sensors

Article
Sensors Driven AI-Based Agriculture
Recommendation Model for Assessing
Land Suitability
Durai Raj Vincent 1 , N Deepa 1 , Dhivya Elavarasan 1 , Kathiravan Srinivasan 1, * ,
Sajjad Hussain Chauhdary 2, * and Celestine Iwendi 3
1 School of Information Technology and Engineering, Vellore Institute of Technology,
Vellore 632 014, Tamil Nadu, India
2 College of Computer Science and Engineering, University of Jeddah, Jeddah 21577, Saudi Arabia
3 Department of Electronics, BCC of Central South University of Forestry and Technology,
Changsha 410004, China
* Correspondence: [email protected] (K.S.); [email protected] (S.H.C.)




Received: 18 July 2019; Accepted: 17 August 2019; Published: 23 August 2019


Abstract: The world population is expected to grow by another two billion in 2050, according to
the survey taken by the Food and Agriculture Organization, while the arable area is likely to grow
only by 5%. Therefore, smart and efficient farming techniques are necessary to improve agriculture
productivity. Agriculture land suitability assessment is one of the essential tools for agriculture
development. Several new technologies and innovations are being implemented in agriculture as
an alternative to collect and process farm information. The rapid development of wireless sensor
networks has triggered the design of low-cost and small sensor devices with the Internet of Things
(IoT) empowered as a feasible tool for automating and decision-making in the domain of agriculture.
This research proposes an expert system by integrating sensor networks with Artificial Intelligence
systems such as neural networks and Multi-Layer Perceptron (MLP) for the assessment of agriculture
land suitability. This proposed system will help the farmers to assess the agriculture land for
cultivation in terms of four decision classes, namely more suitable, suitable, moderately suitable,
and unsuitable. This assessment is determined based on the input collected from the various sensor
devices, which are used for training the system. The results obtained using MLP with four hidden
layers is found to be effective for the multiclass classification system when compared to the other
existing model. This trained model will be used for evaluating future assessments and classifying the
land after every cultivation.

Keywords: smart agriculture; multi-layer perceptron; agricultural data; IoT in agriculture; land
suitability using sensors; sensor data in agriculture

1. Introduction
Agriculture farming is considered as the base for human living because it is the primary source
of food and income for most of the countries in the world. The economy of the country depends
on agriculture production as it provides food, raw materials, employment, etc., to the people living
in that country [1]. It has been observed in recent times that there is no significant development in
crop production in the agriculture sector. Also, there is a rapid increase in the price of food because
the production of the crop is not meeting demand [2]. One of the causes for the decrease in crop
production is the farmers use of the traditional way of cultivation, which leads to less crop yield.
The farmers new to the agriculture field have insufficient knowledge about the characteristics of soil

Sensors 2019, 19, 3667; doi:10.3390/s19173667 www.mdpi.com/journal/sensors

Sensors 2019, 19, 3667 2 of 16

for crop cultivation [3]. They are not conscious of the fact that agriculture land needs to be assessed
before cultivation [4].
Land suitability analysis is a mandatory prerequisite for crop cultivation, which helps to obtain
maximum production. In order to acquire the properties of the soil, farmers depend on soil testing
labs, which are not sufficient enough to help them, and sometimes the data provided by the labs are
inaccurate [5]. In order to get enough knowledge about cultivation, data need to be collected manually,
which is very difficult for farmers. The solution is the replacement of traditional methods of data
collection with Internet of Things (IoT)-based sensors [6]. Sensors play a significant role in collecting
information about various factors such as soil, water, climate, etc., for agriculture development. With
the help of data gathered from different sensors, land suitability analysis could be done, which would
help farmers identify the current status of their agriculture land and improve their crop production.
Several decision models have been developed so far to assist the farmers in taking a decision on crop
cultivation to maximize their profit [7].
Ever-changing real-world conditions like monsoons to temperate are the dominant factors in
affecting agriculture productivity. Soil parameters to weather prediction are essential in knowing
the future yield [8]. The key factors that are influencing the crop yield are climatic conditions, soil
productivity, and groundwater characteristics, as well as its availability. An extensive study on
phenotyping is helping researchers to understand these factors [9]. One of the significant challenges
involved in identifying the different land conditions is based on the influence of various soil parameters.
Over the past few decades, soil testing has been practised, which is widely accepted by agriculture
experts and farmers to determine the properties of soil for agriculture production [10]. The Agriculture
Management Information System (AMIS) is mandatory to store the data acquired in various formats
and exchanges the data for digital agriculture technology [11]. Currently, AMIS has enhanced its
facilities by accumulating the latest technologies, such as the Internet and the Internet of Things (IoT).
Introduction of the usage of sensors actually reduces the cost and time involved in assessing the
land suitability in the traditional manner [12]. The uses of sensors are plenty, and it is possible to use
as many sensors in the field of agriculture. Soil sensors, water sensors, and biosensors are few that
have been shown to have a significant role in measuring nature. These sensors contribute to the smart
farming system, especially in the handling of appropriate irrigation systems to help farmers. The Radio
Frequency Identification (RFID) technology is used for the detection of animals entering into the
farmland. The Global Positioning System (GPS) position-based seeding recommendation is emerging
as another benefit in the field of agriculture benefiting from the latest technological developments [13].
Some of the benefits it has achieved in recent days are improved efficiency, sustainable growth, as well
as cost-effectiveness. It is considered to be the connection between the collection of physical devices
through internet connection such as WiFi, Data cards, etc.
As a revolutionary change, the sensor-driven network has a considerable role in precision
agriculture [14]. Timely forecast based on soil parameters will ensure proper irrigation systems. IoT
has a significant role in handling the recorded data from sensors. General Purpose Input/Output
facilitates various sensor inputs and the corresponding outputs in Raspberry Pi. Further, this could
really help in hosting different sensors. Cloud computing is helping the IoT sensor networks to record
the measured data through its sensors. The wireless medium is transferring the data to the cloud,
which is used later for data analysis [15].
Data Analytics has brought impactful results for future predictions in almost every single
application. Artificial intelligence-based model building is a challenging task as the model should
replicate the observed parameters in the dataset. Parametric estimations are helpful in solving the
problems in a futuristic way, and agriculture problems are among those [16]. Agriculture data analysis
is done with different machine learning algorithms. Maximizing the outcome is the agenda for any
machine learning model, and model assessment metrics are helpful in analyzing the results obtained.
Appropriate training of the machine learning model will produce the results with utmost accuracy.
Sensors 2019, 19, 3667 3 of 16

Neural network-based models are used widely, as they have the capability to improve the efficiency
of the model over every iteration [17]. Classification problems are handled by various algorithms,
such as support vector machines, decision tree algorithms, etc. Multiclass classifications are expertly
handled by neural network algorithms [18]. When the model is trained with the appropriate number
of inputs and outputs, it is expected to give better results compared to other machine learning models.
The integration of IoT, along with machine learning models, are providing the farmer recommendation
system with appropriate inputs [19].
In this work, an agriculture dataset has been obtained through various IoT sensor devices such as
a pH sensor, soil moisture sensor, salinity sensor, and electromagnetic sensor. A sensor is a device
which is used to detect and respond to some type of input obtained from the physical environment [20].
As it requires less labor and consumes less time, it is used in many real-time applications. As the
Internet plays a mediatory role in various kinds of communication and data exchange, it is advisable to
integrate the agriculture data with the cloud platform. The data acquired from the various IoT devices
could be stored to the cloud platform [21].
The key contribution of this work is as follows

• The selection of the relevant attributes and the collection of data from the appropriate sources for
better agriculture land suitability classification, and these data are heterogeneous in nature.
• Integration of different data collection and bringing it to a form suitable for future predictions.
• The noble initiation towards helping the farmers build a recommendation model, especially for
the low economic regions, in a country like India.
• Generalization of the proposed model, which could be used for any land, as the current test inputs
are sufficient for the classification of the land suitability with this developed model.
• The multiclass classification is another crucial contribution as the majority of the past work is
based on binary classification, which includes two decision classes, namely suitable and unsuitable.
These multiclass results are helpful for the farmers when compared to the binary classification, to
take suitable action based on the outcome of the results.

This paper presents a model that will automate the process of analyzing various deciding factors
in the assessment of present and future conditions for the better crop yield. The different classification
will help decide well in advance how to proceed to start a new session. Besides, this would ensure
minimal loss to the farmers, and the different sections presented here elucidate the model further.

2. Related Work
The integration of the Internet of Things into the agricultural system has led to the Internet of
Agriculture of Things along with advanced computing techniques. The researchers apply this to
obtain maximum benefit and to also improve the production of agriculture, artificial intelligence, and
IoT [22]. The agriculture domain is experiencing new evolution and revolution motivated by sensors,
IoT, big data, and cloud technology [23]. The IoT based on a smart agriculture system was developed
using deep reinforcement learning integrated with a cloud environment, in which four layers were
included, namely data collection, edge computing, data transmission, and cloud computing [24]. IoT
has been integrated with agriculture systems to yield maximum profit, and the applications of IoT in
agriculture have been categorized into restricted environment planning, open farm planting, livestock
farm monitoring, and aquaculture development [25].
A new proposal was presented for agriculture applications by investigating various integrated
platforms, which include cloud computing, IoT, and data mining techniques [26]. A scalable
network-based architecture was proposed to monitor and regulate agriculture farms in rural areas
with the help of IoT-based WiFi with long-distance network and fog computing [27]. The development
of IoT in agriculture data analytics has transformed from a specific crop to any kind of crop. Also,
the system is able to support a various number of applications from controlling and supervising the
agriculture crops to promoting the products in the market [28]. An irrigation monitoring and control
Sensors 2019, 19, 3667 4 of 16

system was developed using a fuzzy approach to generate a moisture content distribution map of soil
by applying an interpolation technique.
The IoT-based Wireless Sensor Network (WSN) was used to generate a distribution map with the
measurements of soil moisture content taken at nine places of the selected farmland [29]. An IoT-based
system was developed to water agriculture crops using a wireless sensor network, and a web application
was also designed to control the information about the crop and field [30]. Several machine learning
algorithms have been used so far for prediction and forecasting. A new forecasting model was designed
by combining the gravitational search algorithm with Multilayer Perceptron (MLP) to forecast the level
of water in Winnipesaukee and Cypress Lakes in the USA [31]. A neural network-based model was
proposed using MLP to predict the bid price of the keyword in a particular search engine ranking [32].
A hybrid model was proposed by combining the firefly algorithm and MLP to predict the salinity value
of soil with the help of remote sensing data [33]. An algorithm was proposed for the identification of
weed diseases using MLP and automatic relevance determination. Three multi-criteria yield models
were proposed using neural networks with MLP for the prediction of winter rapeseed [34].
An IoT-based hydroponic intelligent system was developed using deep neural networks to control
the actions in the hydroponic environment characterized by multiple parameters gathered using IoT
sensors [35]. An IoT-based smart decision model was developed to help the farmers to get information
such as fertilizer requirements, crop analysis, and market requirements. IoT-based sensors were used to
obtain the data related to farming, and it is stored in a cloud framework where information is processed
and sent to farmers’ mobile devices by applying some data mining techniques [36]. An Intelligent
model was developed for the diagnosis of faults in the aquaculture ponds located in remote places
through IoT sensors. A fuzzy neural network algorithm was applied for fault detection, and the model
detects many types of faults in aquaculture ponds [37]. An IoT-based prototype was designed for
monitoring the water level in sources like borewells, water tanks, etc. This prototype could be used to
avoid damage in pump motors when the water level goes beyond the threshold level [38].
A predictive platform for precision agriculture was presented using IoT devices to improve
crop productivity at low-cost. Deep learning algorithms were used to predict the diseases in plants
which were hidden in the images of the leaves, and they were classified with maximum accuracy [39].
An intelligent system was developed to predict the fruit-melon image of skin lesions and alert for new
planting environments using the convolutional neural network and deep learning rule. The image is
obtained by using an infrared video sensor, and the system has obtained a satisfactory accuracy of
97.5%. A modern automatic agriculture monitoring system was developed using the ZigBee Wireless
Sensor Network and deep learning algorithms. Wireless sensor networks are also used to monitor the
temperature of soil and humidity. Also, the system classifies the moisture content in the soil.
A decision tree-based model was proposed for optimizing water and energy consumption [40].
Further, this was accomplished with the assistance of the Internet-of-things. Moreover, to help the
farmers in saving the precious water resource, which is declining day by day, several soil parameters
were considered. Besides, for measuring the water requirement, and factors such as soil moisture
content and soil temperature were utilized for making a suitable decisions concerning water release.
The cloud-based IoT sensors are helpful in this work to control the overall setup. This in turn also
saves the energy.
The below section presents the proposed system which is a recommendation model for using
artificial intelligence and sensors.

3. Proposed System—Sensors Driven Artificial Intelligence (AI) - Based Agriculture

Recommendation Model
The sensor network-based proposed model is discussed in this section. The architecture diagram
of the sensor-based data management is shown in Figure 1. Sensor-based data collection requires three
main steps: data acquisition, data communication, and data processing. For collecting the various
parameter values concerning the properties of soil suitable for agriculture development, various
Sensors 2019, 19, 3667 5 of 16

sensors are used [41]. The data acquisition is made using various sensor devices such as the pH sensor,
soil moisture
Sensors sensor,
2019, 19, x salinity
FOR PEER REVIEWsensor, and an electromagnetic sensor. 5 of 16

Figure
Figure 1. Architecture diagram
1. Architecture diagram of
of the
the proposed
proposed model.
model.

The
The raspberry
raspberry Pi Pi 33 system
system is is used
used here
here toto handle
handle inputs
inputs fromfrom multiple
multiple sensors,
sensors, and and the
the data
data are
are
sent
sent to a cloud for storage, since it has the most powerful CPU comparatively, as well as the802.11
to a cloud for storage, since it has the most powerful CPU comparatively, as well as the IEEE IEEE
wireless standard.
802.11 wireless A wi-fiAfacility
standard. is also available
wi-fi facility and is used
is also available and is toused
transfer the datathe
to transfer fromdatathefrom
remote
the
agriculture land. For
remote agriculture the For
land. better
thehandling of data, of
better handling thedata,
datatheis sent
datafurther
is sent to the cloud
further to thewith
cloudthewith
helpthe
of
the internet. The cloud facility used here is Amazon Web Serive (AWS),
help of the internet. The cloud facility used here is Amazon Web Serive (AWS), and the stored data and the stored data is used
for machine
is used learninglearning
for machine for the purpose of analysis.
for the purpose The dataThe
of analysis. is accessed on the local
data is accessed machine
on the through
local machine
athrough
cloud facility.
a cloudThe algorithm
facility. is developed
The algorithm in the machine
is developed and it isand
in the machine testedit ison the collected
tested data to
on the collected
verify
data to verify the accuracy of the results obtained. The Raspberry Pi controller is used to collectfrom
the accuracy of the results obtained. The Raspberry Pi controller is used to collect the data the
various sensors for a sampling period of one day. Then the average
data from various sensors for a sampling period of one day. Then the average values of various values of various sensors are
moved
sensorsto arethe AWS to
moved Cloud
the AWSNetworkCloud with the help
Network of the
with the Internet.
help of the From the AWS,
Internet. Fromthe thedata
AWS, is accessed
the data
by the proposed model, which is developed in the system for normalization
is accessed by the proposed model, which is developed in the system for normalization and training and training purposes.
Thus, the data
purposes. obtained
Thus, the data fromobtained
various sensors for six months
from various sensorsare forconsidered
six months forarethe considered
developmentfor of the
the
sensor-driven AI model.
development of the sensor-driven AI model.
The
The various
various parameters
parameters used used to to find
find agriculture
agriculture landland suitability
suitability analysis
analysis in in this
this work
work are are soil
soil
texture,
texture, granular
granular fragments
fragments (percentage
(percentage of of sand
sand particles
particles in in the
the soil),
soil), structure
structure of of the
the soil,
soil, compact
compact
and
and cementation,
cementation, cnternal
cnternal drainage,
drainage, available
available water
water content,
content, porousness,
porousness, organic organic matter,
matter, cation
cation
exchange capacity, degree saturation, pH value, salinity, and carbonates.
exchange capacity, degree saturation, pH value, salinity, and carbonates. Moreover, this is the Moreover, this is the collected
information which is which
collected information essential to maketoproper
is essential classification
make proper of land
classification ofsuitability.
land suitability. These inputs
These are
inputs
collected from sensors, as well as from the anonymous farmers and agriculture
are collected from sensors, as well as from the anonymous farmers and agriculture experts. Their experts. Their collective
responses and inputsand
collective responses helped
inputsto construct
helped tothe required
construct thedataset.
required dataset.
Electromagnetic
Electromagnetic sensors have been used to obtain the
sensors have been used to obtain the measurements
measurements for for soil
soil texture,
texture, internal
internal
drainage, available water content, organic matter, cation exchange capacity,
drainage, available water content, organic matter, cation exchange capacity, carbonates, and degree carbonates, and degree of
saturation.
of saturation. pHpH sensors
sensorsare used to obtain
are used the pH
to obtain thevalue, and salinity
pH value, sensorssensors
and salinity are used aretoused
get theto salinity
get the
salinity value. The porousness value is obtained based on the measurement of the soil moisture
sensor and granularity of the soil is obtained using the measurement acquired from the moisture
content of the soil. The values of the structure of the soil and the compact and cementation are
obtained from the agriculture experts.
Sensors 2019, 19, 3667 6 of 16

value. The porousness value is obtained based on the measurement of the soil moisture sensor and
granularity of the soil is obtained using the measurement acquired from the moisture content of the
soil. The values of the structure of the soil and the compact and cementation are obtained from the
agriculture experts.
The type of dataset acquired from the sensors are numerical values. Based on the values obtained
from the sensor, suitability range values (obtained from experts) are assigned to the parameters
considered for the land suitability model. For example, the pH value ranges from 7.2–6.7 for favorable
land conditions, 6.6–5.5 for less favorable land conditions, 5.4–4.5 for unfavorable land conditions,
and less than 4.5 for the very unfavorable land condition. The salinity value for the favorable land
condition is <2, 2–6 for less favorable land conditions, 6–12 for unfavorable land conditions, and
greater than 12 for the very unfavorable condition. The sampling period of data acquisition is one
day. The average of the one-day data obtained from various sensors is stored in the cloud network as
part of the preprocessing of the data. As the data may contain missing and noisy values, the mean of
the data is considered. Since the data contain different units of measurement (categorical, numerical),
normalization is done before applying the proposed model.
Figure 1 states the architectural diagram of the proposed model. The steps involved in this process
are shown in this figure. The data is collected from the farmland using sensors, and is then transferred
and stored with the help of the Raspberry Pi system in the cloud. The stored data is used here to
build the artificial intelligence recommendation model shown in the figure. The constructed model is
expected to classify the results into four different categories. Amazon Web Service (AWS) is used here
for the storage of the data.
The learning model construction is explained in the algorithm part given below. The built model
will be assessed with inputs received after every cultivation period, and accordingly, assessment results
will be provided. Different classes are considered here for land suitability assessment as most suitable
(class 1), suitable (class 2), moderately suitable (class 3), and unsuitable (class 4).
The Algorithm 1 presented here explains the steps involved in the processing the collected data
from the different sensors. Since the size of the data is quite high in size, it is necessary to handle the
data with the algorithm that is capable of handling the expected size of the data. That is the reason
why a neural network approach is considered here for processing the data. The algorithm presented
here is further explained below.
Artificial Intelligence Model building is given in the algorithm. Processing of the neural networks
relies upon the network topology, adjustments of weights, and activation functions. The network
topology defines the network arrangement together with its connecting links and nodes. In neural
networks, learning refers to the process of modifying weights of neural connections among the neurons
of the specific network. Activation functions are additional efforts applied over the input to achieve
the exact output. In the proposed work, a supervised learning-based neural network is adopted where
the learning process is dependent. The input vector is offered to the network model and obtains an
output vector, which is then compared to the desired or target vector. An error signal is developed
in case of a variation existing between the actual and the desired output vector. Depending on the
error signal the weights are adapted until the desired and the actual output matches. Neural networks
can be precisely defined as complex adaptive architecture, which can transform its internal structure
depending on the information passing through them. The following section explains the data collection
and further processes.
Sensors 2019, 19, 3667 7 of 16

Algorithm 1 MLP Algorithm for Machine Learning Model

1: Initialize the weights (w) to small random numbers, Bias (b) and Learning rate (α)
2: While (stopping criteria not met) do
3: For each training pattern (xi , yi ) do
4: Process the input forward:
5: Activate each input: xi = si , i ∈ 1 to n
Net input: yi = b + ni=1 xi .wi
P
6:
7: where b = bias, n = number of neurons, y—actual output, t—target output
8: Final output
 with the following activation
 function:


 1 i f yin > θ 


 0 i f − θ ≤ yin ≤ θ 
 
9: f yin = 
 

−1 i f yin < −θ

 

 
10: Adjust weight and bias:
11: i f y , t,
12: wi (new) = wi (old)+ ∝ txi
13: b(new) = b(old)+ ∝ t
14: i f y = t,
15: wi (new) = wi (old)
16: b(new) = b(old)
17: Calculating the average gradient:
18: Gradient calculation of each error with respect to wi
P P  2
19: E(w) = 21 i k∈y ti k − yi k
20: where ti k , yi k are the target and the actual output at the ith input instance
21: end
22: Calculation proceeds to the hidden layer
23: end

4. Implementation of a Proposed Sensor Network-Based Model

4.1. Description of the Dataset and Study Area

The data used for this experimentation are collected from the different villages of the Vellore and
Tiruvannamalai district in the state of Tamil Nadu, India. With expert opinions, we have shortlisted
the various parameters that are really influence the crop yield. This collection of the dataset includes
a combination of distinctive climate, soil, and groundwater properties. Four decision classes are
considered here for land suitability assessment, namely most suitable (class 1), suitable (class 2),
moderately suitable (class 3), and unsuitable (class 4).

4.2. Data Preparation

It is essential to prepare the information astutely before training the neural networks (NN) and
multilayer perceptron (MLP) models. The present real-world data obtained from sensors are distributed
non-uniformly, and henceforth, th data cannot be utilized legitimately amid training and testing of the
NN and MLP models. In this way, the input features are normalized, and categorical variables are
converted to numeric data through the data label encoder for efficient processing. The standardized
dataset is then sub-divided into training and independent test sets in the ratio of 75:25. The results
obtained are observed in the perspective of multiclass classification, as well as for the individual class
basis. The process of training and testing is repeated for a variable number of iterations starting from
30, 50, and so on, till 100 until the convergence optimization is met, and also for a variable number of
neurons and hidden processing layers for each model.
Sensors 2019, 19, 3667 8 of 16

4.3. Performance Measures

The performance parameters which are assessed for multiclass classification are characterized as
follows. For a class ‘C’, the classifier efficiency and performance could be assessed with true positive
(TP), true negative (TN), false positive (FP), and false negative (FN) that can be determined from the
counts of testing samples belonging to ‘C’ [42]. Various performance metrics appropriate for assessing
NN and MLP for multiclass classification problems are calculated [42]. Classification accuracy scores
could be used to determine the accuracy of the model by defining the number of correct predictions
made from all the predictions. Also, this enables us to determine the per-class effectiveness of the
model and is calculated as,
TP + TN
Accuracy = (1)
TP + TN + FP + FN
The performance metrics such as precision, recall, F1 score, etc., used for binary classification
needs to be generalized for the multiclass classification setting. To meet the appropriate needs, we
calculate the micro-average and the macro average. In micro-averaging, the performance is evaluated,
i.e., the precision, recall, and the F1 score are obtained from the individual TP, TN, FP, and FN of
the n-class model [42]. In macro-averaging, the mean of the performance of each individual class is
determined, providing equal weight to each of the classification labels. Precision is the ratio of the
predicted positive events that are actually positive.
Pn
i=1 TPi
Precisionmicro = Pn (2)
i=1 (FPi + TPi )
Pm TPi
i=1 (FPi +TPi )
Precision macro = (3)
n
Recall, also known as sensitivity, is the ratio of the positive events that are predicted correctly.
Pn
i=1 TPi
Recallmicro = Pn (4)
i=1 (FNi + TPi )
Pm TPi
i=1 (FNi +TPi )
Recall macro = (5)
n
F1 score is the harmonic mean of the recall and the precision. The higher the F1 score, the better
the model [27].
2 × Precisionmicro × Recallmicro
F1 scoremicro = (6)
Precisionmicro + Recallmicro
2 × Precisionmacro × Recallmacro
F1 scoremacro = (7)
Precisionmacro + Recallmacro
The Receiver Operating Characteristics (ROC)-Area Under the Curve (AUC) curve is the
performance measurement for classification problems under various threshold settings. ROC represents
the probability curve, and AUC is the degree or measure of separability distinguishing the classes.
The AUC-ROC score is represented in Table 2 under various iterations for the NN and MLP learning
models. Figures 2–4 represent the AUC-ROC curve for the multiclass classification problem under
various iterations for the learning models. The AUC-ROC curve indicates how well the probabilities of
the positive class are separated from the negative class.
Sensors
Sensors2019,
2019,19,
19,x 3667
FOR PEER REVIEW 9 of
9 of1616
Sensors 2019, 19, x FOR PEER REVIEW 9 of 16

(a) (b)
(a) (b)

(c)
(c)
Figure 2. Multiclass Area Under Curve (AUC)-Receiver Operating Characteristics (ROC) curve for
Neural Network
Multiclass
Figure 2. Multiclass
Figure Area (NN)Curve
Area Under
Under Curve (a) 𝑁 = 30,Operating
model(AUC)-Receiver
for:
(AUC)-Receiver (b) 𝑁 = 50,
Operating (c) 𝑁 =80. (𝑁
Characteristics
Characteristics − curve
(ROC)
(ROC) curve
forfor
Neural Network (NN)
Neural model (NN)
Network Number
Nh = 30,
for: (a)model of(a)
for:(b)hidden =neurons)
Nh𝑁= 50, Nh =𝑁80.= (N
(c) (b)
30, 50,h (c) 𝑁 =80.of
− Number −
(𝑁hidden neurons)
Number of hidden neurons)

(a) (b)
(a) Figure 3. Cont.
(b)
Sensors 2019, 19, 3667 10 of 16
Sensors
Sensors2019,
2019,19,
19,xxFOR
FORPEER
PEERREVIEW
REVIEW 10
10 of
of 16
16

(c)
(c)
Figure
Figure 3.
Figure 3. Multiclass
Multiclass AUC-ROC
Multiclass AUC-ROC
AUC-ROCcurve curve
curveforfor
for MultiLayer
MultiLayer
MultiLayer Perceptron
Perceptron
Perceptron (MLP)
(MLP)
(MLP) model
withwith
model
model threethree
with three
hiddenhidden
hidden
layers
layers
for: (a)for:
layers Nh(a)
for: (a) 𝑁
𝑁 (b)
= 30, == 30,
30, (b) 𝑁
𝑁(c)==N
= 50,
Nh (b) 50,
50,
h =
(c)
(c) 𝑁
𝑁
80. =
= 80.
80.

(a)
(a) (b)
(b)

(c)
(c)
Figure
Figure4.
Figure 4.4.Multiclass
Multiclass
Multiclass AUC-ROC
AUC-ROC curve
curve
AUC-ROC for
forMLP
curve MLP
for model
model
MLP with four
fourhidden
withwith
model hidden layers for:
for:(a)
layerslayers
four hidden 𝑁𝑁 =(a)
(a)for: =30,N(b)
30, =𝑁
h(b) 𝑁
30,
==(b)
50,
50,N(c)
(c)
h = 𝑁𝑁50,= 80.
(c)
= N
80. h = 80.

5. Besides the Results

5. Experimental
Experimental above-mentioned micro and macro-measures, the errors—Mean squared error (MSE),
Results
and root mean squared error (RMSE)—are also determined during the resulting process. Nevertheless,
The
The proposed
proposed sensor-based
sensor-based AI AI model
model evaluates
evaluates the
the experimental
experimental agriculture
agriculture land
land based
based onon the
the
it is prudent to mention that the stochastic gradient descent does not require MSE and RMSE during
considered attributes for cultivating the crop, which is expected to give better yield. The
considered attributes for cultivating the crop, which is expected to give better yield. The agriculture agriculture
training, instead it needs the error term between the random sample and its prediction for refurbishing
data
dataobtained
obtainedfromfromsensors
sensorsconsists
consistsof
of1000
1000alternatives
alternativeswithwith14
14attributes.
attributes.Out
Outof of1000
1000data
datainstances,
instances,
750
750 data instances are used for training the model, and the remaining 250 are used for testing the
data instances are used for training the model, and the remaining 250 are used for testing the
model.
model. The
The test
test dataset
dataset has
has been
been collected
collected separately
separately after
after every
every crop
crop yield
yield in
in aa span
span of
of one
one year.
year. The
The
proposed
proposed model
model classifies
classifies the
the experimental
experimental dataset
dataset into
into four
four decision
decision classes,
classes, namely
namely most
most suitable
suitable
(class
(class 1), suitable (class 2), moderately suitable (class 3), and unsuitable (class 4). The agriculture site
1), suitable (class 2), moderately suitable (class 3), and unsuitable (class 4). The agriculture site
Sensors 2019, 19, 3667 11 of 16

the weights of the neural networks. MSE and RMSE for determining the average error of convergence
of the model could be calculated as follows.
N
1 X
MSE = (yi − y0i )2 (8)
N
i=1

Here N is the total number of data samples, yi is the target or the hypothesis for ith data instance
and y0i is the outcome or the output of the ith data instance of the learning model.

5. Experimental Results
The proposed sensor-based AI model evaluates the experimental agriculture land based on the
considered attributes for cultivating the crop, which is expected to give better yield. The agriculture
data obtained from sensors consists of 1000 alternatives with 14 attributes. Out of 1000 data instances,
750 data instances are used for training the model, and the remaining 250 are used for testing the model.
The test dataset has been collected separately after every crop yield in a span of one year. The proposed
model classifies the experimental dataset into four decision classes, namely most suitable (class 1),
suitable (class 2), moderately suitable (class 3), and unsuitable (class 4). The agriculture site belongs to
class 1, and class 2 could be used for crop cultivation in the present form, whereas the agriculture farm
belongs to class 3 needs to be further processed by applying the appropriate manure before cultivating
the crop, and the land that comes under class 4 cannot be used for cultivation. The collected agriculture
dataset here is applied to the MLP algorithm for training purpose. The performance of the MLP-based
multiclass classification model is compared with the results obtained from neural networks.
The number of hidden layers and the number of neurons in each hidden layer are one of the most
significant architectural parameters which legitimately impacts the execution of MLP and NN during
the training and preparation of the information for setting up a learning base. In any case, determining
these parameters prior has been an unsolved problem in ML research. In the proposed work for NN,
the experiments were performed with the number of hidden neurons (Nh ) to be set as 30, 50, and 80
and the results were noted. For MLP, the experiment is repeated for three and four numbers of hidden
layers and the neurons in each layer were 30, 50, and 80. All the results have been epitomized and
defined in tables for varying the Nh values. Tables 1 and 2 defines the results obtained for each class
separately using NN and MLP with Nh = 30. Likewise, Tables 3 and 4 defines the results obtained
for Nh = 50 and Nh = 80. On the other hand, Table 4 describes the results attained by the NN and
MLP based on multiclass classification with Nh = 30. Likewise, Tables 5 and 6 illustrates the multiclass
classification results obtained for Nh = 50 and Nh = 80 accordingly. It needs to be noticed that all the
results that appeared in these three tables are averaged and equated over ten independent simulations
for each of the learning models.

Table 1. Performance of Neural Networks (NN) and MultiLayer Perceptron (MLP) with Nh = 30 for
each class averaged over ten independent runs.

Performance Performance of MLP with Three Performance of MLP with Four

Performance of NN
Measures Hidden Layers Hidden Layers
Class 1 Class 2 Class 3 Class 4 Class 1 Class 2 Class 3 Class 4 Class 1 Class 2 Class 3 Class 4
Accuracy 0.86 0.88 0.85 0.83 0.93 0.96 0.94 0.945 0.91 0.93 0.924 0.912
Recallmicro 0.85 0.87 0.862 0.884 0.951 0.94 0.92 0.938 0.92 0.914 0.93 0.911
Recallmacro 0.20 0.215 0.223 0.218 0.23 0.22 0.23 0.217 0.22 0.229 0.21 0.23
Precisionmicro 0.875 0.864 0.881 0.87 0.951 0.96 0.95 0.959 0.92 0.926 0.931 0.93
Precisionmacro 0.22 0.204 0.213 0.221 0.239 0.24 0.23 0.248 0.22 0.236 0.221 0.23
F1 Scoremicro 0.88 0.892 0.887 0.891 0.948 0.96 0.95 0.946 0.91 0.92 0.924 0.918
F1 Scoremacro 0.21 0.224 0.217 0.23 0.219 0.23 0.22 0.235 0.22 0.232 0.221 0.23
AUC-ROC
0.86 0.88 0.88 0.89 0.97 0.95 0.97 0.969 0.95 0.952 0.96 0.952
Score
MSE 0.235 0.241 0.24 0.237 0.047 0.04 0.04 0.042 0.08 0.087 0.091 0.09
RMSE 0.26 0.27 0.274 0.265 0.121 0.10 0.11 0.12 0.27 0.31 0.33 0.29
Sensors 2019, 19, 3667 12 of 16

Table 2. Performance of NN and MLP with Nh = 50 for each class averaged over ten independent runs.

Performance Performance of MLP with Three Performance of MLP with Four

Performance of NN
Measures Hidden Layers Hidden Layers
Class 1 Class 2 Class 3 Class 4 Class 1 Class 2 Class 3 Class 4 Class 1 Class 2 Class 3 Class 4
Accuracy 0.97 0.982 0.971 0.969 0.98 0.99 0.98 0.987 0.99 0.99 0.998 0.99
Recallmicro 0.98 0.975 0.981 0.974 0.97 0.98 0.98 0.976 0.9 0.98 0.99 0.987
Recallmacro 0.23 0.24 0.248 0.25 0.24 0.25 0.24 0.247 0.24 0.25 0.25 0.25
Precisionmicro 0.981 0.97 0.984 0.975 0.97 0.98 0.97 0.979 0.98 0.99 0.994 0.99
Precisionmacro 0.25 0.257 0.249 0.25 0.25 0.24 0.25 0.242 0.25 0.249 0.25 0.25
F1 Scoremicro 0.972 0.979 0.98 0.981 0.99 0.98 0.99 0.991 0.98 0.991 0.997 0.995
F1 Scoremacro 0.25 0.242 0.239 0.24 0.249 0.25 0.25 0.24 0.25 0.249 0.24 0.25
AUC-ROC
0.99 0.987 0.982 0.993 0.99 0.98 0.99 0.99 0.99 0.987 0.99 0.98
Score
MSE 0.018 0.02 0.014 0.01 0.001 0.00 0.00 0.001 0.00 0.001 0.001 0.001
RMSE 0.11 0.101 0.112 0.104 0.01 0.02 0.01 0.019 0.04 0.039 0.04 0.04

Table 3. Performance of NN and MLP with Nh = 80 for each class averaged over ten independent runs.

Performance Performance of MLP with Three Performance of MLP with Four

Performance of NN
Measures Hidden Layers Hidden Layers
Class 1 Class 2 Class 3 Class 4 Class 1 Class 2 Class 3 Class 4 Class 1 Class 2 Class 3 Class 4
Accuracy 0.98 0.99 0.987 0.99 0.99 0.98 0.99 0.985 0.99 0.99 0.987 0.99
Recallmicro 0.99 0.995 0.993 0.984 0.994 0.99 0.99 0.992 0.98 0.99 0.996 0.989
Recallmacro 0.25 0.245 0.249 0.25 0.25 0.25 0.24 0.248 0.25 0.249 0.237 0.25
Precisionmicro 0.98 0.986 0.99 0.992 0.99 0.99 0.98 0.991 0.99 0.986 0.992 0.99
Precisionmacro 0.25 0.248 0.24 0.25 0.25 0.25 0.25 0.24 0.25 0.251 0.25 0.248
F1 Scoremicro 0.992 0.984 0.99 0.991 0.997 0.99 0.98 0.99 0.99 0.987 0.989 0.99
F1 Scoremacro 0.25 0.247 0.25 0.249 0.25 0.24 0.24 0.25 0.24 0.248 0.25 0.25
AUC-ROC
0.99 0.97 0.99 0.985 0.997 0.99 0.99 0.987 0.99 0.97 0.985 0.99
Score
MSE 0.01 0.03 0.01 0.01 0.001 0.00 0.09 0.001 0.00 0.00 0.01 0.001
RMSE 0.058 0.06 0.05 0.068 0.004 0.01 0.00 0.004 0.00 0.005 0.002 0.004

Table 4. Performance of NN and MLP with Nh = 30 for multiclass classification averaged over ten
independent runs.

Performance of MLP Performance of MLP

Performance Measures Performance of NN with Three Hidden with Four Hidden
Layers Layers
Accuracy 0.89 0.959 0.926
Recallmicro 0.897 0.96 0.92
Recallmacro 0.22 0.24 0.23
Precisionmicro 0.88 0.963 0.93
Precisionmacro 0.21 0.241 0.23
F1 Scoremicro 0.895 0.956 0.926
F1 Scoremacro 0.223 0.239 0.231
Multiclass AUC-ROC Score 0.89 0.972 0.95
MSE 0.24 0.04 0.09
RMSE 0.275 0.116 0.3
Sensors 2019, 19, 3667 13 of 16

Table 5. Performance of NN and MLP with Nh = 50 for multiclass classification averaged over ten
independent runs.

Performance of MLP Performance of MLP

Performance Measures Performance of NN with Three Hidden with Four Hidden
Layers Layers
Accuracy 0.98 0.99 0.999
Recallmicro 0.983 0.988 0.99
Recallmacro 0.245 0.25 0.25
Precisionmicro 0.989 0.989 0.99
Precisionmacro 0.26 0.254 0.25
F1 Scoremicro 0.986 0.992 0.99
F1 Scoremacro 0.246 0.25 0.25
Multiclass AUC-ROC Score 0.991 0.998 0.998
MSE 0.016 0.001 0.001
RMSE 0.103 0.01 0.04

Table 6. Performance of NN and MLP with Nh = 80 for multiclass classification averaged over ten
independent runs.

Performance of MLP Performance of MLP

Performance Measures Performance of NN with Three Hidden with Four Hidden
Layers Layers
Accuracy 0.99 0.998 0.999
Recallmicro 0.993 0.995 0.998
Recallmacro 0.248 0.25 0.25
Precisionmicro 0.99 0.99 0.99
Precisionmacro 0.247 0.25 0.25
F1 Scoremicro 0.993 0.998 0.99
F1 Scoremacro 0.25 0.249 0.25
Multiclass AUC-ROC Score 0.99 0.998 0.99
MSE 0.01 0.001 0.001
RMSE 0.06 0.004 0.003

6. Result Analysis
The obtained classification results, which are delineated in Tables 2–6, are basically dependent on
different performance parameters. The proposed work compiles various performance parameters for
the evaluation and assessment of NN and MLP for the multiclass classification dataset. The training
performance for the NN for the proposed dataset is found to be varying for different values of Nh such
as Nh = 30, Nh = 50, and Nh = 80. The performance measures of the NN model are found to improve as
the number of neurons in the hidden layer Nh increases. With the increase in Nh , the NN is capable
of predicting the test data hypothesis with improved accuracy. Moreover, the MSE and the RMSE
are found to be very low and decrease accordingly in all three cases. Further, this results in better
convergence of error, leading to possibly proper learning of features. The performance parameters,
like the accuracy, precision, and others mentioned, seems to follow a similar manner.
Hence, it could be presumed that predictive accuracy could be a good measure for this dataset.
The performance results obtained using the MLP with three hidden layers are found to be much better
than that of NN. Similar to the NN model, this MLP model shows improved performance results
with the increasing number of Nh . However, since the number of hidden layers has increased the
architectural complexity, the convergence optimization is achieved with less number of iterations when
compared to the NN.
The accuracy and other performance measures are found to improve accordingly with an increase
in the Nh . On observing the performance measures of MLP with four hidden layers, the model
individually is found to provide better results with improved performance with an increase in Nh .
Sensors 2019, 19, 3667 14 of 16

However, when observed for the test case of Nh = 30, MLP with three hidden layers is found to produce
better results than the MLP with four hidden layers. Further, this may be due to the increase in
architectural complexity with an increased number of hidden layers. Also. if the Nh is set to as high as
100, there might be a high risk of over-fitting the training data which would lead to poor performance
than those obtained with Nh values of 50 and 80. Besides, provided the current experimental setup,
one could likewise accomplish slightly different results due to the underlying fact that the biases and
initial weights of the neural networks are determined randomly. If an appropriate weight set is fixed
initially, one could obtain superior results for similar settings. This assertion could be endorsed by the
fact that the gradient descent may not generally ensure a near-ideal weight set towards the end of the
neural network training process.
According to the results obtained inTable 6, the performance of MLP with four hidden layers
is high compared to the performance of the Neural Network and the MLP with three hidden layer
approaches used in some of the existing literature. Therefore, the proposed model assesses the given
agriculture land and provides better results for sustainable agriculture development compared to other
approaches. This model provides a reliable decision on the suitability level of the agriculture land in
four different categories, helping agriculturists to assess their land appropriately. Thus, this proposed
model could be used as a recommendation model for land suitability to improve the crop production
for sustainable agriculture development.

7. Conclusions
Since agriculture is the backbone of any country, it is necessary to ensure its sustainable growth
over the years. This work has presented a model that would be as accurate as 99%, which would be the
most desirable. The data collected through various sensors, handled here with MLP with four hidden
layers, has ensured better efficiency. A proper advisory system with precise instructions would always
deliver better results. Thus, the accuracy score to precision score presented depicts the efficiency of this
proposed approach, which will ensure appropriate classification. Multiclass classification in agriculture
would further fine-tune the recommendation system to guide farmers appropriately. Rather than
binary classification, this one would guide the farmers’ precisely. Thus, this approach would provide
real-time data to ensure better crop yield productivity.

Author Contributions: This research specifies below the individual contributions. Conceptualization—D.R.V.
and N.D.; Data curation—K.S.; Formal analysis—N.D.; Funding acquisition—S.H.C.; Investigation—D.R.V.;
Methodology—K.S.; Project administration—K.S.; Resources—S.H.C.; Software—D.E.; Supervision—D.E., K.S.
and C.I.; Validation—D.R.V. and K.S.; Visualization—D.R.V. and C.I.; Writing—review & editing—C.I.
Funding: This research received no external funding.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Ahmed, A.N.; de Hussain, I.D. Internet of Things (IoT) for smart precision agriculture and farming in rural
areas. IEEE Internet Things J. 2018, 5, 4890–4899. [CrossRef]
2. Elijah, O.; Rahman, T.A.; Orikumhi, I.; Leow, C.Y.; Hindia, M.N. An overview of Internet of Things (IoT) and
data analytics in agriculture: benefits and challenges. IEEE Internet Things J. 2018, 5, 3758–3773. [CrossRef]
3. Somov, A.; Shadrin, D.; Fastovets, I.; Nikitin, A.; Matveev, S.; Hrinchuk, O. Pervasive Agriculture: IoT-Enabled
Greenhouse for Plant Growth Control. IEEE Pervasive Comput. 2018, 17, 65–75. [CrossRef]
4. Bu, F.; Wang, X. A smart agriculture IoT system based on deep reinforcement learning. Future Gener.
Comput. Syst. 2019, 99, 500–507. [CrossRef]
5. Alreshidi, E. Smart Sustainable Agriculture (SSA) solution underpinned by Internet of Things (IoT) and
Artificial Intelligence (AI). arXiv 2019, arXiv:1906.03106. [CrossRef]
6. Kamienski, C.; Soininen, J.-P.; Taumberger, M.; Dantas, R.; Toscano, A.; Cinotti, T.S.; Maia, R.F.; Neto, A.T.
Smart water management platform: Iot-based precision irrigation for agriculture. Sensors 2019, 19, 276.
[CrossRef] [PubMed]
Sensors 2019, 19, 3667 15 of 16

7. Stočes, M.; Vaněk, J.; Masner, J.; Pavlík, J. Internet of Things (iot) in agriculture-selected aspects. Agris-line
Pap. Econ. Inf. 2016, 8, 83–88. [CrossRef]
8. Guardo, E.; Di Stefano, A.; La Corte, A.; Sapienza, M.; Scatà, M. A fog computing-based iot framework for
precision agriculture. J. Internet Technol. 2018, 19, 1401–1411.
9. Keswani, B.; Mohapatra, A.G.; Mohanty, A.; Khanna, A.; Rodrigues, J.J.; Gupta, D.; de Albuquerque, V.H.C.
Adapting weather conditions based IoT enabled smart irrigation technique in precision agriculture
mechanisms. Neural Comput. Appl. 2019, 31, 277–292. [CrossRef]
10. Liu, S.; Guo, L.; Webb, H.; Ya, X.; Chang, X.; Yao, X. Internet of Things Monitoring System of Modern
Eco-Agriculture Based on Cloud Computing. IEEE Access 2019, 7, 37050–37058. [CrossRef]
11. Khanna, A.; Kaur, S. Evolution of Internet of Things (IoT) and its significant impact in the field of Precision
Agriculture. Comput. Electron. Agric. 2019, 157, 218–231. [CrossRef]
12. Ray, P.P. Internet of things for smart agriculture: Technologies, practices and future direction. J. Ambient.
Intell. Smart Environ. 2017, 9, 395–420. [CrossRef]
13. Muangprathub, J.; Boonnam, N.; Kajornkasirat, S.; Lekbangpong, N.; Wanichsombat, A.; Nillaor, P. IoT and
agriculture data analysis for smart farm. Comput. Electron. Agric. 2019, 156, 467–474. [CrossRef]
14. Himesh, S. Digital revolution and Big Data: a new revolution in agriculture. CAB Rev. 2018, 13, 1–7.
[CrossRef]
15. Mukherjee, A.; Misra, S.; Raghuwanshi, N.S.; Mitra, S. Blind Entity Identification for Agricultural IoT
Deployments. IEEE Internet Things J. 2018, 6, 3156–3163. [CrossRef]
16. Park, H. MLP modeling for search advertising price prediction. J. Ambient. Intell. Humaniz. Comput. 2019, 1–7.
[CrossRef]
17. Pouladi, N.; Jafarzadeh, A.A.; Shahbazi, F.; Ghorbani, M.A. Design and implementation of a hybrid MLP-FFA
model for soil salinity prediction. Environ. Earth Sci. 2019, 78, 159. [CrossRef]
18. Tamouridou, A.A.; Pantazi, X.E.; Alexandridis, T.; Lagopodi, A.; Kontouris, G.; Moshou, D. Spectral
identification of disease in weeds using multilayer perceptron with automatic relevance determination.
Sensors 2018, 18, 2770. [CrossRef]
19. Niedbała, G. Application of artificial neural networks for multi-criteria yield prediction of winter rapeseed.
Sustain. 2019, 11, 533. [CrossRef]
20. Ruan, J.; Wang, Y.; Chan, F.T.S.; Hu, X.; Zhao, M.; Zhu, F.; Shi, B.; Shi, Y.; Lin, F. A Life Cycle Framework
of Green IoT-Based Agriculture and Its Finance, Operation, and Management Issues. IEEE Commun. Mag.
2019, 57, 90–96. [CrossRef]
21. Khishe, M.; Safari, A. Classification of Sonar Targets Using an MLP Neural Network Trained by Dragonfly
Algorithm. Wirel. Pers. Commun. 2019, 1–20. [CrossRef]
22. Chatterjee, S.; Dey, N.; Sen, S. Soil moisture quantity prediction using optimized neural supported model for
sustainable agricultural applications. Sustain. Comput. Inform. Syst. 2018. [CrossRef]
23. Ghorbani, M.A.; Deo, R.C.; Karimi, V.; Kashani, M.H.; Ghorbani, S. Design and implementation of a hybrid
MLP-GSA model with multi-layer perceptron-gravitational search algorithm for monthly lake water level
forecasting. Stoch. Environ. Res. Risk Assess. 2019, 33, 125–147. [CrossRef]
24. Dewi, C.; Chen, R.C. Decision Making Based on IoT Data Collection for Precision Agriculture. In Asian
Conference on Intelligent Information and Database Systems; Springer: Cham, Germany, 2019.
25. Nawandar, N.K.; Satpute, V.R. IoT based low cost and intelligent module for smart irrigation system.
Comput. Electron. Agric. 2019, 162, 979–990. [CrossRef]
26. Morais, R.; Silva, N.; Mendes, J.; Adão, T.; Pádua, L.; López-Riquelme, J.; Pavón-Pulido, N.; Sousa, J.J.;
Peres, E. Mysense: A comprehensive data management environment to improve precision agriculture
practices. Comput. Electron. Agric. 2019, 162, 882–894. [CrossRef]
27. Yang, Z.; Ding, Y.; Hao, K.; Cai, X.; Caia, X. An adaptive immune algorithm for service-oriented agricultural
Internet of Things. Neurocomputing 2019, 344, 3–12. [CrossRef]
28. Kale, A.P.; Sonavane, S.P. IoT based Smart Farming: Feature subset selection for optimized high-dimensional
data using improved GA based approach for ELM. Comput. Electron. Agric. 2019, 161, 225–232. [CrossRef]
29. Baseca, C.C.; Sendra, S.; Lloret, J.; Tomas, J. A Smart Decision System for Digital Farming. Agronomy
2019, 9, 216. [CrossRef]
30. Shi, X.; An, X.; Zhao, Q.; Liu, H.; Xia, L.; Sun, X.; Guo, Y. State-of-the-Art Internet of Things in Protected
Agriculture. Sensors 2019, 19, 1833. [CrossRef]
Sensors 2019, 19, 3667 16 of 16

31. Rad, P.; Choo, K.-K.R.; Roopaei, M. Cloud of Things in Smart Agriculture: Intelligent Irrigation Monitoring
by Thermal Imaging. IEEE Cloud Comput. 2017, 4, 10–15.
32. Mehra, M.; Saxena, S.; Sankaranarayanan, S.; Tom, R.J.; Veeramanikandan, M. IoT based hydroponics system
using Deep Neural Networks. Comput. Electron. Agric. 2018, 155, 473–486. [CrossRef]
33. Rajeswari, S.; Suthendran, K.; Rajakumar, K. A smart agricultural model by integrating IoT, mobile and
cloud-based big data analytics. In Proceedings of the 2017 International Conference on Intelligent Computing
and Control (I2C2), Coimbatore, India, 23–24 June 2017; pp. 1–5.
34. Chen, Y.; Zhen, Z.; Yu, H.; Xu, J. Application of Fault Tree Analysis and Fuzzy Neural Networks to Fault
Diagnosis in the Internet of Things (IoT) for Aquaculture. Sensors 2017, 17, 153. [CrossRef] [PubMed]
35. Malche, T.; Maheshwary, P. Internet of Things (IoT) Based Water Level Monitoring System for Smart Village.
In Proceedings of the Advances in Intelligent Systems and Computing, Lviv, Ukraine, 6–10 September 2016;
Springer Science+Business Media: Berlin/Heidelberg, Germany, 2017; Volume 508, pp. 305–312.
36. Hossam, M.; Kamal, M.; Moawad, M.; Maher, M.; Salah, M.; Abady, Y.; Hesham, A.; Khattab, A. PLANTAE:
An IoT-Based Predictive Platform for Precision Agriculture. In Proceedings of the 2018 International
Japan-Africa Conference on Electronics, Communications and Computations (JAC-ECC), Alexandria, Egypt,
16–18 December 2018; pp. 87–90.
37. Tan, W.; Zhao, C.; Wu, H. Intelligent alerting for fruit-melon lesion image based on momentum deep learning.
Multimed. Tools Appl. 2016, 75, 16741–16761. [CrossRef]
38. Geng, L.; Dong, T. An Agricultural Monitoring System Based on Wireless Sensor and Depth Learning
Algorithm. Int. J. Online Eng. (iJOE) 2017, 13, 127–137. [CrossRef]
39. Li, X.; Yuan, J.; Ma, H.; Yao, W. Fast and parallel trust computing scheme based on big data analysis for
collaboration cloud service. IEEE Trans. Inf. Forensics Secur. 2018, 13, 1917–1931. [CrossRef]
40. Ferrández-Pastor, F.; García-Chamizo, J.; Nieto-Hidalgo, M.; Mora-Martínez, J. Precision agriculture design
method using a distributed computing architecture on internet of things context. Sensors 2018, 18, 1731.
[CrossRef] [PubMed]
41. Torres, A.F.; Walker, W.R.; McKee, M. Forecasting daily potential evapotranspiration using machine learning
and limited climatic data. Agric. Water Manag. 2011, 98, 553–562. [CrossRef]
42. Salman, M.S.; Kukrer, O.; Hocanin, A. Recursive inverse algorithm: Mean-square-error analysis.
Digit. Signal Process. 2017, 66, 10–17. [CrossRef]

© 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).