2016). Our imaging analysis, INVESTIMENTO 1.4 – D.D. 1032 17/06/2022, CN00000022). This
Fig. 2, indicates the formation of cork tissue around the inserted needle, work was supported by the Open Access Publishing Fund of the Free
a phenomenon previously observed in Aspen trees when sensors are University of Bozen-Bolzano. This manuscript reflects only the authors’
implanted (Diacci et al., 2021). However, the impact on our measure- views and opinions, neither the European Union nor the European
Commission can be considered responsible for them. The graphical
ments was mild in comparison to that study. Despite this, the plants
abstract and Fig. 1 were created with Biorender.
in our study remained healthy, capable of flowering, and produced
tomato fruits, showing that the needle insertion did not significantly
Appendix A. Supplementary data
hinder their normal physiological functions. This outcome is consis-
tent with another study where cotton thread was inserted into the
Supplementary material related to this article can be found online
stem of tomato plants without inducing observable stress (Coppedè at [Link]
et al., 2017). To address concerns regarding invasiveness, we reference
the work on the use of microneedle plant electrodes. These devices References
provide a less invasive approach while maintaining high sensitivity
in impedance measurements (Bukhamsin et al., 2021). Additionally, Alfosea-Simón, M., Simón-Grao, S., Zavala-Gonzalez, E.A., Navarro-Morillo, I., Martínez-
innovative advancements have been made in the integration of printed Nicolás, J.J., Alfosea-Simón, F.J., Simon, I., García-Sánchez, F., 2022. Ionomic,
metabolic and hormonal characterization of the phenological phases of different
electronics with biocompatible materials, such as cryogels. This has led tomato genotypes using omics tools. Sci. Horticult. 293, 110697.
to the development of hydrogel-based bioelectronic devices, which can Araújo, S.O., Peres, R.S., Barata, J., Lidon, F., Ramalho, J.C., 2021. Characterising
be implanted in plant tissue for extended periods without significant the agriculture 4.0 landscape—emerging trends, challenges and opportunities.
Agronomy 11 (4), 667.
disruption (Bihar et al., 2023). We acknowledge the potential for tissue
Baluška, F., Mancuso, S., Van Volkenburgh, E., 2021. Barbara G. Pickard-queen of plant
response and the need for less invasive electrodes in the future. electrophysiology. Plant Signal. Behav. 16 (6), 1911400.
Bar-On, L., Shacham-Diamand, Y., 2021. On the interpretation of four point impedance
spectroscopy of plant dehydration monitoring. IEEE J. Emerg. Sel. Top. Circuits
CRediT authorship contribution statement Syst. 11 (3), 482–492.
Barezzi, M., Cum, F., Garlando, U., Martina, M., Demarchi, D., 2022. On the impact of
the stem electrical impedance in neural network algorithms for plant monitoring
Saleh Hamed: Writing – review & editing, Writing – original draft, applications. In: 2022 IEEE Workshop on Metrology for Agriculture and Forestry
Methodology, Investigation, Formal analysis, Data curation, Concep- (MetroAgriFor). IEEE, pp. 131–135.
tualization. Antonio Altana: Writing – review & editing, Writing – Bihar, E., Strand, E.J., Crichton, C.A., Renny, M.N., Bonter, I., Tran, T., Atreya, M.,
Gestos, A., Haseloff, J., McLeod, R.R., et al., 2023. Self-healable stretchable printed
original draft, Validation, Software, Methodology, Conceptualization. electronic cryogels for in-vivo plant monitoring. NPJ Flex. Electron. 7 (1), 48.
Paolo Lugli: Writing – review & editing, Supervision, Resources. Luisa Bukhamsin, A., Moussi, K., Tao, R., Lubineau, G., Blilou, I., Salama, K.N.,
Petti: Writing – review & editing, Supervision, Resources, Funding Kosel, J., 2021. Robust, long-term, and exceptionally sensitive microneedle-based
bioimpedance sensor for precision farming. Adv. Sci. 8 (16), 2101261.
acquisition. Pietro Ibba: Writing – review & editing, Writing – orig-
Caird, M.A., Richards, J.H., Hsiao, T.C., 2007. Significant transpirational water loss
inal draft, Visualization, Validation, Supervision, Investigation, Formal occurs throughout the night in field-grown tomato. Funct. Plant Biol. 34 (3),
analysis, Data curation, Conceptualization. 172–177.

8
S. Hamed et al. Computers and Electronics in Agriculture 226 (2024) 109347

Calvo, S., Barezzi, M., Demarchi, D., Garlando, U., 2023. In-vivo proximal monitoring Janni, M., Coppede, N., Bettelli, M., Briglia, N., Petrozza, A., Summerer, S., Vurro, F.,
system for plant water stress and biological activity based on stem electrical Danzi, D., Cellini, F., Marmiroli, N., et al., 2019. In vivo phenotyping for the early
impedance. In: 2023 9th International Workshop on Advances in Sensors and detection of drought stress in tomato. Plant Phenom..
Interfaces. IWASI, IEEE, pp. 80–85. Jinyang, L., Meiqing, L., Hanping, M., Wenjing, Z., 2016. Diagnosis of potassium
Cammarano, D., Jamshidi, S., Hoogenboom, G., Ruane, A.C., Niyogi, D., Ronga, D., nutrition level in solanum lycopersicum based on electrical impedance. Biosyst.
2022. Processing tomato production is expected to decrease by 2050 due to the Eng. 147, 130–138.
projected increase in temperature. Nat. Food 3 (6), 437–444. Kuromori, T., Fujita, M., Takahashi, F., Yamaguchi-Shinozaki, K., Shinozaki, K., 2022.
Cole, K.S., 1940. Permeability and impermeability of cell membranes for ions. In: Inter-tissue and inter-organ signaling in drought stress response and phenotyping
Cold Spring Harbor Symposia on Quantitative Biology. Vol. 8, Cold Spring Harbor of drought tolerance. Plant J. 109 (2), 342–358.
Laboratory Press, pp. 110–122. Liu, Y., Li, D., Qian, J., Di, B., Zhang, G., Ren, Z., 2021. Electrical impedance
Coppedè, N., Janni, M., Bettelli, M., Maida, C.L., Gentile, F., Villani, M., Ruotolo, R., spectroscopy (EIS) in plant roots research: a review. Plant Methods 17 (1), 1–25.
Iannotta, S., Marmiroli, N., Marmiroli, M., et al., 2017. An in vivo biosensing, Maver, M., Trevisan, F., Miras-Moreno, B., Lucini, L., Trevisan, M., Cesco, S.,
biomimetic electrochemical transistor with applications in plant science and Mimmo, T., 2022. The interplay between nitrogenated allelochemicals, mineral
precision farming. Sci. Rep. 7 (1), 16195. nutrition and metabolic profile in barley roots. Plant Soil 479 (1–2), 715–730.
Costa, J.M., Heuvelink, E., 2018. The global tomato industry. In: Tomatoes. CABI Meiqing, L., Jinyang, L., Hanping, M., Yanyou, W., 2016. Diagnosis and detection of
Wallingford UK, pp. 1–26. phosphorus nutrition level for solanum lycopersicum based on electrical impedance
Cum, F., Calvo, S., Demarchi, D., Garlando, U., 2023. Machine learning models compar- spectroscopy. Biosyst. Eng. 143, 108–118.
ison for water stress detection based on stem electrical impedance measurements. Mizukami, Y., Sawai, Y., Yamaguchi, Y., 2006. Moisture content measurement of tea
In: 2023 IEEE Conference on AgriFood Electronics. CAFE, IEEE, pp. 108–112. leaves by electrical impedance and capacitance. Biosyst. Eng. 93 (3), 293–299.
Dellis, J.-L., 2023. Zfit. MATLAB Central File Exchange. Muñoz-Huerta, R.F., de J. Ortiz-Melendez, A., Guevara-Gonzalez, R.G., Torres-
Diacci, C., Abedi, T., Lee, J.W., Gabrielsson, E.O., Berggren, M., Simon, D.T., Niittylä, T., Pacheco, I., Herrera-Ruiz, G., Contreras-Medina, L.M., Prado-Olivarez, J., Ocampo-
Stavrinidou, E., 2021. Diurnal in vivo xylem sap glucose and sucrose monitoring Velazquez, R.V., 2014. An analysis of electrical impedance measurements applied
using implantable organic electrochemical transistor sensors. iScience 24 (1). for plant N status estimation in lettuce (Lactuca sativa). Sensors 14 (7),
Eckardt, N.A., Cutler, S., Juenger, T.E., Marshall-Colon, A., Udvardi, M., Verslues, P.E., 11492–11503.
2023. Focus on climate change and plant abiotic stress biology. Plant Cell 35 (1), Najdenovska, E., Dutoit, F., Tran, D., Rochat, A., Vu, B., Mazza, M., Camps, C.,
1–3. Plummer, C., Wallbridge, N., Raileanu, L.E., 2021. Identifying general stress in
Ehosioke, S., Nguyen, F., Rao, S., Kremer, T., Placencia-Gomez, E., Huisman, J.A., commercial tomatoes based on machine learning applied to plant electrophysiology.
Kemna, A., Javaux, M., Garré, S., 2020. Sensing the electrical properties of roots: Appl. Sci. 11 (12), 5640.
A review. Vadose Zone J. 19 (1), e20082. Ogden, A.J., Bhatt, J.J., Brewer, H.M., Kintigh, J., Kariuki, S.M., Rudrabhatla, S.,
Elli, G., Hamed, S., Petrelli, M., Ibba, P., Ciocca, M., Lugli, P., Petti, L., 2022. Field-effect Adkins, J.N., Curtis, W.R., 2020. Phloem exudate protein profiles during drought
transistor-based biosensors for environmental and agricultural monitoring. Sensors and recovery reveal abiotic stress responses in tomato vasculature. Int. J. Mol. Sci.
22 (11), 4178. 21 (12), 4461.
Francini, A., Sebastiani, L., 2019. Abiotic stress effects on performance of horticultural Presti, D.L., Di Tocco, J., Massaroni, C., Cimini, S., De Gara, L., Singh, S., Raucci, A.,
crops. Horticulturae 5 (4), 67. Manganiello, G., Woo, S.L., Schena, E., et al., 2022. Current understanding,
Garlando, U., Calvo, S., Barezzi, M., Sanginario, A., Ros, P.M., Demarchi, D., 2022. challenges and perspective on portable systems applied to plant monitoring and
Ask the plants directly: Understanding plant needs using electrical impedance precision agriculture. Biosens. Bioelectron. 115005.
measurements. Comput. Electron. Agric. 193, 106707. Randles, J.E.B., 1947. Kinetics of rapid electrode reactions. Discuss. Faraday Soc. 1,
González-Araiza, J.R., Ortiz-Sánchez, M.C., Vargas-Luna, F.M., Cabrera-Sixto, J.M., 11–19.
2017. Application of electrical bio-impedance for the evaluation of strawberry Reynolds, J., Taggart, M., Lobaton, E., Daniele, M., Rufty, T., Bozkurt, A., 2020. An
ripeness. Int. J. Food Prop. 20 (5), 1044–1050. environmental station with bioimpedance capabilities for agricultural deployment.
Grimnes, S., Martinsen, Ø.G., 2006. Bioimpedance. Wiley Encyclopedia Biomed. Eng.. In: 2020 IEEE SENSORS. IEEE, pp. 1–4.
Hamed, S., Ibba, P., Altana, A., Lugli, P., Petti, L., 2023. Towards tomato plant Reynolds, J., Taggart, M., Martin, D., Lobaton, E., Cardoso, A., Daniele, M., Bozkurt, A.,
iron stress monitoring through bioimpedance and circuit analysis. In: 2023 IEEE 2023. Rapid drought stress detection in plants using bioimpedance measurements
Conference on AgriFood Electronics. CAFE, IEEE, pp. 20–24. and analysis. IEEE Trans. AgriFood Electron..
Hamed, S., Ibba, P., Petrelli, M., Ciocca, M., Lugli, P., Petti, L., 2021. Transistor- Roper, J.M., Garcia, J.F., Tsutsui, H., 2021. Emerging technologies for monitoring plant
based plant sensors for agriculture 4.0 measurements. In: 2021 IEEE International health in vivo. ACS Omega 6 (8), 5101–5107.
Workshop on Metrology for Agriculture and Forestry (MetroAgriFor). IEEE, pp. Schwarz, D., Thompson, A.J., Kläring, H.-P., 2014. Guidelines to use tomato in
69–74. experiments with a controlled environment. Front. Plant Sci. 5, 625.
Hoagland, D.R., Arnon, D.I., et al., 1950. The water-culture method for growing plants Tiitta, M., Tiitta, V., Heikkinen, J., Lappalainen, R., Tomppo, L., 2020. Classification of
without soil. Circ. Calif. Agric. Exp. Station 347 (2nd edit). wood chips using electrical impedance spectroscopy and machine learning. Sensors
Ibba, P., 2021. Fruit Quality Evaluation Using Electrical Impedance Spectroscopy (Ph.D. 20 (4), 1076.
thesis). Free University of Bozen-Bolzano. Vermeulen, K., Steppe, K., Linh, N.S., Lemeur, R., De Backer, L., Bleyaert, P., Dekock, J.,
Ibba, P., Cantarella, G., Abera, B.D., Petti, L., Falco, A., Lugli, P., 2019. Selection of Aerts, J., Berckmans, D., 2007. Simultaneous response of stem diameter, sap flow
cole model bio-impedance parameters for the estimation of the ageing evolution of rate and leaf temperature of tomato plants to drought stress. In: International Sym-
apples. In: International Conference on Electrical Bioimpedance. Springer Singapore posium on High Technology for Greenhouse System Management: Greensys2007
Singapore, pp. 25–32. 801. pp. 1259–1266.
Ibba, P., Crepaldi, M., Cantarella, G., Zini, G., Barcellona, A., Rivola, M., Petrelli, M., Verslues, P.E., Agarwal, M., Katiyar-Agarwal, S., Zhu, J., Zhu, J.-K., 2006. Methods and
Petti, L., Lugli, P., 2021a. Design and validation of a portable AD5933–based concepts in quantifying resistance to drought, salt and freezing, abiotic stresses that
impedance analyzer for smart agriculture. IEEE Access 9, 63656–63675. affect plant water status. Plant J. 45 (4), 523–539.
Ibba, P., Falco, A., Abera, B.D., Cantarella, G., Petti, L., Lugli, P., 2020. Bio-impedance Von Arx, G., Crivellaro, A., Prendin, A.L., Čufar, K., Carrer, M., 2016. Quantitative
and circuit parameters: An analysis for tracking fruit ripening. Postharvest Biol. wood anatomy—practical guidelines. Front. Plant Sci. 7, 781.
Technol. 159, 110978. Warburg, E., 1899. Ueber das verhalten sogenannter unpolarisirbarer elektroden gegen
Ibba, P., Falco, A., Rivadeneyra, A., Lugli, P., 2018. Low-cost bio-impedance analysis wechselstrom. Ann. Phys., Lpz. 303 (3), 493–499.
system for the evaluation of fruit ripeness. In: 2018 IEEE SENSORS. IEEE, pp. 1–4. Ward, L.C., Brantlov, S., 2023. Bioimpedance basics and phase angle fundamentals.
Ibba, P., Lugli, P., Petti, L., 2022. Cole bioimpedance parameters influence on straw- Rev. Endocr. Metab. Disord. 24 (3), 381–391.
berry ripening binary classification using decision trees. Int. J. Bioelectromagn. Zhang, M., Willison, J., 1991. Electrical impedance analysis in plant tissues11. J. Exp.
24. Bot. 42 (11), 1465–1475.
Ibba, P., Tronstad, C., Moscetti, R., Mimmo, T., Cantarella, G., Petti, L., Martinsen, Ø.G., Zhou, R., Yu, X., Ottosen, C.-O., Rosenqvist, E., Zhao, L., Wang, Y., Yu, W., Zhao, T.,
Cesco, S., Lugli, P., 2021b. Supervised binary classification methods for strawberry Wu, Z., 2017. Drought stress had a predominant effect over heat stress on three
ripeness discrimination from bioimpedance data. Sci. Rep. 11 (1), 11202. tomato cultivars subjected to combined stress. BMC Plant Biol. 17 (1), 1–13.

View publication stats