Neural Computing and Applications (2019) 31 (Suppl 1):S277–S292

https://doi.org/10.1007/s00521-018-3737-1

S . I . : M A C H I N E L E A R N I N G A P P L I CA T I O N S F O R S E L F - O R G A N I Z E D W I R E L E S S
NETWORKS

Adapting weather conditions based IoT enabled smart irrigation

technique in precision agriculture mechanisms
Bright Keswani1 • Ambarish G. Mohapatra2 • Amarjeet Mohanty3 • Ashish Khanna4 • Joel J. P. C. Rodrigues5,6,7 •

Deepak Gupta4 • Victor Hugo C. de Albuquerque8

Received: 15 July 2018 / Accepted: 10 September 2018 / Published online: 26 September 2018
 The Natural Computing Applications Forum 2018

Abstract
Precision agriculture is the mechanism which controls the land productivity and maximizes the revinue and minimizes the
impact on sorroundings by automating the complete agriculture processes. This projected work relies on independent
internet of things (IoT) enabled wireless sensor network (WSN) framework consisting of soil moisture (MC) probe, soil
temperature measuring device, environmental temperature sensor, environmental humidity sensing device, CO2 sensor,
daylight intensity device (light dependent resistor) to acquire real-time farm information through multi-point measurement.
The projected observance technique consists of all standalone IoT-enabled WSN nodes used for timely data acquisitions
and storage of agriculture information. The farm history is additionally stored for generating necessary action throughout
the whole course of farming. The work summarizes the optimum usage of irrigation by the precise management of water
valve using neural network-based prediction of soil water requirement in 1 h ahead. Our proposed irrigation control
scheme utilizes structural similarity (SSIM)-based water valve management mechanism which is used to locate farm
regions having water deficiency. Moreover, a close comparative study of optimization techniques, like variable learning
rate gradient descent, gradient descent for feedforward neural network-based pattern classification, is performed and the
best practice is employed to forecast soil MC on hourly basis together with interpolation method for generating soil
moisture content (MC) distribution map. Finally, SSIM index-based soil MC deficiency is calculated to manipulate the
specified valves for maintaining uniform water requirement through the entire farm area. The valve control commands are
again processed using fuzzy logic-based weather condition modeling system to manipulate control commands by con-
sidering different weather conditions.

Keywords Soil moisture content  Wireless sensor network  Internet of things  Variable learning rate gradient descent 
Gradient descent  Structural similarity index (SSIM)  Interpolation  Fuzzy logic

& Victor Hugo C. de Albuquerque Joel J. P. C. Rodrigues

[email protected] [email protected]
Bright Keswani Deepak Gupta
[email protected] [email protected]
Ambarish G. Mohapatra Extended author information available on the last page of the article
[email protected]
Amarjeet Mohanty
[email protected]
Ashish Khanna
[email protected]

123
S278 Neural Computing and Applications (2019) 31 (Suppl 1):S277–S292

1 Introduction maximize the outcome for real time forecast. The experi-
mental results reveal that the system suggested by Liang
Wireless sensor networks (WSNs) are constantly used from et al. [11] gets the benefits of low-cost, higher precision
the scientific community since the potential of ecological and suitable maintenance. The process implied by Liang
monitoring software composed of cost-effective possibility et al. [11] is a straight sensing of soil MC using WSN
to construct and process all kinds of information with platform, and this content does not signify the mechanism
greater resolution, these systems are regarded as a impor- for water conservation by simply contemplating weather
tant part of ubiquitous computing [1, 2]. conditions. Similarly, the work does not emphasize on
Therefore, using IoT-driven WSN platform in precision remote access like IoT-enabled control [12, 13].
agriculture can revolutionize the information assortment in At the similar circumstance, additional property-asso-
agricultural area and encourage the sought after excep- ciated parameters and ecological parameters might impact
tionally machine-driven agriculture method which the irrigation demands. Timely forecast of soil moisture
demands intensive understanding of ecological conditions content may soon demonstrate much improved irrigation
in the base level and speedy communication of the infor- optimization by considering all of the crucial parameters
mation to a distant or local server where the source of that impact the soil water flow. Further, the prediction
storage and process, identification of insects inside the methodology for a large agriculture land requires high
plants, burial or hyperbolic moist, the alternative generat- computing resources which is possible in a cloud-enabled
ing, and so the direction of plantation instrumentation is services. Thus, IoT comes into the picture with data ana-
completed instantly (automatic propulsion apparatus such lytics and machine learning framework [14–18]. The sug-
as sprinklers, foggers, valve-controlled irrigation method, gested method is determined by establishing wireless
and so forth is used for management irrigation, and fertil- sensors in the farm to acquire real-time observation and
ization and pest control in order to offset the negative states evaluate soil moisture content (MC) in the test site. The test
forming wireless sensor in addition to actuator network site is of length 50 m and width 50 m. It is implanted with
(WSAN). The above system is accessible globally due to Bermuda grass that is generally used as an ayurvedic
the involvement of IoT technology [3–5]. medicine for colitis and gut infections. The distributed
A fast evaluation and observation of soil moisture con- sensor network nodes are utilized to acquire real-time soil
tent (MC) of this massive area is necessary so that optimum as well as environmental parameters. The general param-
irrigation will be possible at agriculture plantation region. eters considered are soil moisture content (MC), soil tem-
Gravimetric method is called the standard to determine soil perature, temperature of the surrounding, humidity level of
water content version, but it is not effective in providing the surrounding, CO2 level of air, as well as sunlight
large-scale fast data collection and can be limited to a intensity level. The WNS environment is configured as
compact dimension quantity. nine numbers of wireless sensor nodes and a single gate-
Geophysical techniques like ground penetrating radar way node (GN) or the coordinator node. The gateway node
(GPR) and electromagnetic induction (EMI) quantify a is also called as the coordinator node and others behave as
bigger sample quantity and therefore are accommodated to router nodes. The coordinator node is responsible to collect
cover huge areas economically, which could not triumph all the sensor information and forward it to the cloud
from different procedures [6–8]. Additional techniques of platform for necessary prediction of soil water require-
spectroscopic scanning-based soil moisture material dis- ments. The coordinator node is designed with Raspberry
covery are most frequently used to find soil MC map of a Pi3, Zigbee tranceiver (S2 series) and wifi circuitry (On
huge agriculture property, but utilizing this way is fre- board). Further, gateway node connects the internet appli-
quently expensive work, and also they are not used to find cation via the wifi network for the prediction of the soil
soil MC variations through the summertime [9, 10]. water requirements.
Within exactly the same circumstance, wireless sensor The essential data set is obtained over wifi connectivity
network (WSN)-based land MC evaluation can be utilized to the internet server unit for the estimation of soil MC.
for real-time observation and evaluation of dirt MC Further, it is used obtain the soil MC distribution map of
dynamics throughout the whole cropping time. Liang et al. the land. With this work structural similarity index (SSIM)-
[11] have suggested a plan and employ a real-time irriga- based soil MC deficiency is calculated to manipulate irri-
tion forecast using GPRS and WSN system . The addition gation valves to maintain the required water content uni-
of the general system uses WSN to obtain soil MC infor- form over the farm area.
mation. The backend system employs the hereditary back- The performance of two different optimization algo-
propagation neural system to assess in addition to process rithms such as VLRGD and GD-based neural network
sensor information with simulated annealing algorithm to (NN) data prediction strategies is compared by computing
few statistical errors like RMSE, R-squared error and MSE.

123
Neural Computing and Applications (2019) 31 (Suppl 1):S277–S292 S279

Similarly, Majone et al. [19] have used wireless sensor • Development of a low-cost and high accuracy dis-
network to supervise MC dynamics of the top soil in both tributed IoT-enabled wireless sensor network environ-
time and space that is appropriate to examine the rela- ment for accurately monitoring the soil and
tionship between MC dynamics and plant physiology. The environmental parameters which are used for predicting
wireless sensor network deployment consists of twenty- timely water requirements.
seven locations (in verticals direction) configured by multi- • Development of a precise optimization mechanism for
hop configuration. The every wireless node contains 5 timely forecast of MC level.
numbers of soil moisture sensor probes those are deployed • Development of a robust mechanism for generating soil
in the varied depths like 10, 20, 30, 50 and 80, along with a moisture content distribution map using interpolation
temperature sensing device (temperature sensor PT100) at technique. The distribution map is generated by mea-
a distance downwards 25 cm [3, 17]. The method projected suring soil moisture content at nine positions of the
by Majone et al. [19] does not provide the mechanism for farmland using IoT enabled WSN environment (land
site-specific irrigation valve control for maintaining effi- dimension is defined in the experimental site section of
cient irrigation requirement. Irrigation scheduling and the article).
decision support system will provide better control over • Structural similarity index (SSIM) mechanisms that
farm land water requirement. And also the water require- associate interpolated data for generating zone-specific
ment is not precisely controlled by considering site-specific irrigation control commands to control irrigation
weather data. There are several weather condition-depen- valves.
dent irrigation controllers available commercially. All • Implementation of the weather condition-dependent
these instruments are based on few environmental param- irrigation control mechanism using fuzzy logic-based
eters like temperature, humidity and solar radiation weather model.
[20–24]. But other environmental parameters like wind
The detailed explanation of the above methodology is
speed, CO2 and soil-related parameters are also affects the
explained in the rest of the sections.
farm irrigation requirements [25–27]. The complete intel-
ligent weather-dependent irrigation control scheme is also
discussed in this article. In the similar perspective, Lenka
3 Experimental site
et al. [28] projected the MC prediction using partial-least-
square-regression and fuzzy logic technique to manage the
The test site is situated at Bhubaneswar (Latitude:
land irrigation requirements. The method produces MSE
20.351323, Longitude: 85.806547) in the eastern region of
about 0.183193 and RMSE about 0.428011. The method
India. The cultivated farm has an area of almost 5000 m2.
proposed by Lenka et al. [28] controls the farm irrigation
The profile consists of a layer of soppy soil, quite 2 m deep
valve in a real-time approach, but zone-wise irrigation map
with sandy loam surface. Here the key top soil quality was
is not included in the work. Similarly Lenka et al. [28]
found by laboratory study of untouched sample of 3.5 kg.
proposed BFGS-quasi-newton feedforward NN to forecast
Further, the dry bulk density and consistency of the soil
MC level in 1 h advance, and it is integrated with fuzzy
sample are about 1.43 g-cm-3 and 0.46 g-cm-3,
logic model to generate appropriate irrigation valve control
respectively.
commands. The land soil moisture requirement map is not
proposed in this work which will be produce better zone-
wise as well as crop-wise irrigation control. The moisture
map shows a better view over the soil moisture content
deficiency of the farm land [29, 30]. The irrigation
requirements can easily be addressed quickly by the IoT-
driven decision support system (DSS). Our SSIM-based
zone and crop specific irrigation control mechanism will
produce added advantage for efficient irrigation control
scheme over a large farming area.

2 Proposed methodology

The main aim of the research work described in this article

Fig. 1 Deployment of wireless sensor network nodes in a Bermuda
is to design a precise irrigation monitoring and control grass (Cynodon dactylon) planting site located at Bhubaneswar
mechanism consisting of the following objectives. (latitude: 20.351323, longitude: 85.806547)

123
S280 Neural Computing and Applications (2019) 31 (Suppl 1):S277–S292

The deployment of WSN nodes is shown in Fig. 1 where 4 Proposed IoT-driven WSN system
all nodes are connected with each other using multi-hop
mesh network. The architecture of a single WSN node is The battery capacity and backup are two major factors of
shown in Fig. 2. All sensors are interfaced with the node the WSN nodes. All sensors are connected to the analog-to-
via internal eight channels analog-to-digital converter digital converter (ADC) of the microcontroller. The battery
(ADC) of the microcontroller (ATMEGA328P-PU). The capacity chosen in this work is of 6 V/4.5 A/h lead acid
microcontroller communicates with the Zigbee S2 unit via type. The solar panel provides necessary charging current
RS232 interface. Each Zigbee S2 device is configured to the batter unit during day time. The solar panel chosen in
using mesh network topology. The digital out pin of the this work is of 7 V/250 mA. The fabricated WSN node is
Zigbee S2 is used to activate corresponding sprinkler valve shown in Fig. 2b. Similarly, the energy supply is also
unit. In this work, the irrigation mechanism consists of monitored in a real-time approach via ADC of the con-
sprinkler pipe lines with sprinkler heads at 0.41 m spacing troller. The data acquisition is carried out in every 10 min
as well as with a flow of 2.2 Lh-1. and during non-transmitting mode the Zigbee modem is
kept in a sleep mode which eventually increases the battery

Fig. 2 a Single node

architecture of the WSN
environment b proposed Single
node fabricated for acquiring
farm data. The IP67 housing
used for the sensor node is
shown in the right side figure

123
Neural Computing and Applications (2019) 31 (Suppl 1):S277–S292 S281

backup. During deep sleep mode, the Zigbee generally After the successful prediction of soil MC at every location
consumes a very low current as 0.02 lA. The Zigbee of the WSN nodes connected in the field area, bilinear
transceivers are configured in a mesh network pattern by interpolation is used to generate soil MC distribution
updating the device firmware. Finally, the gateway is throughout the farm land. The soil MC distribution map
responsible to send the data to the internet web server for generated using interpolation is compared with required
real-time analysis of sensor parameters. The complete unit soil MC using SSIM, which plays an important role in
is enclosed in a IP65 enclosure and mounted in a panel box. controlling zone-specific irrigation valves. The valve con-
The solar panel is inclined at 45 for maximum power trol command generated using SSIM is used to evaluate the
efficiency. actual necessity of water by implementing fuzzy logic-
In the similar context, Yunseop James et al. (2008) have based weather-dependent irrigation mechanism. A detailed
proposed a 295-m-long linear movement irrigation system flow model of the methodology is shown in Fig. 4. Further,
at Valley, Valmont Irrigation, Valley, NE [26]. It is the performance analysis of the above two prediction
observed that Bluetooth radio has many limitations as methodologies is listed in Table 1A (Fig. 5).
compared to Zigbee.
The wireless nodes are utilized to acquire sensor data
and forward it to the sink node for storing in its own 6 Variable learning rate gradient descent
database. The detailed architecture of the sink node is (VLRGD)
shown in Fig. 3. In this work, MATLAB is used to com-
municate with the sink node for collecting all data which is The neural network performance depends on learning rate;
used for soil MC prediction and irrigation valve control therefore, the error rate converges soon. The bigger
using SSIM technique. learning rate may converge the error rate of the neural
network training. But much bigger learning rate will make
the neural network unstable. Similarly, much smaller
5 Proposed soil MC prediction and irrigation learning rate will take long time to converge [14, 33]. To
control methodology improve the convergence time, the VLRGD neural network
assumes bigger learning rate a. Further, the new weight
The proposed MC prediction methodology is based on vector Viþ1 is modified with a varying ai . Again,gk is the
feedforward neural network where the input parameters are gradient of the error with respect to the weight matrix
the real-time wireless sensor data [31, 32]. [14, 34–36]. There is momentum (l) of gradient descent
In this article, the performance analysis of the training for a varying ak [33]. Normally, the updated weight vector
algorithms such as VLRGD and GD is deliberated for Viþ1 is defined in Eq. 1.
implementation of IoT-WSN bidirectional DSS (decision Viþ1 ¼ Vi  aiþ1 gi þ lVi1 ð1Þ
support system) in PA (precision agriculture) mechanism.

Fig. 3 Sink node architecture of

the WSN environment

123
S282 Neural Computing and Applications (2019) 31 (Suppl 1):S277–S292

Fig. 4 Fuzzy logic model for

the prediction of weather-
dependent soil MC requirement

Table 1 (A) Input and outputs used for soil moisture content pre- wkþ1 ¼ wk  ak gk ð2Þ
diction (B) Approximate values of seasonal crop water needs
The updated weight matrix depends on three factors such
(A) Input Output
as current weight matrix (wk), updated weight matrix (wk?1),
Current time Sunlight intensity Soil moisture learning rate (a) and gradient error (gk ) with respect to the
Environment temperature Sunrise time Content after weight matrix [28, 37, 38]. The negative sign of the weight
Environment humidity Air flow rate 1h matrix signifies that it is following the opposite direction.
Present soil moisture Air CO2 Duration
Content UV index 6.2 Bilinear interpolation on the grid data
Soil temperature
Soil type The bilinear transformation is a type of linear interpolation
technique where the interpolation of two variables (x and
(B) Crop Crop water need Crop type Crop water need
type (mm/total for the (mm/total for the
y) on a 2D plane. This interpolation first performs in one
complete cropping complete cropping direction and then again in another direction. For example,
period) period) an unknown function ‘f’ at two different points (a, b) need
to be interpolated where four discrete points are available
Alfalfa 800–1600 Onion 350–550
which is mentioned below.
Banana 1200–2200 Peanut 500–700
D11 = (a1, b1), D12 = (a1, b2), D21 = (a2, b1), and
Barley/ 450–650 Pea 350–500
oats/ D22 = (a2, b2)
wheat Firstly, the interpolation is performed in the x-direction
Bean 300–500 Pepper 600–900 and then in y-direction according to Eqs. 3 and 4. These
Cabbage 350–500 Potato 500–700 yields.
Citrus 900–1200 Rice 450–700 a2  a a  a1
f ða; y1 Þ  f ðD11 Þ þ f ðD21 Þ ð3Þ
(paddy) a2  a1 a 2  a1
Cotton 700–1300 Sorghum/ 450–650 a2  a a  a1
millet f ða; y2 Þ  f ðD12 Þ þ f ðD22 Þ ð4Þ
a2  a1 a 2  a1
Maize 500–800 Soybean 450–700
Melon 400–600 Sugarbeet 550–750 Further, the interpolation is performed from y-direction
to estimate the interpolated data according to Eq. 5.
b2  b b  b1
f ða; bÞ  f ða; b1 Þ þ f ða; b2 Þ ð5Þ
aiþ1 ¼ bai b2  b1 b2  b1

0:75 if updated error [ 1:05 ðfor previous errorÞ Finally, the same result is used to interpolate the values
b¼ in x-direction.
1:05 if updated error\1:05 ðfor previous errorÞ
6.1 Gradient descent 6.3 Structural similarity (SSIM) index
This optimization technique updates the weight matrix and The SSIM conveys the visual characteristics by calculating
bias values in the direction of the negative gradient [28]. a similarity measure between two spots such as x and y.
The updated weight matrix is based on Eq. 2.

123
Neural Computing and Applications (2019) 31 (Suppl 1):S277–S292 S283

Fig. 5 Complete architecture of the proposed irrigation requirement modeling

The SSIM is calculated by multiplying mean m(x, y), whatsoever. Uncertainty that is clear is carried by them. In
variance v(x, y), along with cross-correlation r(x, y) as the event the form of doubt occurs to appear as a conse-
mentioned in Eq. 6 [39]. quence of ambiguity; then the factor could be portrayed by
means of a membership function. Similarly, defuzzification
SSIMðx; yÞ ¼ mðx; yÞa vðx; yÞb cðx; yÞc ð6Þ
is the transformation of a fuzzy figure into an explicit
!a !b  
2lx ly þ C1 2dx dy þ C2 dx dy þ C3 c amount, as fuzzification is the conversion of an equal
¼  2  amount to some fuzzy quantity. Similarly, the output of a
l2x þ l2y þ C1 dx þ d2y þ C2 dx dy þ C3
fuzzy method may be the logical union of 2 or additional
where lx is mean of x, ly is mean of y, dx is standard fuzzy membership functions outlined on the universe of
deviation of x, dy is standard deviation of y, dxy is cross- discourse of the output variable. The fuzzy logic is a rule-
correlation (inner product) of the mean shifted matrix x–lx based model where the inputs and outputs are bounded to
and y–ly , as well as the Ci for i = 1, 2, 3 are small positive the rule sets. The outputs of the model are mathematically
bounded to the input parameters. In this article, we have
constants.
considered the real-time sensor parameters as the inputs to
the model and valve control commands are generated
6.4 Fuzzification and defuzzification
according to the rule set defined in the fuzzy logic model.
In this work, the weather information as well as the
Fuzzification is that the practice of building a sharp fuzzy
predicted soil MC is used to generate appropriate irrigation
set. This is sometimes accomplished by simply recognizing
valve control commands. The fuzzification and defuzzifi-
that a number of these amounts will be thought about to be
cation process is shown in Fig. 4.
eloquent and settled that are really not deterministic

123
S284 Neural Computing and Applications (2019) 31 (Suppl 1):S277–S292

Fig. 6 Flow chart of the

irrigation control mechanism
using hybrid decision model

6.5 Proposed flow model for irrigation valve mechanism takes into account of all the IoT enabled WSN
control node data for analyzing soil MC distribution using bilinear
interpolation. Finally, precise irrigation valve control is
The proposed flow model of the irrigation control done using SSIM index of the interpolated data with the
scheme is depicted in Fig. 6. The irrigation control reference soil MC. The agriculture irrigation valve control
is efficiently performed using real-time acquisition of IoT-
enabled WSN data, neural network prediction module and
Load Prediction
Soil Moisture Content in %

30
Actual
28 Neural Network Pattern Recognition 2
Best Training Performance is 0.0013267 at epoch 1999
10
26 Train
Best
24 1
10
Mean Squared Error (mse)

22
0 5 10 15 20 25
Time is Hours
0
10
Prediction Error
0.2
Neural Network Pattern Recognition Error -1
0.1 10
Residuals

0
-2
-0.1 10

-0.2
0 5 10 15 20 25
-3
Time is Hours 10
0 200 400 600 800 1000 1200 1400 1600 1800 2000
Fig. 7 Actual soil moisture content and predicted soil moisture 2000 Epochs
content versus time in hours. Error graph plotted between residuals
and time in hours [gradient descent optimization-based neural Fig. 8 Variable learning rate gradient descent-based neural network
network pattern classification] pattern classification training performance

123
Neural Computing and Applications (2019) 31 (Suppl 1):S277–S292 S285

5
Gradient = 0.032748, at epoch 2000 Best Training Performance is 0.038914 at epoch 2000
2
10 10
Train
gradient

0
10 Best

-5

Mean Squared Error (mse)

10 1
10
Validation Checks = 0, at epoch 2000
1
val fail

0 0
10

-1

Learning Rate = 0.017357, at epoch 2000

-1
2 10

1
lr

0 -2
0 200 400 600 800 1000 1200 1400 1600 1800 2000 10
2000 Epochs 0 200 400 600 800 1000 1200 1400 1600 1800 2000
2000 Epochs
Fig. 9 Training state of the variable learning rate gradient descent
neural network pattern classification training Fig. 11 The training performance of Gradient Descent optimiza-
tion based neural network pattern classification technique

fuzzy logic-based decision-making module. The DSS

Gradient = 0.034636, at epoch 2000
model is integrated with a GSM modem for SMS sending. 10
2

The interfacing is done using suitable AT command gradient

sequences through RS232 interface with the server station. 0

10

7 Results and discussions 10

-2

The optimization techniques utilized within the scope of Validation Checks = 0, at epoch 2000
1
this neural network-based prediction methodology are
variable learning rate gradient descent and gradient des- 0.5
val fail

cent. The parameters used to establish a relationship 0

between wireless sensor data and soil MC in 1 h advance
-0.5
are shown in Table 1A. Further, the crop water need can be
-1
0 200 400 600 800 1000 1200 1400 1600 1800 2000
2000 Epochs
Load Prediction
Soil Moisture Content in %

30
Fig. 12 Training state of the gradient descent optimization neural
Actual
28 Neural Network Pattern Recognition
network pattern classification training

26
studied from Table 1B where the water requirement of
24
some seasonal crops are listed. The performance analysis
22 of the prediction algorithm is done by evaluating
0 5 10 15 20 25
Time is Hours R-squared, MSE and RMSE.
Prediction Error It is observed that VLRGD-based feedforward NN pat-
0.5
Neural Network Pattern Recognition Error tern classification generates minimum MSE, RMSE, and
better R-squared value. Finally, this is used for the gener-
Residuals

0
ation of MC distribution map of the farm land by taking all
-0.5 the IoT enabled WSN node data. Few input parameters
such as sunrise time, UV index, and air flow rate are taken
-1
0 5 10 15 20 25 from weather forecast data, and soil type is taken from
Time is Hours
manual soil test of the farm land.
Fig. 10 Actual soil moisture content and predicted soil moisture VLRGD and NN pattern classification is used to predict
content versus time in hours. Error graph plotted between residuals soil MC by obtaining soil and environmental information.
and time in hours [gradient descent optimization-based neural
network pattern classification]

123
S286 Neural Computing and Applications (2019) 31 (Suppl 1):S277–S292

Table 2 Different errors

Sl. No. Algorithm type MSE RMSE R-squared
obtained prediction of soil MC
1. Gradient descent 0.1111 0.0123 0.9999
2. Variable learning rate gradient descent 0.0468 0.0021 0.9999

Fig. 15 Complete required Soil MC distribution map in every region

Fig. 13 Soil MC distribution map using Variable Learning Rate of the farm land using bilinear interpolation with finer grid technique
Gradient Descent neural network pattern classification training

Fig. 16 Soil MC deficiency map using SSIM index

Fig. 14 Complete soil MC distribution map at every region of the
farm land using bilinear interpolation with finer grid technique Secondly, the GD optimization-based NN pattern clas-
sification training is performed to analyze training perfor-
The soil MC prediction performance in addition to error mance. The soil MC prediction response and error graph of
graph is shown in Fig. 7. GD-based NN pattern classification model is shown in
Actual soil moisture content and predicted soil moisture Fig. 10. The GD optimization-based NN pattern classifi-
content versus time in hours error graph plotted between cation training performance is shown in Fig. 11. The
resuduals and time in hours [variable learning rate gradient accurate training is performed after 2000 epochs.
descent-based neural network pattern classification]. The training performance of Gradient Descent optimiza-
Similarly, the performance of VLRGD-based NN pat- tion based neural network pattern classification technique.
tern classification training is shown in Fig. 8. The accurate Furthermore, the validation check of GD optimization
training performance is found after 2000 epochs. The for water requirement is shown in Fig. 12. Then, the per-
training state of VLRGD-based NN pattern classification is formance analysis is established by evaluating general
shown in Fig. 9. It is observed that the best training was statistical errors like MSE, R-squared value and RMSE
performed after 1999 epochs. which is shown in Table 2. The MSE and RMSE are
The training state of the variable learning rate gradient required to analyze the behavior of the two prediction
descent neural network pattern classification training. techniques during the estimation of required soil moisture

123
Neural Computing and Applications (2019) 31 (Suppl 1):S277–S292 S287

(a) (b) L H
NL NR PL 1
1

0.8

Degree of membership
0.8
Degree of membership
0.6 0.6

0.4 0.4

0.2 0.2

0 0

-100 -80 -60 -40 -20 0 20 40 60 80 100

0 10 20 30 40 50 60 70
Soil MC Difference
Environment Temperature

(c) (d) LP DPDD CS SC VS SCS

L H 1
1

0.8

Degree of membership
0.8
Degree of membership

0.6 0.6

0.4 0.4

0.2 0.2

0
0

0 10 20 30 40 50 60 70 80 90 100
0 20 40 60 80 100 120 140 160 180 200
Environment Humidity
Wind Speed

(e) CL MR HR (f) L N H ST DP LP
1 1
Degree of membership

0.8 0.8
Degree of membership

0.6 0.6

0.4 0.4

0.2 0.2

0 0

-4 -3 -2 -1 0 1 2 3 4 5 2 4 6 8 10 12 14 16 18 20
valve sms

(g) (h)
0.5 11.6

0 11.4

11.2
valve

-0.5
sms

-1 11

10.8
-1.5
10.6
12 100
10 200
8 150 200
x 10
4 6 150
4 100 50
2 50 100
sunlight 0 50
windspeed 0 0
Envronment Humidity windspeed
Fig. 17 Membership functions of the inputs and outputs for fuzzy d Triangular membership functions for wind speed, e Triangular
inference system as well as surface plot between different inputs and membership function for valve position, f Triangular membership
outputs. a Triangular membership function for soil MC difference, function for SMS type, g Surface plot of wind speed, sunlight
b Triangular membership function for environmental temperature, intensity versus valve position, h Surface plot of wind speed,
c Triangular membership function for environmental humidity, environmental humidity versus sms_index

123
S288 Neural Computing and Applications (2019) 31 (Suppl 1):S277–S292

Valve ON/OFF region of the sample area • Similarly, the variable learning rate gradient descent
0.5 1
(VLRGD) produces less MSE and RMSE compared to
0.9 GD-based neural network soil moisture content predic-
1 OFF OFF OFF
0.8
tion algorithm.

0.7 It is found from Table 2 that the performance of the

1.5
VLRGD is far better than simple GD-based neural network
0.6
pattern classification method. This method is used to cal-
2 OFF OFF OFF 0.5 culate MC in 1 h advance of all the farm regions wherever
IoT-enabled WSN nodes are connected. The predicted MC
0.4
using VLRGD-based neural network pattern classification
2.5
0.3 of all the regions is stored for generating soil MC distri-
0.2
bution map in 1 h advance as shown in Fig. 13.
3 OFF OFF ON Soil MC distribution map using variable learning rate
0.1
gradient descent neural network pattern classification
3.5 0 training.
0.5 1 1.5 2 2.5 3 3.5
The soil MC distribution map data are used for bilinear
Fig. 18 Valve status in nine regions of the farm land and right side interpolation for generating complete soil MC distribution
corner region has low soil MC level map at each location of the farm land. The complete soil
MC distribution map is shown in Fig. 14. The required soil
content. Less is the MSE and RMSE, more better is the MC distribution of the farm land is shown in Fig. 15.
performance of the prediction algorithm. Similarly, Complete required soil MC distribution map in every
R-squared value more closely to 1 represents the better region of the farm land using bilinear interpolation with
estimation of the target values during the training of neural finer grid technique.
network algorithm. The SSIM index-based soil MC deficiency map is gen-
• Gradient descent (GD) has higher MSE and RMSE than erated by considering interpolated soil MC distribution
VLRGD-based neural network soil moisture content data and required soil MC distribution data. The SSIM
prediction algorithm. index-based soil MC deficiency map is shown in Fig. 16.
Furthermore, the soil MC deficiency data of the agriculture

Table 3 Few fuzzy logic rules established for SMS notifications and irrigation valve control
1. (soilmc_diff ==PL) & (env_temp ==H) & (env_humidity ==L) & (windspeed ==LP) & (sunlight ==BR) =[ (valve = HR)(sms = L) (1)
2. (soilmc_diff ==PL) & (env_temp ==H) & (env_humidity ==L) & (windspeed ==DP) & (sunlight ==BR) =[ (valve = HR)(sms = L) (1)
3. (soilmc_diff ==PL) & (env_temp ==H) & (env_humidity ==L) & (windspeed ==DD) & (sunlight ==BR) =[ (valve = HR)(sms = L) (1)
4. (soilmc_diff ==PL) & (env_temp ==H) & (env_humidity ==L) & (windspeed ==LP) & (sunlight ==BS) =[ (valve = HR)(sms = L) (1)
5. (soilmc_diff ==NL) & (env_temp ==L) & (env_humidity ==H) & (windspeed ==DD) & (sunlight ==SR) =[ (valve = CL)(sms = H) (1)
6. (soilmc_diff ==NR) & (env_temp ==L) & (env_humidity ==H) & (windspeed ==LP) & (sunlight ==SI) =[ (valve = MR)(sms = N) (1)
7. (soilmc_diff ==NR) & (env_temp ==L) & (env_humidity ==H) & (windspeed ==DP) & (sunlight ==SI) =[ (valve = MR)(sms = N) (1)
8. (soilmc_diff ==NR) & (env_temp ==L) & (env_humidity ==H) & (windspeed ==DD) & (sunlight ==SI) =[ (valve = MR)(sms = N) (1)
9. (env_temp ==L) & (env_humidity ==H) & (windspeed ==LP) & (sunlight ==TP) =[ (valve = CL)(sms = LP) (1)
10. (env_temp ==L) & (env_humidity ==H) & (windspeed ==LP) & (sunlight ==EX) =[ (valve = CL)(sms = LP) (1)
11. (env_temp ==L) & (env_humidity ==H) & (windspeed ==LP) & (sunlight ==SR) =[ (valve = CL)(sms = LP) (1)
12. (windspeed ==CS) | (sunlight ==EX) =[ (valve = CL)(sms = ST) (1)
13. (windspeed ==SC) | (sunlight ==EX) =[ (valve = CL)(sms = ST) (1)
14. (windspeed ==VS) | (sunlight ==EX) =[ (valve = CL)(sms = ST) (1)
15. (windspeed ==SCS) | (sunlight ==EX) =[ (valve = CL)(sms = ST) (1)
16. (env_humidity ==H) & (windspeed ==CS) & (sunlight ==EX) =[ (valve = CL)(sms = ST) (1)
17. (env_humidity ==H) & (windspeed ==SC) & (sunlight ==EX) =[ (valve = CL)(sms = ST) (1)
18. (env_humidity ==H) & (windspeed ==VS) & (sunlight ==EX) =[ (valve = CL)(sms = ST) (1)
FIS (fuzzy inference system) input sequence: [soil MC difference (soilmc_diff), environment temperature (env_temp), environment humidity
(env_humidity), wind speed, sunlight intensity]
FIS output sequence : [valve position, SMS index]

123
Neural Computing and Applications (2019) 31 (Suppl 1):S277–S292 S289

Table 4 Membership functions of the fuzzy logic model

Soil MC difference Membership index Description Environment temperature Membership index Description

NL 1 Negative large L 1 Low condition (\ 35)

NR 2 Normal H 2 High condition ([ 35)
PL 3 Positive large
Environment humidity Membership index Description Wind speed Membership index Description

L 1 Low humidity LP 1 Low pressure

H 2 High humidity DP 2 Depression
DD 3 Deep depression
CS 4 Cyclonic storm
SC 5 Severe cyclonic storm
VS 6 Very severe cyclonic storm
SCS 7 Super cyclonic storm

Sunlight intensity Membership index Description Valve control Membership index Description

BR 1 Brightest CL 1 Close
BS 2 Bright sunlight MR 2 Medium regulated
SR 3 Sunrise and sunset clean sky HR 3 High regulated
SI 4 Shade illuminated by clear sky, midday
TP 5 Typical overcast day
EX 6 Extreme dark, storm, cloud, midday

SMS events Membership index Description

L 1 Irrigation is required
N 2 Irrigation is not required
H 3 Irrigation overflow, and extraction pump is required to be activated
ST 4 Cyclonic storm and the valve is closed
DP 5 Deep depression occurred and valve is closed
LP 6 Low pressure occurred and valve is closed

area are impressed to fuzzy logic-based weather module to

generate actual irrigation valve commands (Fig. 17).
Valve ON/OFF region of the sample area
0.5 1 The membership functions defined in the fuzzy logic
model are shown in Fig. 18a–f. Further, the surface plot of
0.9
1 OFF OFF OFF
the membership functions is also shown in Fig. 18g, h. The
0.8 rules are listed in Table 3. All the parameters for the fuzzy
0.7 version are recorded in Table 4.
1.5
Finally, the irrigation valve control commands are
0.6
generated by considering fuzzy inference system with
2 OFF ON OFF 0.5 SSIM index-based soil MC deficiency data and the valve
status of nine regions of the farm land is shown in Figs. 18
0.4
2.5
and 19. The net irrigation from 01-Feb-2018 to 12-May-
0.3 2018 and irrigation status for 02-Feb-2018 is shown in
0.2 Fig. 20. A sample valve control command matrix for single
3 OFF OFF ON
region soil MC deficiency is shown in Table 5. It is
0.1
observed from Table 5 that the valve control commands
3.5 0 are generated in a matrix format where 0 and 1 are used.
0.5 1 1.5 2 2.5 3 3.5
The ‘0’ represents OFF condition of the valve, and ‘1’
Fig. 19 Valve status in nine regions of the farm land. Right side represents the ON condition of the valve.
corner and center region are having low soil MC level

123
S290 Neural Computing and Applications (2019) 31 (Suppl 1):S277–S292

Net Irrigation from 01-Feb-2018 to 12-May-2018

2
Irrigation Pump Activation

1.5

0.5

0
Feb 01, 18 Feb 08, 18 Feb 16, 18 Feb 24, 18 Mar 03, 18 Mar 11, 18 Mar 19, 18 Mar 26, 18 Apr 03, 18 Apr 11, 18 Apr 18, 18 Apr 26, 18 May 04, 18 May 12, 18
Date

Net Irrigation on 02-Feb-2018

1.5
Irrigation Pump Activation

0.5

0
0 5 10 15 20
Time

Fig. 20 Net irrigation from 01-Feb-2018 to 12-May-2018 and irrigation status for 02-Feb-2018

bilinear interpolation and SSIM-based approach. The irri-

Table 5 Sample valve control command matrix for soil MC defi-
gation valve control commands are generated successfully
ciency in different regions
with fuzzy logic weather model to fulfil uniform farm
Sl. Irrigation valve control Soil MC deficiency region irrigation requirement almost all the weather conditions.
no. command matrix

1. 000 Right corner region is having low

000 soil MC level
001
References
2. 000 Right corner and center region are
1. Panchard J, Hubaux J-P (2008) Wireless sensor networks for
010 having low soil MC level
marginal farming in India, Ph.D. thesis, EPFL, Lausanne. http://
001 dx.doi.org/10.5075/epfl-thesis-4172
2. Gupta D, Julka A, Jain S, Aggarwal T, Khanna A, de Albu-
querque VHC (2018) Optimized cuttlefish algorithm for diagno-
8 Conclusion sis of parkinson’s disease. Cogn Syst Res 52:36–48. https://doi.
org/10.1016/j.cogsys.2018.06.006 (ISSN: 1389-0417 SCIE (IF
1.18))
The valve control is done successfully by contemplating 3. Khanna A, Singh AK, Swaroop A (2016) A token based solution
soil MC prediction technique utilizing VLRGD-based to group local mutual exclusion problem in mobile Ad Hoc
optimization algorithm. A comparative analysis between networks. Arab J for Sci Eng 41(12):5181–5194 (ISSN
2193-567X)
VLRGD and GD is also formed by acquiring MSE, RMSE 4. Lakshmanaprabu SK, Shankar K, Gupta D, Khanna A, Rodrigues
and R-squared value. During this test, we found that JJPC, Plácido RP, de Albuquerque VHC (2018) Ranking analysis
VLRGD has a much reduced MSE and RMSE such as for online customer reviews of products using opinion mining
0.0468 and 0.0021, respectively. Similarly, GD-based NN with clustering. Complexity 2018:1–9. https://doi.org/10.1155/
2018/3569351 (SCIE (IF 4.62))
prediction system produces much higher MSE and RMSE 5. Ashish Khanna (2017) Image segmentation and edge detec-
values such as 0.1111 and 0.0123, respectively. The com- tion neuro-fuzzy logic system. Int J Adv Res Trends Eng Technol
prehensive soil MC requirement map is obtained using (IJARTET) 47–54 (ISSN 2394-3785)

123
Neural Computing and Applications (2019) 31 (Suppl 1):S277–S292 S291

6. Rodrigues JJPC, Segundo DPDR, Junqueira HA, Sabino MH, Microw Technol (MECS) 5(4):1–14 [ISSN: 2076–1449 (Print),
Prince RM, Al-Muhtadi J, De Albuquerque VHC (2018) Enabling ISSN: 2076–9539 (Online)]
technologies for the internet of health things. IEEE Access, vol 6, 22. Gómez-Melendez D, Lopez-Lambrantilde A, Herrera-Ruiz G,
pp 13129–13141 Fuentes C, Rico-Garcia E, Olvera-Olvera C, Alaniz-Lumbrerasc
7. da Cruz MA, Rodrigues JJPC, Al-Muhtadi J, Korotaev VV, de D, Teobaldis MF, Verlinden S (2011) Fuzzy irrigation green-
Albuquerque VHC (2018) A reference model for internet of house control system based on a field programmable gate array.
things middleware. IEEE Internet Things J 5:871–883 Afr J Agric Res 6(11):2544–2557
8. Lakshmanaprabu SK, Shankar K, Khanna A, Gupta D, Rodrigues 23. Jia X, Yao L, Zhang Y (2011) Design of field integrative irri-
JJPC, Pinheiro PR, De Albuquerque VHC (2018) Effective fea- gation control system based on fuzzy control and PlC. Commun
tures to classify big data using social internet of things. IEEE Comput Inf Sci 237:295–301
Access, vol 6, pp 24196–24204 24. Schulze K, Spreer W, Keil A, Ongprasert S, Müller J (2013)
9. Mahmoud MME, Rodrigues JJPC, Ahmed SH, Shah SC, Al- Mango (Mangifera indica L. cv. Nam Dokmai) production in
Muhtadi JF, Korotaev VV, De Albuquerque VHC (2018) Northern Thailand—costs and returns under extreme weather
Enabling technologies on cloud of things for smart healthcare. conditions and different irrigation treatments. Agric Water
IEEE Access, vol 6, pp 31950–31967 Manag 126:46–55
10. Hussein AF, Kumar A, Burbano-Fernandez M, Ramirez-Gonza- 25. Poch-Massegú R, Jiménez-Martı́nez J, Wallis KJ, de Cartagena
lez G, Abdulhay E, de Albuquerque VHC (2018) An automated FR, Candela L (2014) Irrigation return flow and nitrate leaching
remote cloud-based heart rate variability monitoring system. under different crops and irrigation methods in Western
IEEE Accessed 2018 Mediterranean weather conditions. Agric Water Manag 134:1–13
11. Liang R, Ding Y, Zhang X, Zhang W (2008) A real-time pre- 26. Migliaccio KW, Morgan KT, Fraisse C, Vellidis G, Andreis JH
diction system of soil moisture content using genetic neural- (2015) Performance evaluation of urban turf irrigation smart-
network based on annealing algorithm. In: IEEE international phone app. Comput Electron Agric 118:136–142
conference on automation and logistics (ICAL-2008), Computer 27. Khanna A, Singh AK, Swaroop A (2016) h-group local mutual
and Information Engineering College, Hohai University, exclusion algorithm in MANETs, CSI Transaction on the ICT
Changzhou, 1–3 Sept, pp 2781–2785 (Springer), pp 227–234, ISSN 2277–9078
12. Khanna A, Singh AK, Swaroop A (2014) A leader-based k-local 28. Lenka SK, Mohapatra AG (2016) Gradient descent with
mutual exclusion algorithm using token for MANETs. J Inf Sci momentum based neural network pattern classification for the
Eng 30(5):1303–1319 (SCI) (IF 0.54) prediction of soil moisture content in precision agriculture. In:
13. Gupta D, Shirish Sundaram, Ashish Khanna, Aboul Ella Has- Proceedings of the IEEE international symposium on nanoelec-
sanien, Victor Hugo C de Albuquerque (2018) Improved diag- tronic and information systems (iNIS), 17 March 2016, pp 63–66
nosis of parkinson’s disease based on optimized crow search 29. Zhengjun Q, Xiaoxing T, Jiehui S, Yidan B (2007) Irrigation
algorithm. Comput Electr Eng, SCIE (IF 1.57) decision-making system based on the fuzzy-control theory and
14. Mohapatra AG, Lenka SK (2016) Hybrid decision model for virtual instrument. Nongye Gongcheng Xuebao/Trans Chin Soc
weather dependent farm irrigation using resilient backpropaga- Agric Eng 23(8):165–169
tion based neural network pattern classification and fuzzy logic. 30. Mohapatra AG, KeswaniB Lenka SK (2018) ICT specific tech-
In: Proceedings of the Springer smart innovation, systems and nological changes in precision agriculture environment. Int J
technologies (SIST) Book series, Chapter 30, pp 1–12 Comput Sci Mob Appl 6:1–16
15. Mohapatra AG, Keswani B, Lenka SK (2018) Neural network 31. Svozil D, KvasniEka V, Pospichal J (1997) Introduction to multi-
and fuzzy logic based smart DSS model for irrigation notification layer feed-forward neural networks. Chemom Intell Lab Syst
and control in precision agriculture. In: Proceedings of the 39:43–62
National Academy of Sciences, India Section A: Physical Sci- 32. Phillips AJ, Newlands NK, Liang SHL, Ellert BH (2014) Inte-
ences, Springer, Berlin, pp 1–10. https://doi.org/10.1007/s40010- grated sensing of soil moisture at the field scale: measuring,
017-0401-6 modeling and sharing for improved agricultural decision support.
16. Mohapatra AG, Keswani B, Lenka SK (2017) Optimizing farm Comput Electron Agric 104:73–88
irrigation mechanism using feedforward neural networkand 33. Mohapatra AG, Lenka SK (2016) Hybrid decision support system
structural similarity index. Int J Comput Appl 4(7):135–141 using PLSR-fuzzy logic for GSM based site specific irrigation
17. Mohapatra AG, Keswani B, Lenka SK (2017) Soil n-p-k pre- notification and control in precision agriculture. Int J Intell Syst
diction using location and crop specific random forest classifi- Technol Appl Indersci 15(1):4–18
cation technique in precision agriculture. Int J Adv Res Comput 34. Forouzanfar M, Dajani HR, Groza VZ, Bolic M (2010) Com-
Sci 8:1–6 parison of feed-forward neural network training algorithms for
18. Lakshmanaprabu SK, Shankar K, Khanna A, Gupta D, Rodrigues oscillometric blood pressure estimation. In: 4th international
JJPC, Plácido RP, de Albuquerque VHC (2018) Effective fea- workshop on soft computing applications (SOFA), Sch. of IT and
tures to classify big data using social internet of things. IEEE Eng., Univ. of Ottawa, Ottawa, ON, Canada, 15–17 July 2010,
Accessed, SCIE (IF 3.24) pp 119–123
19. Majone B, Viani F, Filippi E, Bellin A, Massa A, Toller G, Robol 35. Zhang M, Li M, Wang W, Liu C, Gao H (2013) Temporal and
F, Salucci M (2013) ‘‘WSN deployment for monitoring soil spatial variability of soil moisture based on WSN. Math Comput
moisture dynamics at the field scale, International Conference on Model 58(3–4):826–833
Four Decades of Progress in Monitoring and Modeling of Pro- 36. Popa CA (2014) Enhanced gradient descent algorithms for
cesses in Soil Plant Atmosphere System: Applications and complex-valued neural networks. In: 16th international sympo-
Challenges’’, Procedia Environmental Sciences, vol 19, sium on symbolic and numeric algorithms for scientific com-
pp 426–435 puting. SYNASC 2014, Dept. of Comput. and Software Eng.,
20. Water Sense Labeled Weather-Based Irrigation Controllers Polytech. Univ. Timisoara, Timisoara, Romania, 22–25 Sept,
(2012) WTR-SENS. pp 987–7367. http://www3.epa.gov/water pp 272–279
sense/docs/irrigation_controller_rpt_minireport_508.pdf 37. Zhang G-L, Wang Y, de Silva CW (2008) Multisensor gripper
21. Khanna A, Singh AK, Swaroop A (2015) Dynamic request set positioning in unstructured urban environments using neural
based mutual exclusion algorithm in MANETs. Int J Wirel network. In: IEEE international conference on automation and

123
S292 Neural Computing and Applications (2019) 31 (Suppl 1):S277–S292

logistics (ICAL), Dept. of Mech. Eng., Univ. of British Columbia, 39. Rehman A, Wang Z (2011) SSIM-based non-local means image
BC Vancouver, 1–3 Sept, pp 1474–1479 denoising. In: 18th IEEE international conference on image
38. Gupta A, Reddy BVR, Ghosh U, Khanna A (2012) A permission- processing, Department of Electrical and Computer Engineering,
based clustering mutual exclusion algorithm for mobile Ad-Hoc University of Waterloo, Waterloo, ON, Canada, pp 217–220
networks. Int J Eng Res Appl (IJERA) 2(4):19–26 (ISSN:
2248–9622)

Affiliations

Bright Keswani1 • Ambarish G. Mohapatra2 • Amarjeet Mohanty3 • Ashish Khanna4 • Joel J. P. C. Rodrigues5,6,7 •

Deepak Gupta4 • Victor Hugo C. de Albuquerque8

1 5
Department of Computer Applications, Suresh Gyan Vihar National Institute of Telecommunications (Inatel), Instituto
University, Mahal Jagatpura, Jaipur, India de Telecomunicações, Santa Rita do Sapucaı́/MG, Brazil
2 6
Department of Electronics and Instrumentation Engineering, Instituto de Telecomunicações, Lisbon, Portugal
Silicon Institute of Technology, Bhubaneswar, Odisha, India 7
ITMO University, St. Petersburg, Russia
3
Department of Information Technology, Silicon Institute of 8
Graduate Program in Applied Informatics, University of
Technology, Bhubaneswar, Odisha, India
Fortaleza, Fortaleza, CE, Brazil
4
Maharaja Agrasen Institute of Technology, GGSIP
University, Dwarka, India

123