Computers and Electronics in Agriculture 223 (2024) 109126

Contents lists available at ScienceDirect

Computers and Electronics in Agriculture

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

Artificial intelligence to optimize water consumption in agriculture: A

predictive algorithm-based irrigation management system
Luca Preite *, Giuseppe Vignali
University of Parma, Department of Engineering and Architecture, Parco Area delle Scienze 181/A, Parma 43124, Italy

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

Keywords: Water scarcity is a pressing global issue that needs to be faced. The United Nations highlights that only about 31
Water management percent of the population is not characterized by water stress, meaning that the world’s freshwater resources are
Precision agriculture unevenly distributed and unsustainably managed. This phenomenon is both natural and man-made, as the
Sustainable farming
human footprint is linked to many disciplines, with the largest impact coming from the agricultural sector.
Machine learning
Predictive algorithm
Scientific literature highlights how 4.0 technologies are the key drivers to reduce this sector’s impact on valuable
resources, such as water and soil.
Based on these premises, the proposed work aims at improving the irrigation management in agriculture by
implementing a three-layer architecture system to optimize water consumption and prevent soil percolation, by
avoiding situations where the soil moisture exceeds its capacity point. To achieve this, experimental activities
allow the evaluation of the soil capacity point and thus the definition of a confidence interval to guide watering
decisions. The latter interval, soil and environmental data, and three-day weather forecasts are aggregate to
create a consistent dataset for training and testing three different machine learning algorithms based on a
classification problem to predict the state of the irrigation network.
As a result, the implemented multi-layer perceptron neural network, support vector machine, and k-neighbors
classifier achieved an accuracy of nearly 99%. Despite this, the neural network produced superior decision region
boundaries, resulting in fewer false predictions. A Monte Carlo simulation was then applied to evaluate the water
and energy savings, which were up to 27 % and 57 %, respectively.
In summary, the predictive algorithm-based irrigation management system is a cost-effective solution for
optimizing water management in agriculture that it is truly scalable to any crop by assessing the appropriate soil
capacity level.

1. Introduction to improve the sustainability of the agricultural sector with multiple

sources, such as blockchain, internet of things technologies (IoT), deep
Water scarcity is a critical challenge for sustainable development learning and machine learning algorithms, and other computer appli­
that is truly influenced by the human footprint and climate change. cations (Preite et al., 2023b). Predictive algorithms are straightforward
According to the United Nations, a state of water stress is defined as a 25 solutions widely used in a 4.0 scenario across multiple disciplines
% of water withdrawal from the renewable freshwater resources. Sus­ because they provide highly scalable automation (Mazzei & Ramjattan,
tainable Development Goal indicator 6.4.2 states that in 2018, 18.4 % of 2022). The scientific literature shows a few of machine learning ap­
the global renewable freshwater was withdrawn by all economic activ­ proaches in agriculture, which can be categorized into three groups: pre-
ities (United Nations, 2024). One of the largest contributions is strongly harvesting, harvesting, and post harvesting (Liakos et al., 2018,
related to agricultural activities, as the irrigation management accounts Meshram et al., 2021). Applications identified for irrigation purposes,
for nearly 70 % of the global water withdrawals (Food and Agriculture such as water scarcity recognition, water demand predictions, and irri­
Organization of the United Nations, 2023). In this framework, future gation scheduling were in the first group. The irrigation is traditionally
research aims at addressing the water scarcity issue from different action scheduled by considering a fixed period and overlooking the variability
points. Therefore, 4.0 technologies are emerging as a potential solution in environmental and plant properties. Specifically, water scarcity

* Corresponding author.
E-mail address: [email protected] (L. Preite).

https://doi.org/10.1016/j.compag.2024.109126
Received 26 March 2024; Received in revised form 30 May 2024; Accepted 1 June 2024
Available online 6 June 2024
0168-1699/© 2024 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
L. Preite and G. Vignali Computers and Electronics in Agriculture 223 (2024) 109126

recognition deals with the evaluation of the drought stress, stem water Evapotranspiration is an important variable for implementing ac­
potential and the plant water content by processing thermal infrared curate irrigation decisions, which reflects both the metabolic activities
images, meteorological and soil data. In this case, convolutional neural of the plant and the meteorological data. The authors of (Aly et al.,
network, random forest, and gradient-boosted decision trees were the 2024) exploited extra tree regressor, support vector regressor, k-nearest
main employed models (Zhou et al., 2021). Multispectral and thermal neighbor and AdaBoost regression for feeding data into a super learner
images in combination with meteorological and soil data were also used ensemble to estimate the reference evapotranspiration, achieving high
to estimate the soil moisture content, reference evapotranspiration and accuracy with limited meteorological data. The latter issue was also
sap flow properties by implementing support vector machine, artificial highlighted by (Yong et al., 2024) as the main limitation for predicting
neural network, and gradient-boosting decision trees algorithms (Ali­ evapotranspiration rate, pointing out a hybrid neuro-fuzzy inference
babaei et al., 2021). The authors of (Corell et al., 2016) provide an system for counteract this problem. In this perspective, the author of
irrigation framework based on sensors data for comparing three re­ (Adnan et al., 2021) investigated the applicability of hybrid support
gressor models to determine the suitable irrigation volume for an olive vector regression models, which combine machine learning techniques
cultivation. Corn, kiwi, and potato crops were monitored to implement a with optimization meta-heuristic algorithms (i.e., Particle Swarm Opti­
fuzzy decision support system for evaluating the proper irrigation mization, Whale Optimization, Differential Evolution, and Covariance
amount by considering the growing degree days, the amount of water Matrix Adaptation Evolution Strategy). Following the same logical path,
delivered to the plants, and the evapotranspiration rate (Giusti & (Považanová et al., 2023) improved prediction accuracy by combining
Marsili-Libelli, 2015). A partial least-square regression and an adaptive machine learning and feature engineering for estimating reference
neuro fuzzy inference system were applied to provide watering recom­ evapotranspiration, providing valuable insight into the effectiveness and
mendations for lemon trees by processing soil moisture, evapotranspi­ the generalization of the proposed models. The authors of (Youssef et al.,
ration, and humidity (Navarro-Hellín et al., 2016). Soil moisture trends 2024) showed how several machine learning algorithms, such as k-
in both vertical depth and time were assessed by (Chandrappa et al., nearest neighbors, support vector machine, multilinear regression, and
2022), which develops both machine learning (support vector regression decision tree achieved an accuracy close to 99 % in estimating reference
and linear regression) and deep learning algorithms (long short-term evapotranspiration.
memory). In this background, a significant correlation between wind The scientific literature also outlines the main challenges of applying
speed and soil moisture in multi depth was highlighted. A water saving machine learning to irrigation. First, these applications require high
of 20 % was achieved by (Gu et al., 2021) developing an artificial neural quality data sources for ensuring an effective learning rate and accurate
network capable to determine the irrigation duration working on data predictions. Thereby implementing a dataset for an entire irrigation
provided by soil sensors and weather station. Artificial intelligence season is challenging because the growth period is truly long, and many
application was also explored by (Kavya et al., 2023) for predicting short parameters influence the crop quality and water demand. In this context,
term water demand. Specifically, the predictive performance of the scalability is the other pressing issue as many predictive models fail
machine learning and deep learning was evaluated with both univariate when crop, soil, environmental and irrigation conditions differ (Sharma
and multivariate time series. The univariate series has been imple­ et al., 2021).
mented considering only the flow meter data, meanwhile meteorolog­ The main aim of this study is to address the water shortage issue by
ical data were also considered for the multivariate scenario. Time series proposing a machine learning application to schedule the irrigation by
forecasting to estimate soil moisture for irrigation purposes has been processing soil, environmental and weather forecast data. The experi­
also explored by (Custódio & Prati, 2024), which compares the perfor­ mental activities were carried out in a living lab developed within the
mance of modern and traditional modeling methods by testing three framework of the National Research Center for Technology in Agricul­
different scenarios. The first aimed at gaining insight into the impact of ture. The living lab focused on a tomato crop (Solanum Lycoperium L.
the training window size on the prediction performance of the multi­ Cv. HEINZ 1301) irrigated according to the watering recommendation
variate algorithms. In the second scenario, the same test was repeated provided by the IRRIFRAME platform, which amounts to 324.5 mm for a
but the frequency of fitting the algorithms was modified, and finally, in cultivated surface of 0.0132 ha (Irriframe, 2023). In this framework, a
the third scenario, the most performing multivariate algorithm was LoRaWAN-based IoT network has been installed to real-time monitoring
compared with univariate algorithms by evaluating ten datasets from soil and environmental conditions and controlling the water delivery.
different regions of the world. As a result, the authors stated that lower The collected data were aggregate with the three-day weather forecast
prediction capabilities resulted with algorithms trained with short- for implementing a dataset to train and validate a predictive algorithms-
period samples. Furthermore, multivariate random forest, and extreme based irrigation management system able to avoid overwatering by
gradient boosting showed higher performance than statistical methods, maintaining an adequate soil moisture level, which ranges around the
such as vector autoregression. Finally, the univariate methods out­ soil capacity point. In this framework, the soil capacity point is experi­
performed the multivariate methods, suggesting that exogenous vari­ mentally evaluated by computing the time derivative of the soil mois­
ables have limited influence in predicting soil moisture time series data. ture to identify the area where the rate of decline is slower after an
In (Mokhtar et al., 2023), different machine learning algorithms were irrigation or a precipitation event. From this perspective, an irrigation
developed to evaluate irrigation water requirements starting from the confidence interval has been defined, which allows water and energy
maximum and minimum temperature, relative humidity, wind speed, consumptions to be minimized to increase the sustainability of the
precipitation, root depth, basal crop coefficient, evapotranspiration, and irrigation at the farm level. Specifically, the proposed framework re­
the fraction of the wetted and exposed soil. High performance was sponds to a multi-classification problem, where the outputs are used to
reached with the convolutional neural network and the long short-term trigger the control valves or to send warning messages requiring farmer
memory, meanwhile the support vector machine resulted in the lowest actions. Deeping into the technical features, three different supervised
accuracy. The authors of (Srivastava et al., 2024) developed a proba­ models have been selected by considering the most investigated appli­
bilistic framework to determine irrigation strategies based on three cations in the scientific literature previously described. Specifically, two
different parameters: soil moisture, leaf area index, and evapotranspi­ of the most widely used machine learning algorithms for classification
ration. The described indicators reflect the soil water deficit, crop water purposes, k-nearest neighbors, support vector machine, were compared
stress, and the water demand, respectively. In this framework, a random with a multi-layer perceptron artificial neural network to define the
forest regressor was used to determine suitable predictors for each more suitable solution. Therefore, the first has been selected for its
parameter, while side the predictions were entrusted to a recurrent flexibility in handling different types of data, but it was highly sensitive
artificial neural network (long short-term memory). Finally, the result­ to outliers. Support vector machine is more complicated to train for
ing weights were adjusted by comparing the predicted and actual values. multi-classification problems, but it provides robustness in dealing with

2
L. Preite and G. Vignali Computers and Electronics in Agriculture 223 (2024) 109126

outliers and avoiding overfitting. In addition, it encompasses kernel 2. Materials and methods
functions to increase its effectiveness in high-dimensional spaces.
Finally, multi-layer perceptron neural network is one of the most 2.1. Living lab description
widespread feedforward deep learning algorithms due to its reliability
and its capability to maximize the accuracy by using a back-propagation An experimental tomato crop has been tested throughout the 2023
algorithm during the training phase (Hemalatha K. et al., 2017). Pre­ summer season in a living lab located in Parma, northern Italy
diction performance is evaluated by defining the resulting decision (44◦ 48′34″N 10◦ 16′28″E). The living lab was developed by the University
boundary regions, the confusion matrix and calculating various in­ of Parma in collaboration with the Azienda Sperimentale Stuard within
dicators, such as accuracy, precision, recall, F1-score, the corresponding the framework of the National Research Center for Agricultural Tech­
macro-average precision, and the weighted-average precision. nologies. Specifically, the Solanum Lycoperium L. Cv. HEINZ 1301 was
Summarizing the results, all the models associate the classification grown according to the organic guidelines and specific instructions
label to unknown data with high degree of accuracy, close to 99 %. given by the involved research teams to carry out various experimental
Although the high accuracy achieved by the three algorithms, signifi­ activities on different issues. In this perspective, a set of parameters were
cant differences are highlighted when considering the resulting decision continuously monitored throughout the experimental season by a three-
boundaries and the resulting confusion matrix. Indeed, the artificial layer LoRaWAN network-based framework. Specifically, it provides
neural network shows more effective decision regions that allow to irrigation recommendations based on a three-day weather forecast and
minimize false predictions. In addition, this framework can be truly data collected from the device layer installed in the living lab. Fig. 1
scaled to other crops and soil conditions since the new operating con­ shows how the proposed system was characterized by an acquisition
ditions can be simulated by evaluating the new soil capacity point. layer, which enables real-time data monitoring and data collection into a
Water and energy savings are determined by performing a Monte Carlo network server. An intermediate layer allows data to be processed and
simulation of 1000 tomato seasons and comparing the results with the stored in cloud, where they can be used by user applications or pro­
amount of water recommended by the IRRIFRAME platform. The first cessing nodes.
one ranges from 14.5 % to 27.6 %, while the second ranges from 49.2 %
to 57 %.
The methodology adopted for the present work is described in the 2.2. Acquisition data layer: LoRaWAN network and weather forecast API
next section, while the results are reported in section 3 and then dis­
cussed in section 4. The final section summarizes the conclusion, high­ The acquisition data layer was characterized by two different data
lighting future development and limitations. sources. On the one hand, a few of end nodes (ENs) were installed in the
field to collect data on soil and environmental conditions, and to manage

Fig. 1. Three interrelated layers architecture based on crop data, three-day weather forecast and machine learning algorithms.

3
L. Preite and G. Vignali Computers and Electronics in Agriculture 223 (2024) 109126

the irrigation network. Specifically, signals were transmitted every 10 Data analysis represents a crucial step to provide high quality infor­
min over LoRa communication protocol between the end nodes and a mation to make predictions.
gateway (GW) located in a warehouse in the vicinity of the experimental ∑
(xi − xm )(yi − ym )
field. The latter communicated with a network server (NS) using an IP- r = √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ (1)
∑
based protocol that allows data storage and processing. A summary of (xi − xm )2 (yi − ym )2
IoT devices characterizing the end nodes block is provided below:
∑
(6 d2 )
rs = 1 − (2)
• Water meters to monitor water consumption (MClimate, 2023). n(n2 − 1)
• Water valves for the collection of data on the status of the water
supply and the automation of irrigation strategies (ON/OFF) (MCli­ 2.3.2. Processing node: irrigation policy management
mate, 2023). The amount of water available to the plant held in the soil ranges
• 8-inch depth soil sensors to gain insights into the volumetric water from a lower limit, known as the permanent wilting point, to an upper
content, temperature, and soil conductivity (MClimate, 2023). limit, known as the soil capacity. The magnitude of this range is highly
• Environmental and weather sensors to collect weather and local dependent on soil structure and texture. Specially, the soil is charac­
environmental data (MClimate, 2023). terized by macropores and micropores, which regulates the drainage
flow and the amount of water available for the plant, respectively
Concurrently, water management was optimized by aggregating (Skovdal Christiansen et al., 2004). Irrigation or precipitation fills mi­
real-time data with three-day weather forecasts. Specifically, an open- cropores that supply water to the plant. Meanwhile, water drains out of
source web application programming interface (API) was used to the macropores by gravity, trapping air. From this perspective, when the
collect hourly precipitation, relative humidity, and evapotranspiration soil moisture exceeds the field’s capacity, overwatering occurs because
reference rate (Open Meteo, 2024). The API provide a set of weather the water percolating to the lower layers of the soil cannot be used for
forecast based on various weather model dependent on the geographic the growth (Warrick, A.W., 2001).
region, such as the Integrated Forecasting System (IFS) for the European The proposed framework provides an irrigation policy management
Union and COSMO 2I & 5 M for Italy (Bottazzi et al., 2021). based on the significant parameters resulting by the preprocessing steps.
The API request has been formulated by setting the living lab co­ Specifically, over irrigations were performed to study soil saturation at
ordinates, i.e., latitude = 44,1125 and longitude = 10,411, the European the depth of tomato root. Water content in the soil has been monitored
time zone (GMT + 1) and the reference period starting from July 28th to by the IoT soil sensors. As shown in Fig. 2, soil moisture declines at a
September 3rd, 2023. faster rate after the irrigation event ends up to the field capacity point.
From this point, the water content decreases more slowly because the
macropores are already drained and the water from the micropores is
2.3. Water management framework withdrawn by evapotranspiration, which includes evaporation and
transpiration. The first is the percentage of water that evaporates from
2.3.1. Data analysis the soil surface and the capillary fringe of the groundwater. The second
Using Python Pandas library, a data set having 16,187 samples for is the process by which plants release water to the atmosphere due to
each parameter was implemented by removing redundant and inaccu­ their metabolic activities. The above-mentioned API provide the hourly
rate data. The raw data was scrubbed for errors, duplicates or missing reference evapotranspiration evaluated according to the Penman –
data, providing high quality information for further analyses and pro­ Monteith method as set in (Pereira et al., 2015). It represents a recog­
cessing. The implemented dataset is described in Table 1, which reports nized evapotranspiration rate calculated by considering clipped grass
the main statistical metrics. under given parameters. Hence, the specific crop evapotranspiration can
Statistical methods were used to examine patterns among the data be calculated by multiplying the reference value by the crop coefficient,
collected. Specifically, the associations between the measured variables which reflects the different contribution given by each specific crop.
were identified by calculating two different correlation coefficients. According to equation (3), reference evapotranspiration identifies the
First, the Pearson correlation coefficient (r) was evaluated according to local climate and the crop drought conditions by considering the net
equation (1) to quantify the strength of the linear relationship between surface radiation Rn, soil heat flux density G, the mean hourly air tem­
the parameters and to identify the corresponding direction. In this case, perature at 1.5 to 2.5 m T, mean hourly wind speed at 2 m height u2,
xi is the value of an analyzed variable and yi represents the value of pressure vapor saturation es at 1.5 to 2.5 m height, the mean actual
another one. At the same time, xm and ym are the corresponding mean vapor pressure at the same height ea, the slope of the pressure vapor
values. saturation vs temperature Δ, psychrometric constant γ, and two con­
Furthermore, the correlation between the ranks associated to the raw stants related to the crop type Cn and Cd.
data was explored by evaluating the Spearman correlation coefficient
(rs) according to equation (2), where d represents the differences be­
tween the ranks for each item and n is the size of the samples analyzed.

Table 1
Statistical metrics of the dataset used to implement the predictive system.
End Nodes Data Weather Forecast Crop Data
Irrigation Soil Soil Soil Electrical Daily Environmental Environmental Rainfall Environmental Evapotranspiration
(ON/OFF) Moisture Temperature Conductivity Hour Temperature Humidity [RH [mm] humidity [RH [mm]
[RH%] [◦ C] [◦ C] %] %]

Count 16,187 16,187 16,187 16,187 16,187 16,187 16,187 16,187 16,187 16,187
Mean 0.59 23.28 23.27 277. 84 11.49 25.66 56.07 0.098 56.25 0.177
Std 0.49 4.2 1.43 148.42 6.93 5.1 17.65 0.067 17.81 0.20
Min 0 16.45 19.6 58 0 14 18 0 18 0
25 % 0 20.11 22.4 148 5 22 42 0 42 0
50 % 1 22.66 23.4 261 11 25.1 55 0 55 0.5
75 % 1 25.51 24.2 375 18 29.5 69 0 70 0.33
Max 1 41.26 27.1 1484 23 38.9 96 10.6 96 0.67

4
L. Preite and G. Vignali Computers and Electronics in Agriculture 223 (2024) 109126

Fig. 2. a) Soil moisture properties derived from the data acquisition layer. The estimated soil capacity level is highlighted by the red line. b) Confidence interval used
to drive irrigation decisions, where the lower and upper boundaries are calculated by subtracting and adding the half of the standard deviation to the corresponding
mean value.

0.408Δ(Rn − G) + γ(Cn /T + 273)u2 (es − ea ) Irrigation and precipitation events influenced the amount of water
ETref = (3)
Δ + γ(1 + Cd u2 ) retained by the soil. For the former, irrigation timing and volume
affected water delivery. In the proposed living lab, the irrigation
As a result, the soil field capacity was considered as a reference value
network was characterized by ON/OFF control valves that allowed to
to guide the irrigation policy management, since it represents the limit
deliver a constant flow rate, close to 300 liters per hour (Preite et al.,
to avoid overwatering. Specifically, for its experimental evaluation, the
2023a). In this perspective, the timing was the only parameter used to
time derivative of the soil moisture was calculated to identify the zone
control the amount of water available to the plant.
where the rate of decline is slower after an irrigation event. As shown in
The proposed approach also includes the rainfall contribution by
Fig. 2, the resulting data were filtered for evaluating statistical metrics,
exploring the three-day weather forecast. The main purpose is to define
such as mean (μ) and standard deviation (σ) to design a confidence in­
a reference value that would allow non-irrigation when the predicted
terval for an irrigation framework based on soil capacity point. This
event will have a significant saturating effect on the soil. Precipitation
interval focused on a limited region of the normal distribution by
amount is expressed in millimeters, which corresponds to the volume of
calculating the lower and upper confidence limits by subtracting and
water (expressed in liters) per square meter of water on the surface. The
adding the half of σ to μ, respectively.
reference value was calculated considering the amount of water

Fig. 3. Flow diagram of the adopted irrigation framework, which defines four different irrigation status: on, off, no adjustment and alert.

5
L. Preite and G. Vignali Computers and Electronics in Agriculture 223 (2024) 109126

distributed by the irrigation system. Specifically, 300 liters of water per outliers. The right value of ξ is evaluated by using C as regularization
hour was delivered to an area of 132 square meters, almost 2 mm of parameter (Lasso regularization). According to equation (4) smaller C
water. As a result, if the predicted precipitation is higher than 2 mm the emphasizes the influence of ξ resulting in wider margin that can cause
irrigation is not activate in the next days. In this case, the soil moisture is misclassification, meanwhile, larger C reduced the impact of ξ resulting
anyway monitored to avoid significant drought stress to the crop by in smaller margin with a less tolerance to outliers (Awad, M., et al.,
considering a lower confidential limit of μ – σ. An alert message is sent 2015).
for manual handling when the soil moisture falls below this critical ∑
value. The adopted irrigation framework is summarized in Fig. 3. C ξn (4)

If the data cannot be linearly separable, kernel functions are used to

2.3.3. Irrigation predictive algorithms move through another dimensional space to define the better hyper­
Predictive algorithms were developed with Python using scikit-learn plane. In the proposed framework, the model has been trained by
library to enhance the sustainability of the irrigation management. considering the gaussian kernel function, which has one parameter γ.
Specifically, three machine learning algorithms were trained using the According to equation (5) small value of γ results in a linear behavior
data set implemented according to the proposed irrigation framework. and large value can fall in overfitting. The model has been implemented
In this case, the issue deals with a classification problem, where four by tuning the hyperparameters C and γ.
classes label are predicted starting from proper input data: )
rbf(Xi, Xj) = exp( − γ‖Xi − Xj ‖2 (5)
• Irrigation OFF = 0, to close the control valve and turn off the wa­
An artificial neural network has been trained by using back­
tering when the soil moisture exceeds than the upper confidential
propagation learning rules to classify new watering recommendations.
limit previously introduced (3467 records in the dataset).
Specifically, a supervised multi-layer perceptron neural network has
• Irrigation ON = 1, to turn on irrigation on by moving the control
been used to learn a non-linear function providing a classification output
valve when the soil moisture falls below the lower confidential limit
based on the input features (Zou, J., et al., 2008). Two different solvers
(8875 records in the dataset).
were tested to train the classifier, i.e., stochastic gradient descent (sgd)
• No adjustment = 2, if the watering status is not going to be changed
and adaptive moment estimation (adam). Hidden layers perform the
when the soil moisture falls within the confidence range designed
main computational part of the model, as they transform the values
(3780 records in the dataset).
transferred from the previous or input layer by a weighted linear com­
• Alert = 3, a warning message requiring farmer actions, when pre­
bination plus a bias and then computing a non-linear activation func­
cipitations are detected but the soil moisture is below a critical value
tion. In this context, the lower bound of the hidden layer sizing was
(65 records in the dataset). In this framework, the critical value has
evaluated based on the framework provided by (Kuri-Morales, 2017)
been defined as μ – σ .
and it was used for test different combinations. Hence, a single layer
with two, or six, or twelve neurons and a three layers layout with six,
For the sake of simplicity, the output labels were changed to nu­
twelve, and six neurons were considered to perform the cross-validation
merical values during the training and testing phases. Three models
process. Furthermore, two activation function associated to the hidden
were trained by using three supervised classifiers, i.e., k-nearest neigh­
layers, i.e., hyperbolic tangent function (tanh) and the rectified linear
bors, support vector machines, and artificial neural network. To train
unit (relu), were tested to define the better combination of the hyper­
and test each model, the database resulting from the pre-processing was
parameters. The multi-label classification was finalized by applying
split into two datasets following the rules 70/30, respectively. Usually, a
Softmax activation function for the output layer, which is characterized
standardization of the input data is recommended to train the mentioned
by four neurons. Specifically, this function is a generalization of the
algorithms, since they can be truly sensitive to the scale of the input
sigmoid formula used for the logistic regression in binary classification
data. Since, the performance yielded by working with the original
and computes a normalized probability distribution of the values to
dataset was almost identical to that obtained with the z-score normal­
determine the model output. The other hyperparameters were the initial
ized dataset, this issue is not significant. As a result, the data were
learning rate and the regularization index. Lasso regularization was used
processed in the non-normalized format to give predictions directly in a
to define the appropriate value of the weight associated with the values
meaningful measurement.
along the network. Fine tuning of the alpha parameters (regularization
K-nearest neighbors stores the data available in the training dataset
parameter) was crucial to avoid overfitting. At the same time, the initial
and classifies new data based on the feature similarity given by the
learning rate controlled the weight updating during this process.
majority votes of its k neighbors in an n-dimensional space (Kramer,
Each model has been improved through a heuristic approach.
2013). For that purpose, the nearest neighbors of an unknown point are
Indeed, several hyperparameter combinations have been automatically
defined by computing the Euclidean distance from all the points in the
tested by using GridSearchCV, since there is no way to define in advance
dataset. In this case, k is one of the hyperparameters that needs to be
the optimal values that maximize the model performance. This function
tuned to maximize the accuracy of the model. The other is the weighting
is provided by the scikit-learn library and evaluates the model perfor­
function used to make predictions, which can take on two options,
mance via three-fold cross validation, which is a statistical method to
uniform, and distance. The uniform function assigns the same weight to
compare the partitions involved in the analysis.
each neighborhood. Meanwhile, the associated weight is inversely
Once the optimal combinations had been defined, the models were
proportional to the distance.
validated against the test data by means of different metrics, such as
Support vector machine categorizes unknown data by considering
confusion matrices and accuracy score. Specifically, three different pa­
the optimal hyperplane corresponding to the maximum distance, d,
rameters, i.e., precision, recall, and F-1 score, were evaluated by
between the support vectors in each class. Indeed, support vectors
considering the resulting true positive (TP), true negative (TN), false
represent the data points closest to the hyperplane, which are the only
positive (FP), and false negative (FN) indices. The first describes the
parameters influencing the final decision, unlike some other machine
number of samples that were correctly predicted. There were false
learning algorithms such as artificial neural network. A proper weight W
positive and false negative when the predictions belonged to a different
is assigned to the identified support vectors and predicts the output by
class in the real system. Finally, true negative was observed when a class
an optimization problem of the maximum distance d. As the real data are
was not correctly predicted. Table 2 reports the 4x4 confusion matrix
commonly noise, the classifier cannot tolerate outliers and fails in
resulted for each model, since the classification output was character­
optimizing the maximum distance. In this case, a variable ξ is added to
ized by four different classes (i.e., OFF, ON, No adjustment, and Alert).
the constraints represented by the support vectors to consider the

6
L. Preite and G. Vignali Computers and Electronics in Agriculture 223 (2024) 109126

Table 2 2.4. Water and energy savings

4x4 confusion matrix used to assess the algorithms’ performance.
ActualSamples OFF TNALERT TNALERT TNALERT FPALERT The proposed framework allows to regulate the soil moisture in a
TNNAD TNNAD TNON TNNAD neighborhood of the capacity level to avoid percolation phenomena. In
TNON FPON FPNAD TNON this perspective the potential water and energy savings have been
TPOFF FNOFF FNOFF FNOFF
probabilistically determined by performing a Monte Carlo simulation of
ON TNALERT TNALERT TNALERT FPALERT
TNNAD TNNAD FPNAD TNNAD 1000 tomato seasons. For any uncertain variable, this method leverages
FNON TNOFF FNON FNON a probability distribution to develop a model of possible outcomes (Zio,
FPOFF TPON TNOFF TNOFF E., 2013). Specifically, starting from a normal distribution characterized
No TNALERT TNALERT TNALERT FPALERT by the soil capacity level as the mean value and μ-σ/2 as the standard
adjustment TNON FNNAD TNON TNON
deviation, the resulting activation inputs provided by the prediction
FNNAD FPON TPNAD FNNAD
FPOFF TNOFF TNOff TNOff system were used to preliminarily evaluate a range of the potential
Alert FNALERT FNALERT FNALERT TNNAD water savings along a conventional irrigation period of 70 days (i.e.,
TNNAD TNNAD TNON TNON 100,800 min, 10,080 samples considering the operating frequency of the
TNON FPON FPNAD TPALERT
LoRaWAN network). An activation input results in an irrigation time of
FPOFF TNOFF TNOFF TNOFF
OFF ON No Alert 10 min delivering nearly 300 liters/hour. In this way, the total water
adjustment consumption has been evaluated and compared with the watering
Predicted Samples recommendation provided by the IRRIFRAME platform. Following the
same framework, energy savings were evaluated by considering the
power supply of the pump driving the irrigation system, which is 4 kW.
For better classification performance, TP and TN should be maxi­
The output of the Monte Carlo simulation is a normal distribution of
mized and FP and FN should be the smallest. Thereby, precision ranges
the water and energy savings. Specifically, an appropriate confidence
from 0 to 1 and describes what percentage of the specified class is
interval with 95 % coverage probability has been provided by plus and
actually true.
minus two estimated standard deviation from the estimated mean value.
∑
TPi
PrecisionCLASS = ∑ ∑ (6)
TPi + FPi 3. Results

Recall shows the true positive rate, which is the percentage of the
3.1. Data distribution and correlation
predictions belonged to the given class.
∑
TPi Fig. 4 (a-b) reports the correlation matrices describing the strength
RecallCLASS = ∑ ∑ (7)
TPi + FNi and the direction of the correlations of the data collected according to
Pearson and Spearman ranks.
F1 – score considers both false positives and false negatives by As it can be seen before, the performed correlation analyses provided
computing the harmonic mean of precision and recall. consistent results that allow to reshape the data frame to train and test
2*Precision*Recall the predictive algorithms. In this perspective, the parameters charac­
F1 − scoreCLASS = (8) terized by redundant correlations in strength and direction with the
Precision + Recall
irrigation decision were dropped to avoid overfitting.
Macro average and weighted average were also determined to take
According to the proposed framework, the parameters were grouped
into account the influence of all classes. Thus, the former represents the
into three classes considering the corresponding relationship:
arithmetic average of the precision of all classes, while the latter is based
on the precision classes weighted by considering the number of samples
• Soil properties.
contained in each class.
• Environmental properties.

PrecisionOFF + PrecisionON + PrecisionNAD + PrecisionALERT

Macro − Average Precision = (9)
4

PrecisionOFF *NOFF + PrecisionON *NON + PrecisionNAD *NNAD + PrecisionALERT *NALERT

Weighted − Average Precision = (10)
NTOT

• Weather forecast parameters.

Finally, the prediction capabilities are evaluated by computing the
accuracy score according to equation (11), where y represents the real Soil properties and evapotranspiration had the most significant
value and ŷ the corresponding predicted value. correlation. Considering the first one, moisture content showed the
nvalues
highest correlation with the irrigation decision. Electrical conductivity
1 ∑− 1
AccuracyCLASS (y, ŷ) = 1(ŷ = y) (11) had a lower correlation than soil moisture in the same direction, so it
nvalues i=0 was not selected for the final dataset. Although temperature was less
correlated than soil moisture, it was also included because the correla­
The resulting decision regions of each model were evaluated by using
tion was in the opposite direction. Among the environmental charac­
the mlxtend library to visualize the performance achieved. Visualizing
teristics, temperature and relative humidity were included because they
these regions truly shows the complexity of the model, highlights
had weak correlation and were in the opposite direction. Since humidity
overfitting, and gives an indication of the prediction confidence level.
was consistent in direction with a weaker relationship, precipitation and
evapotranspiration were the variables processed for weather prediction.

7
L. Preite and G. Vignali Computers and Electronics in Agriculture 223 (2024) 109126

Fig. 4. a) Pearson rank correlation of the collected data. b) Spearman rank correlation of the collected data.

3.2. Processing node: Irrigation predictive algorithms 3.2.4. Algorithm’s comparison

Confusion matrices and metrics are used to assess the performance in
3.2.1. K-nearest neighbors’ classifier predicting field data. Overall, the three models were characterized by
The cross-validation process allows the hyperparameters to be high accuracy, with the highest value of 99,61 % achieved by the arti­
maximized. Specifically, the model was trained by considering 13-near­ ficial neural network. This was followed by the k-nearest neighbors and
est neighbors and the weights are handled by the distance function. support vector machine with 99,46 % and 99,20 %, respectively.
True positives, false positives, false negatives, and true negatives are
3.2.2. Support vector machine classifier reported in the confusion matrices in Fig. 5 (a-b-c). In addition, the Fig. 5
Kernel radial basis function was implemented for the training phase. (d-e-f) highlights the 2-D decision regions according by soil moisture
The regulation index C and the kernel function parameter γ were set to content and rainfall forecast amount, taking into account a mean value
1000 and 0.1, respectively. for the other features involved.
The comparison between the classifiers based on the calculated
3.2.3. Artificial neural network metrics is reported in Table 3.
Artificial neural network was employed by considering the three
hidden layer configuration resulting from the cross-validation step. 3.2.5. Water and energy savings
Hyperbolic tangent was the corresponding activation function. In this The water savings confidence interval resulting from the Monte Carlo
framework, Adam was the algorithms used with a constant learning rate simulation ranges from 14.5 % to 27.6 %. The lower and upper bounds
and a regulation index of 0.01. for energy savings are 49.2 % and 57 %, respectively. As mentioned

Fig. 5. a-b-c) Confusion matrices for the implemented machine learning algorithms. d-e-f) decision region boundaries for the implemented machine
learning algorithms.

8
L. Preite and G. Vignali Computers and Electronics in Agriculture 223 (2024) 109126

Table 3
Classification metrics for the implemented machine learning algorithms.
Precision Recall F1-score Samples
KNN SVM ANN KNN SVM ANN KNN SVM ANN

OFF 0,9961 0,9757 1 0,9822 0,9918 0,985 0,9891 0,9837 0,992

ON 0,9993 0,9958 0,9982 0,9975 0,9989 0,9993 0,9984 0,9974 0,9988
No Adjustment 0,9830 0,9929 0,9886 0,9959 0,9781 0,9959 0,9894 0,9854 0,9922
Alert 1 1 1 1 0,8889 1 1 0,9412 1
macro avg 0,9946 0,9911 0,9967 0,9939 0,9644 0,9951 0,9942 0,9769 0,9959 4857
weighted avg 0,9947 0,9920 0,9961 0,9946 0,9920 0,9961 0,9947 0,9920 0,9961 4857

earlier, these intervals cover 95 % of the estimated values. savings while side it ranges from 49.2 % to 57 % for energy savings.
This framework provides a cost-effective solution, as the authors of
4. Discussion (Stefanini et al., 2023) assess the economic viability of the data acqui­
sition layer implemented and state that almost 2 years are required to
The data collected from the developed living lab were used to pro­ recoup the initial investment. The data collected by this layer are then
vide a three-layer predictive algorithms-based irrigation system. Spe­ aggregated with the weather forecast data collected by an open-source
cifically, information on soil and environmental properties were gained platform. The main key of the developed framework is the soil capac­
on 2023 tomato crop season by implementing an IoT LoRaWAN-based ity evaluation. Once it is calculated, the dataset was adapted to it and the
network, which transmits data every 10 min to the network server. algorithm can make new predictions based on the new operating con­
These data were stored in database and processed in combination with ditions. In this perspective, there are also some limitations. A wide range
the weather forecasts provided by an open-source API based on inter­ of soil types and thus different soil capacity level must be tested to assess
national recognized weather models. the robustness of the model and the crop status data should also be
The results show how the developed models achieve significant aggregated to affine the model and account for the real water needs of
performance in predicting irrigation status. All the tested algorithms the crop at different growth stages.
achieved comparable accuracy, close to 99 %. A classification problem
characterizes the irrigation framework, which avoids overwatering by 5. Conclusion
defining the soil capacity as thresholds to manage the water delivery.
The outputs are four classes, i.e., OFF, ON, No Adjustment, Alert. The agricultural sector accounts for almost 70 % of global water
In order to train and validate the models, the data were arranged withdrawals, recording one of the major impacts on water scarcity issue.
according to the correlation analyses results. Thereby, the Pearson and Several efforts reported in the scientific literature have proven how the
Spearman correlations yielded consistent results to determine the right 4.0 scenario is a powerful solution to improve agricultural sustainability
combination of parameters to avoid overfitting and maximize the ac­ and minimize water withdrawals. Therefore, there are several cutting-
curacy. A final dataset of 16,187 samples and 6 features was defined by edge solutions, including blockchain, internet of things, digital twins,
considering soil water content and temperature, ambient temperature, deep learning and machine learning algorithms, and other computer
and relative humidity as real-time attributes; precipitation amount and applications, for reducing wastewater and improving irrigation
evapotranspiration reference rate for the next three days were included schedules.
as forecast features. The resulting dataset was split 70/30 into training The proposed work aims at analyzing the application of predictive
and test data. The training was performed by determining the optimal algorithms in agriculture for irrigation purposes. Specifically, the sci­
value of the hyperparameters with cross validation. Overall, the nearest entific literature highlights how machine learning techniques are widely
neighbor of 13, the weight distance function, the high regularization applied in various disciplines because they offer highly scalable solu­
term C, the resulting neural network structure, and the low value of the tions capable of automating processes based on predictions. In addition,
regularization parameter alpha reveal how the noisiness of the data was the scientific literature describes how the proposed applications in
not significant. agriculture focus on three different levels, i.e., pre-harvest, harvest, and
Fig. 5 provide a powerful overview on the classification results. post-harvest. In this perspective, the first level aims to maximize the
Decision region visualizations and confusion matrices highlight how effectiveness of the growing conditions, which also include irrigation
there are significant differences between the models at nearly the same management.
accuracy. Indeed, the multi-layer perceptron neural network define In the proposed framework an artificial intelligence application
reliable decision boundary to associate the right class to unknown data, based on a three-layer architecture (i.e., data acquisition, edge
meanwhile the k-nearest function and support vector machine fail in computing, and cloud layer) has been investigated to address water
differentiating some cases. This issue is also described by the evaluated scarcity. The main purpose is to improve the irrigation schedules to
metrics reporting in Table 3. In this perspective, a higher amount of false avoid overwatering, which causes percolation of the water from the soil
positive and false negative was reached for the k-nearest neighbors and to the lower layers when the water content exceeds than the soil ca­
support vector machine. Moreover, the warning alarm was correctly pacity. In this perspective, the water moves back to the groundwater
predicted by both neural network and k-nearest neighbors, while two table, but it cannot be used directly for the plant growth. The soil ca­
false positives resulted from the support vector machine model. pacity point has been experimentally assessed by defining a Gaussian
Digging deeper into the results, the algorithm is used as a processing distribution of the soil moisture having given mean and standard devi­
node to make predictions on the field data every 10 min. The main aim ation values, which are evaluated by calculating the time derivative of
of the performed classification problem is to reduce the water con­ the soil moisture to identify the area where the rate of decline is slower
sumption by avoiding overwatering during irrigations. Thereby, the after an irrigation or a precipitation event. As a result, an irrigation
scientific literature highlights how a percolation phenomenon occurs confidence interval has been implemented to create four different clas­
when the water content exceeds the soil capacity point. A potential ses (i.e., irrigation OFF, irrigation ON, no adjustment, and alert) that are
water and energy savings are estimated by comparing these preliminary used to set up a multi-classification problem.
results with the irrigation volume recommended by IRRIFRAME. To address this issue, a dedicated living lab monitored the growing
Therefore, the results of the Monte Carlo simulation outline how the conditions of a tomato crop throughout the whole season. The core part
resulting confidence interval ranges from 14.5 % to 27.6 % for water of the data acquisition layer is an IoT network, characterized by several

9
L. Preite and G. Vignali Computers and Electronics in Agriculture 223 (2024) 109126

end nodes installed in the field, which collect information on environ­ MISSIONE 4 COMPONENTE 2, INVESTIMENTO 1.4 – D.D. 1032 17/06/
mental and soil conditions every 10 min. The collected data are trans­ 2022, CN00000022). This manuscript reflects only the authors’ views
mitted via LoRa communication protocol to a gateway, which exploits and opinions, neither the European Union nor the European Commission
an internet protocol to communicate with the network server. At this can be considered responsible for them.
point, the data can be stored into databases and processed by scrubbing
for errors, duplicates, or missing data, and finally aggregated with the References
three-day weather forecast data provided through an open-source API.
The resulting dataset, characterized by 16,187 samples and 12 features Adnan, R.M., Mostafa, R.R., Reza, A., Islam, M.T., Kisi, O., Kuriqi, A., Heddam, S., 2021.
Estimating reference evapotranspiration using hybrid adaptive fuzzy inferencing
(i.e., irrigation decision, soil moisture, soil temperature, soil electrical coupled with heuristic algorithms. Comput. Electron. Agric. 191, 106541 https://
conductivity, daily hour, ambient temperature, ambient humidity, doi.org/10.1016/j.compag.2021.106541.
rainfall, environmental humidity forecasted, and evapotranspiration), Alibabaei, K., Gaspar, P.D., Lima, T.M., 2021. Modeling Soil Water Content and
Reference Evapotranspiration from Climate Data Using Deep Learning Method. Appl.
has been used to gain insights into features correlation by means of Sci. 11 (11), 5029. https://doi.org/10.3390/app11115029.
Pearson and Spearman correlation coefficients. Following this path, Aly, M. S., Saad, •, Darwish, M., & Aly, A. A. (2024). High performance machine learning
redundant parameters are removed to create a consistent dataset for approach for reference evapotranspiration estimation. Stochastic Environmental
Research and Risk Management doi: 10.1007/s00477-023-02594-y.
training and testing the machine learning algorithms. Awad, M., Khanna, R., 2015. Support Vector Machines for Classification. In: Efficient
Deepening into the results, three different classifiers, i.e., multi-layer Learning Machines. Apress, Berkeley, CA. https://doi.org/10.1007/978-1-4302-
perceptron neural network, support vector machine, and k-neighbors 5990-9_3.
Bottazzi, M., Scipione, G., Marras, G.F., Trotta, G., D’Antonio, M., Chiavarini, B.,
classifier have been trained and tested by splitting the resulting dataset
Caroli, C., Montanari, M., Bassini, S., Gascón, E., Hewson, T., Montani, A., Cesari, D.,
into 70 % and 30 %, respectively. Prediction performances have been Minguzzi, E., Paccagnella, T., Pelosini, R., Bertolotto, P., Monaco, L., Forconi, M.,
assessed by evaluating boundary decision regions, confusion matrices, Pieralice, A., 2021. The Italian open data meteorological portal: MISTRAL. Meteorol.
and various metrics, such as accuracy, precision, recall, F1-score, macro- Appl. 28 (4) https://doi.org/10.1002/met.2004.
Chandrappa, V.Y., Ray, B., Ashwatha, N., Shrestha, P., 2022. Spatiotemporal modeling to
average precision, and weighted-average precision. Despite on the three predict soil moisture for sustainable smart irrigation. Internet of Things 21, 100671.
classifiers work with high comparable accuracy (close to 99 %), multi- https://doi.org/10.1016/j.iot.2022.100671.
layer perceptron neural network is the more suitable solution because Corell, M., Pérez-López, D., Martín-Palomo, M.J., Centeno, A., Girón, I., Galindo, A.,
Moreno, M.M., Moreno, C., Memmi, H., Torrecillas, A., Moreno, F., Moriana, A.,
it minimizes false predictions as it is shown by the confusion matrix and 2016. Comparison of the water potential baseline in different locations. Usefulness
revealed by the boundary decision regions. The highest levels of macro for irrigation scheduling of olive orchards. Agric Water Manag 177, 308–316.
and weighted average precision also support this finding. https://doi.org/10.1016/J.AGWAT.2016.08.017.
Custódio, G., Prati, R.C., 2024. Comparing modern and traditional modeling methods for
As the final step, the performance of the proposed framework has predicting soil moisture in IoT-based irrigation systems. Smart Agricultural
also been investigated by evaluating the resulting water and energy Technology 7, 100397. https://doi.org/10.1016/J.ATECH.2024.100397.
saving. Therefore, a Monte Carlo simulation shows that water savings of Food and Agriculture Organization of the United Nations. (2023). 2050: A third more
mouths to feed. Retrieved from: https://www.fao.org/news/story/en/item/35571/
up to 27.6 % and energy savings of up to 57 % can be achieved compared icode/.
to a conventional tomato season. Giusti, E., Marsili-Libelli, S., 2015. A Fuzzy Decision Support System for irrigation and
Summing up the results, the predictive algorithm-based irrigation water conservation in agriculture. Environ. Model. Softw. 63, 73–86. https://doi.
org/10.1016/J.ENVSOFT.2014.09.020.
system is a cost-effective solution to reduce water withdrawals by
Gu, Z., Zhu, T., Jiao, X., Xu, J., Qi, Z., 2021. Neural network soil moisture model for
avoiding soil percolation phenomenon. It can be scaled to other crops, as irrigation scheduling. Comput. Electron. Agric. 180, 105801 https://doi.org/
the core part of the system is the reliable determination of the soil ca­ 10.1016/J.COMPAG.2020.105801.
pacity point, which affects the confidence interval implemented in the Hemalatha, K., Usha, R.K., 2017. Advancements in Multi-Layer Perceptron Training to
Improve Classification Accuracy. Int. J. Recent and Innovation Trends in Comput.
preprocessing stage. Future research should further investigate the Commun. 5 (6) https://doi.org/10.17762/ijritcc.v5i6.776.
robustness of the model by evaluating a wide range of soil types and Irriframe, 2023. Irrinet Canale Emiliano Romagnolo. Retrieved from https://www.
aggregating crop data, such as chlorophyll content or sap ion concen­ irriframe.it/irriframe/home/index_er.
Kavya, M., Mathew, A., Shekar, P.R., P, S., 2023. Short term water demand forecast
tration to also assess plant drought stress. modelling using artificial intelligence for smart water management. Sustain. Cities
Soc. 95, 104610 https://doi.org/10.1016/J.SCS.2023.104610.
CRediT authorship contribution statement Kramer, O., 2013. K-Nearest Neighbors. In: Dimensionality Reduction with Unsupervised
Nearest Neighbors. Intelligent Systems Reference Library, vol 51. Springer, Berlin,
Heidelberg. doi: 10.1007/978-3-642-38652-7_2.
Luca Preite: Conceptualization, Methodology, Validation, Writing – Kuri-Morales, A., 2017. Closed determination of the number of neurons in the hidden
review & editing, Investigation, Formal analysis, Visualization, Re­ layer of a multi-layered perceptron network. Soft Computing, 21(3), pp. 597–609. doi:
10.1007/s00500-016-2416-3.
sources, Software, Writing – original draft. Giuseppe Vignali: Writing – Liakos, K., Busato, P., Moshou, D., Pearson, S., Bochtis, D., 2018. Machine Learning in
review & editing, Visualization, Validation, Supervision, Project Agriculture: A Review. Sensors 18 (8), 2674. https://doi.org/10.3390/s18082674.
administration, Methodology, Investigation, Funding acquisition, Mazzei, D., & Ramjattan, R., 2022. Machine Learning for Industry 4.0: A Systematic
Review Using Deep Learning-Based Topic Modelling. In Sensors,22, 22. MDPI. doi:
Formal analysis, Conceptualization.
10.3390/s22228641.
MClimate, 2023. Retrieved from https://mclimate.eu.
Declaration of competing interest Meshram, V., Patil, K., Meshram, V., Hanchate, D., Ramkteke, S.D., 2021. Machine
learning in agriculture domain: A state-of-art survey. Artificial Intelligence in the
Life Sciences 1, 100010. https://doi.org/10.1016/J.AILSCI.2021.100010.
The authors declare that they have no known competing financial Mokhtar, A., Al-Ansari, N., Wessam El-Ssawy, ⋅, Graf, R., Aghelpour, P., He, H., Salma, ⋅,
interests or personal relationships that could have appeared to influence Hafez, M., & Abuarab, M., 123 C.E. Prediction of Irrigation Water Requirements for
the work reported in this paper. Green Beans-Based Machine Learning Algorithm Models in Arid Region. Water
Resource Management, 37, pp. 1557-1580. doi: 10.1007/s11269-023-03443-x.
Navarro-Hellín, H., Martínez-del-Rincon, J., Domingo-Miguel, R., Soto-Valles, F., Torres-
Data availability Sánchez, R., 2016. A decision support system for managing irrigation in agriculture.
Comput. Electron. Agric. 124, 121–131. https://doi.org/10.1016/J.
COMPAG.2016.04.003.
Data will be made available on request. Open Meteo, 2024. Weather Forecast API. Retrieved from open-meteo.com: https://open-
meteo.com/en/docs.
Acknowledgements Pereira, L.S., Allen, R.G., Smith, M., Raes, D., 2015. Crop evapotranspiration estimation
with FAO56: Past and future. Agric Water Manag 147, 4–20. https://doi.org/
10.1016/J.AGWAT.2014.07.031.
This study was carried out within the Agritech National Research Považanová, B., Čistý, M., Bajtek, Z., 2023. Using feature engineering and machine
Center and received funding from the European Union Next-Generation learning in FAO reference evapotranspiration estimation. J. Hydrol. Hydromech 71,
425–438. https://doi.org/10.2478/johh-2023-0032.
EU (PIANO NAZIONALE DI RIPRESA E RESILIENZA (PNRR) –

10
L. Preite and G. Vignali Computers and Electronics in Agriculture 223 (2024) 109126

Preite, L., Solari, F., Vignali, G., 2023a. A digital model application to optimize water United Nations, 2024. Integrated Monitoring Initiative for SDG 6. Retrieved from: United
consumption in agriculture. Proceedings of the International Food Operations and Nations UnWater.
Processing Simulation Workshop, FOODOPS, 2023-September. doi: 10.46354/ Warrick, A.W. (Ed.), 2001. Soil Physics Companion, (1st ed.). CRC Press. https://doi.org/
i3m.2023.foodops.006. 10.1201/9781420041651.
Preite, L., Solari, F., & Vignali, G., 2023b. Technologies to Optimize the Water Yong, S.L.S., Ng, J.L., Huang, Y.F., Ang, C.K., Ahmad Kamal, N., Mirzaei, M., Najah
Consumption in Agriculture: A Systematic Review. In Sustainability (Switzerland),15, Ahmed, A., 2024. Enhanced Daily Reference Evapotranspiration Estimation Using
7. MDPI. doi: 10.3390/su15075975. Optimized Hybrid Support Vector Regression Models. Water Resour. Manag. https://
Sharma, A., Jain, A., Gupta, P., & Chowdary, V., 2021. Machine Learning Applications doi.org/10.1007/s11269-024-03860-6.
for Precision Agriculture: A Comprehensive Review. In IEEE Access,9, pp. Youssef, M.A., Peters, R.T., El-Shirbeny, M., Abd-ElGawad, A.M., Rashad, Y.M.,
4843–4873. Institute of Electrical and Electronics Engineers Inc. doi: 10.1109/ Hafez, M., Arafa, Y., 2024. Enhancing irrigation water management based on ETo
ACCESS.2020.3048415. prediction using machine learning to mitigate climate change. Cogent Food and
Skovdal Christiansen, J., Thorsen, M., Clausen, T., Hansen, S., Christian Refsgaard, J., Agriculture 10 (1), 2348697. https://doi.org/10.1080/23311932.2024.2348697.
2004. Modelling of macropore flow and transport processes at catchment scale. Zhou, Z., Majeed, Y., Diverres Naranjo, G., Gambacorta, E.M.T., 2021. Assessment for
J. Hydrol. 299 (1–2), 136–158. https://doi.org/10.1016/J.JHYDROL.2004.04.029. crop water stress with infrared thermal imagery in precision agriculture: A review
Srivastava, S., Kumar, N., Malakar, A., Sruti, ⋅, Choudhury, D., Chittaranjan Ray, ⋅, & and future prospects for deep learning applications. Comput. Electron. Agric. 182,
Roy, T., 2024. A Machine Learning-Based Probabilistic Approach for Irrigation 106019 https://doi.org/10.1016/J.COMPAG.2021.106019.
Scheduling. Water Resource Management doi: 10.1007/s11269-024-03746-7. Zio, E., 2013. Monte Carlo Simulation: The Method. In: The Monte Carlo Simulation
Stefanini, R., Preite, L., Bottani, E., Belli, L., Davoli, L., Ferrari, G., & Vignali, G., 2023. Method for System Reliability and Risk Analysis. Springer Series in Reliability
Selection of 4.0 sensors for small holders: the compromise between the advantages Engineering. Springer, London. doi: 10.1007/978-1-4471-4588-2_3.
and the costs of the implementation. Proceedings of the International Food Operations Zou, J., Han, Y., So, S.S., 2008. Overview of Artificial Neural Networks. In: Livingstone,
and Processing Simulation Workshop, FOODOPS, 2023-September. doi: 10.46354/ D.J. (eds) Artificial Neural Networks. Methods in Molecular Biology™, 458. Humana
i3m.2023.foodops.007. Press. doi: 10.1007/978-1-60327-101-1_2.

11