Classi cation and Yield Prediction in Smart

Agriculture System Using IoT

Akanksha Gupta (  [email protected] )
University of Delhi https://orcid.org/0000-0001-7253-3648
Priyank Nahar
Venkateshwara Institute of Management

Research Article

Keywords: IoT, Sensors, Agriculture, Machine learning, Classi cation, Crop yield prediction.

Posted Date: February 21st, 2022

DOI: https://doi.org/10.21203/rs.3.rs-830669/v1

License:   This work is licensed under a Creative Commons Attribution 4.0 International License.
Read Full License

Version of Record: A version of this preprint was published at Journal of Ambient Intelligence and
Humanized Computing on January 5th, 2022. See the published version at
https://doi.org/10.1007/s12652-021-03685-w.
 Classification and Yield Prediction in Smart Agriculture System Using IoT
Ms. Akanksha Gupta*, Department of computer science, University of Delhi
Email: [email protected]
Dr. Priyank Nahar, Shri Venkateshwara University, Gajraula, India.
Email : [email protected]

Abstract evaluation of the proposed prediction

The Modern agriculture industry is data- model. Various metrics are considered for
centred, precise and smarter than ever. the performance evaluation such as
Advanced development of Internet-of- accuracy, RMSE, R2, MSE, MedAE,
Things (IoT) based systems redesigned MAE, MSLE, MAPE and Explained
“smart agriculture”. This emergence in Variance Score (EVS).
innovative farming systems is gradually
enhancing the crop yield, reduces Keywords: IoT, Sensors, Agriculture,
irrigation wastages and making it more Machine learning, Classification, Crop
profitable. Machine learning (ML) yield prediction.
methods achieve the requirement of
Declarations:
scaling the learning performance of the
model. This paper introduces a hybrid ML Funding - Not applicable.
model with IoT for yield prediction. This
work involves three phases : pre-
Conflicts of interest/Competing
processing, feature selection(FS) and
interests (include appropriate
classification. Initially, the dataset is pre- disclosures) - Not applicable
processed and FS is done on the basis of
Correlation based FS (CBFS) and the
Variance Inflation Factor algorithm (VIF). Availability of data and material -
Finally, a two-tier ML model is proposed Available at soilhealth.dac.gov.in
for IoT based smart agriculture system. In
the first tier, the Adaptive k-Nearest Code availability - Custom code
Centroid Neighbour Classifier (aKNCN)
model is proposed to estimate the soil Ethics approval - Not applicable
quality and classify the soil samples into
different classes based on the input soil Consent to participate - The
properties. In the second tier, the crop participant has consented to the
yield is predicted using the Extreme submission of the case report to the
journal
Learning Machine algorithm (ELM). In the
optimized strategy, the weights are
Consent for publication -
updated using modified Butterfly
I give my consent for the publication
Optimization algorithm (mBOA) to of identifiable details, which can
improve the performance accuracy of include photograph(s)
ELM with minimum error values. and/or videos and/or case history
PYTHON is the implementation tool for and/or details within the text to be
evaluating the proposed system. Soil published in the Journal.
dataset is utilized for performance
1. Introduction the data remotely for decision making. The
IoT is an advanced technology for main advantage is that these techniques are
monitoring and controlling devices low cost; hence it is affordable for all
anywhere in the world. In many fields, it farmers [9].
creates a remarkable mark due to its easy
accessibility [1]. Few IoT developed Agricultural supervision, particularly crop
technologies such as remote sensors, yield observation is essential for
drones and robots made human’s life examining the food security in a region.
simpler and beneficial. Moreover, these Due to several difficult aspects, predicting
technologies have experimented on the crop yield manually is a challenging
fundamental needs such as food that is task [10]. Based on the water quality as
obtained from the agricultural field [2, 3]. well as availability, pest infestations,
From the recent survey of World Bank, it genotype, landscape, soil quality, climatic
is approximated that more than 50% food condition, etc., the yield of the crop may
is required to cultivate before 2050 based vary. The strategies and the processes are
on present population rate [4]. However, non-linear in nature, intricate and varied
such huge production of the crop is a with time because of external aspects and
challenging task because of the current correlated factors [11]. Recently, several
climatic changes. In such cases, Smart studies illustrate that ML approaches such
agriculture system plays a vital role to as support vector regression, multilayer
increase the yield by monitoring and perceptron (MLP), etc. have comparatively
predicting the production of the crops [5]. more enhanced potential than the
IoT based crop yield prediction enables the traditional techniques. These approaches
farmers to enhance productivity. In have the ability to predict linear and non-
general, IoT based smart farming system is linear agricultural architecture. From the
deployed in an agriculture field for learning process, these methods were
monitoring the crop field with the help of obtained in ML agricultural framework
sensors namely DHT11 (temperature and [12-14].
humidity sensor), TOC (Total Organic On observing the most outstanding
Carbon) and nitrogen, phosphorus, and frameworks in agriculture, artificial and
potassium (NPK) sensors [6]. Using this deep neural networks are the commonly
setup, farmers can monitor the field utilized models [15]. Artificial neural
conditions from anywhere. Gateways are network (ANN) is a network model that
responsible for receiving data from the generates approximation by bias as well as
crop area and forward them to the storage weight optimization for a node-link
unit. The prediction engine is used to structure including input, hidden and
predict the results and sends information to output layers [16]. Deep learning (DL) is a
the notification server [7]. Therefore, subgroup of ML used to predict the crop
various IoT based techniques can assist the yield based on the varying arrangement of
farmers to produce huge crops and the raw data via the intensive learning process
techniques used are sensors, hydroponic in a deep network [17]. Moreover, these
farming, advanced tractors and drones [8]. DL algorithms have the capability to
These devices assist agriculturists and design a probability model using field
researchers in analyzing and processing
data. Along with this benefit, the data machine learning models have become
about the crop performance under different popular that enable researchers for solving
climatic changes are provided by DL and understanding complex predictions.
approaches [18]. For example, Many ML approaches have been employed
Reinforcement learning is one of the major for crop prediction like decision trees,
areas of artificial intelligence. It is the ANN and Support vector machine (SVM)
preparation of ML models for decision- and in this research work ELM and
making sequences and is the significant aKNCN are used for crop yield prediction
class of algorithm that is used to thereby achieving better results.
streamline logic for dynamic programming
[19]. Besides, extreme learning machine Contributions:
(ELM) is also a ML approach that has the The proposed work’s major contributions
capability to empower neural network are:
training for predicting the crop yield. It
accelerates the learning process and  The proposed work provides an
provided better outcomes. However, these IoT based farming system that
approaches have several disadvantages like ensures the deployment of effective
less sustainability, computationally crop yield prediction model. This
expensive, high complexity and false work involves pre-processing, FS
prediction [20]. To overcome these and classification. The data is pre-
challenges, a more efficient ML based processed and features are selected
crop yield prediction model is proposed in by FS algorithms.
this work.  Then IoT based smart agriculture
system using two-tier ML model is
Motivation: proposed for the better prediction
Agriculture is the major economic of crop yield.
resource of India. To overcome the issues  ML based classification is
of high cost and complex management of performed to classify the soil
conventional agricultural planting, IoT is samples for different classes by
applied for realizing real time detection, considering the properties of soil
crop growth intelligent management and dataset.
changing the conventional agricultural  ELM model is proposed for
planting mode. Various mathematical and predicting the crop yield and the
empirical yield approaches have been weights of the model are updated
evaluated for several crops. These models using mBOA algorithm for the
need huge knowledge about soil and crops performance improvement of the
which make it hard for implementing for prediction system with less error
various localities. Many satellite based values.
remote sensing methods were also  The systematic evaluation of the
developed in yield modelling. But these proposed crop yield prediction
approaches are not able to provide enough system is implemented based on
spatial details of small farms for machine learning performances
optimizing crops. Recent developments in through various experiment
consequences.
Paper outline: Section 1 presents the optimal features for further classification.
introduction and highlights smart PART was a partial decision tree approach
agriculture in brief. Recent related works used for classification and prediction.
during (2019-2021) are discussed in Accuracy, precision, sensitivity and F1
section 2. Section 3 focuses on the score were considered for the experiment
problems in the IoT based crop yield of WPART. The crops taken for
prediction. Section 4 provides the detailed experiment were Sugarcane, Jowar, Bajra
description about IoT devices used in a and Soybean. Some samples in the dataset
smart agriculture system. Section 5 were misleadingly labelled, thus the false
presents the proposed framework along prediction rate was high. In future, some
with important measurables. Section 6 parameters like soil nutrients, agricultural
discusses the experimental analysis and inputs, irrigated area and soil quality will
results. Section 7 concludes the presented be considered for forecasting the crop
work. yield.

2. Related works Bu, et al. [23] developed deep

reinforcement learning based ML
Abbas, et al. [21] predicted the crop yield technique for smart farming IoT system.
via proximal sensing and the ML Cloud computing and artificial intelligence
algorithms. The objective was to extract were combined for the classification and
significant data that are responsible for prediction of crop yield. The key goal of
controlling the yield of crop. The this research was to minimize resource
properties of potato tuber crop and the data consumption and maximize the food
of soil have been gathered by proximal production. A hierarchical Bayesian based
sensing. A large dataset was utilized for multi-task reinforcement learning method
the prediction performance. Support vector has been utilized for modelling the Markov
regression (SVR), k-nearest neighbour (K- decision process. Then, the Q-value
NN), linear regression and elastic net ML regression function was examined using
algorithms were utilized for the policy distillation. However,
classification and prediction of crop yield. computational complexity was considered
R2, MAE and RMSE were determined for as one of the major drawback of this
the performance evaluation. The approach. Also, human-level performance
performance achieved for KNN was poor was not achieved in complex task solution
in crop yield prediction because of higher and in adaptation to dynamic
number of features. environments. In future, this research aims
Rezk, et al. [22] presented an IoT based to design an incremental model and
smart agriculture system using ML transfer learning approaches for the
algorithm. The drought and the crop enhancement of performance efficiency.
productivity were predicted by WPART Nevavuori, et al. [24] proposed a deep
and it was a combination of wrapper and learning technique for crop yield
PART techniques. Feature selection and prediction. The key objectives of this
classification were the two important research were crop yield prediction,
phases in the prediction process. Wrapper biomass evaluation, crop and weed
feature selection technique selected the
detection. Convolutional Neural Network network was updated by the PBIL
(CNN) was modelled for extracting the approach. Hadoop framework has been
features, training, hyperparameter tuning utilized for the prediction performance.
and regularization to predict the yield of Neural network along with ANN and
wheat and barley crops. MAE and MAPE multiple linear regressions (MLR) were
are the evaluation metrics used for implemented for the crop yield prediction.
simulation analysis. But, the presented Low convergence and getting stuck within
CNN does not perform well for the large local minimum were the major drawbacks
dataset. Also, the performance efficiency in this presented model. The future scope
of this method was not good. The future of this research is to use the optimization
scope of this research is to train the approaches for the crop yield forecasting.
developed model for a large set of features
like soil and climate with time series Filippi, et al. [27] proposed the empirical
modelling scheme for forecasting the yield
image data.
of barley, wheat and canola crops. In this
Dos Santos, et al. [25] introduced research, several fields are considered for
AgriPrediction model for IoT based smart the prediction performance instead of
agriculture system. It was an end-to-end single field in isolation. Random forest
model that predicted agricultural crops. It models and publically available data with
was the integration of prediction as well as temporal and spatial data collected on-
short and medium wireless network range farm were combined for the yield
system. The components of AgriPrediction prediction of canola, barley and wheat.
model have been designed according to the The experimental results showed that the
ARIMA prediction model and LoRa IoT accuracy obtained by this predictive model
technology. Initially, the data were was low. In future, this research will be
gathered using sensors, then the discrete extended by exploring enhanced feature
moving average-based prediction has been extraction calibration and more publically
performed. If the predicted crop goes available data sources for forecasting the
wrong, then the notification was given to yield of the crop.
the farmer’s mobile phone. This model
was computationally expensive and less Sun, Jie, et al. [28] predicted the yield of
sustainable. Moreover, the accuracy of this both in-season and end-of-season soybean
AgriPrediction model was not evaluated in using deep CNN-LSTM based on remote
this research. The future scope of this sensing data. The training data such as
research is to generate a mobile application MODIS surface Reflectance (SR) data,
for crop’s real-time monitoring. MODIS land surface temperature (LST)
data and weather data were correlated and
Saranya, C. P., and N. Nagarajan [26] transformed to histogram based tensors
presented a neural network with according to the Google earth engine
population based incremental learning (GEE). The performance of crop yield
(NN-PBIL) method for the prediction prediction at large scale was not evaluated
performance enhancement. The neural in this research. Time and computational
network was used to classify and predict complexity was high and fed the raw
the crop yield. The weight of the neural remote sensing data into DL mode was a
complex task. The future scope of this online prediction model in agri prediction
research is to include more features for the but it has high computational complexity.
yield prediction and the performance will ANN [26] achieves good prediction in
be performed at large scale. crop yield but handling noise in the images
has to be investigated. The method [27]
Sinwar, et al. [29] discussed about achieved higher accuracy but it has to be
different methods of Artificial Intelligence focussed on time series prediction. Deep
(AI)for smart irrigation and crop yield CNN-LSTM [28] model is highly efficient
prediction system. From the research it in error performance at any time node but
was proved that AI based system offered the processing time is large due to its
adequate data regarding the crop yields at complex structure. AI [29] minimizes the
an early stage. Several ML techniques effort of human and improves agricultural
such as Support vector machine (SVM), practices but the system cost is too high.
Linear Discriminant Analysis (LDA), Due to these drawbacks the proposed
ANN, Generative Adversarial network model introduces 2 classification models
(GAN), Deep Boltzmann machine (DBN), with optimal weight selection for crop
K nearest neighbour (KNN), yield prediction.
Backpropagation neural network
(BNN)and Deep neural network (DNN)
4. IoT Devices
performed better in crop yield prediction. The IoT based smart agriculture model is
However, some limitations are there in designed to construct the whole crop
these approaches like computational prediction system. The key goal of this
complexity, high cost, dependencies model is to predict the crop yield based on
between target and input variables, proper the gathered data using IoT devices. It
model representation and accuracy assists the crop yield prediction system for
affected by data quality. better decision making on the yield of crop
3. Problem Statement according to the gathered data from
monitoring camera analysis. The data
The method in [21] attained better obtained from camera is integrated with
performance but in some cases it achieves environment data from IoT devices.
poor result due to the small datasets. The Availability of environment data as well as
method [22] outperformed conventional its timely delivery is significant for
model for 5 datasets for classification of preserving the crops and property at the
drought and productivity of crop but it did time of disasters. The important
not focus on time series analysis. Deep information considered for an effective
reinforcement [23] showed better growth predictive model are Statistical agriculture
in model design but it takes more time for data and contemporary IoT sensing data.
training. CNN [24] achieves better The parameters available in statistical and
accuracy on yield prediction but does not contemporary IoT sensing data are rainfall,
support large features like soil and climate. temperature, pH level, soil nutrients and
The method [25] achieves better result in fertilizers. Figure 1 represents the structure
of IoT services for agriculture.
 Controller

Web

Soil Internet
sensor Environment Prediction
Transmi Receiver Environment database server
tter data handling
PDA
server
Temp.
sensor Data Acquisition and
Handling
IoT sensor

Laptop
Streaming
server
Camera
Data management Monitoring and
Prediction
Figure 1: Service of IoT for Agriculture

The major components in IoT based embedded with IoT system can improve
agriculture model are sensors, data the production of crop by predicting the
acquisition, data management devices, soil type, climatic changes, soil quality and
monitoring and prediction devices. These several other factors. Figure 2 illustrates
devices enhanced the performance of the the example picture representation of IoT
crop yield prediction. Smart farm based smart agriculture system.

Figure 2: Example pictorial representation of IoT based smart agriculture system:

(a) plant (b) moisture sensor (c) temperature and humidity sensor and (d) USP
 The water motor in the field is operated
Sensors: based on the threshold value, i.e., if the
Sensors in IoT are used to track the threshold value is greater than the water
particular soil and water pH value, level, then the needed water is supplied.
temperature and humidity value and Once the water level is attained, the sensor
fertilizer control for the growth of crop reading provides the indication and the
[32].The moisture, temperature and irrigation is stopped.
humidity and pH sensors are discussed in
the following subsections. (ii) Temperature and Humidity sensor:
This is used to measure pending rain,
(i) Soil moisture sensor: temperature changes, rain fallen at a
It is a sensor that determines the water particular period and water consumption. It
content available in soil. Several sensors is comprised of sense of wet NTC
are there in a probe, which save the water temperature measuring devices and
and manage irrigation system resistive element. This sensor provides fast
economically and successfully. Also, the response, high cost performance,
quality and yield of the crop is improved incredible quality and anti-interference
using these sensors. It provides the capability. Figure 4 illustrates the pictorial
accurate outcomes immediately and is less representation of a temperature and
expensive. Figure 3 shows the pictorial humidity sensor.
representation of a soil moisture sensor.

Figure 4: Temperature and humidity

Figure 3:Soil moisture sensor sensor
Using this sensor, the soil’s water content (iii) pH sensor:
can be examined and it averages the water
content over the complete span of the pH is an indicator of a solution’s acidity
sensor. The main purpose of this sensor is and alkalinity. Normally, the range of pH
to monitor the moisture of the soil for level is considered between 0 and 14.
irrigation management, examine the loss of Moreover, it represents the concentration
humidity over time because of plant uptake in certain hydrogen +ions solutions. Figure
and evaporation and to find the optimum 5 depicts the pictorial representation of a
soil moisture stuffing. The sensor readings pH sensor. The potential difference
are then transferred to the transmitter and between two electrodes such as hydrogen-
the threshold value is set in the application. sensitive glass electrode and reference
electrode is detected by this sensor. It can
be used with the microcontroller like Data acquisition:
Arduino. The initial process in data acquisition is to
register each device deployed in the farm
into Mobius. The registration is done
using &cube device and it acts as an
interface between the deployed devices
and the Mobius. Then, the virtual
representation is generated for each device
based on the resource type. The
environmental data gathered by sensors are
transmitted to &cube and the &cube
forwards the data to Mobius [33]. Finally,
the available virtual representations of the
devices are accessed by the farmers or
end-users to monitor and manage their
Figure 5: pH sensor connected farm with the help of IoT
Controller: applications such as tablets, laptops,
Arduino controller is used to monitor and smartphone, etc.
control the farm’s environmental (i) Mobius:
conditions. The inputs provided to the It is an open IoT service platform that
controller are location from the farm and performs operations based on one
the pH value. Besides that, the percentage machine-to-machine (M2M) standard
of nutrients such as Calcium, Zinc, rules. The generation of physical IoT
Manganese, Organic matters, Boron, devices’ virtual representations is assisted
Magnesium, Iron, Sulphur, Nitrogen, by this Mobius. It is designed to conform
Potassium and Phosphorous. Based on the with one M2M specifications and helps
pretrained network or real time values or general M2M/IoT service functions like
both, the decisions regarding the prediction security, device registration, subscription
are made by the controller. Figure 6 as well as notification and data repository
indicates the arduino controller. and management. Moreover, the data
resources preserved in Mobius are
accessed and the IoT devices are
controlled by REST APIs.
(ii) &cube
The &cube is a device software platform
that is installed into IoT gateways.
Through the standard REST APIs, the
gathered data from physical devices is
transmitted to Mobius. Several protocols
are assisted by &cube such as CoAP,
MQTT and HTTP. Raspberry-Pi is utilized
in this work for the connected farm. It is a
Figure 6: Arduino controller
single-board Linux installed computer.
 atmospheric temperature etc. play a vital
role in cultivation. ML methods are an
essential decision support device for the
prediction of crop yield like supporting
decision on what crops to grow. Many ML
algorithms are employed to support the
prediction of crop yield. In the proposed
smart agricultural framework, pre-
Figure 7: Raspberry-Pi processing, FS and a two-tier model is
Data Management: implemented for crop yield prediction.
In data management service, the required Here, akNCN is proposed to be deployed
external data are gathered from the sensors which is an improved version of KNCN.
and the servers preserve the collected data Generally ELM is an influential model
for further process. Also, the data gathered with more fast learning methods, higher
from camera is stored to predict the yield performance and less training error when
of the crop. Storage, maintenance and compared with other algorithms. Therefore
management of data for maintenance of these two classification algorithms are
corresponding services are provided in this proposed in this work to improve the
data management system. accuracy of the system and provide better
results than the existing models.
5. Proposed Framework
The ingredients of soil like Phosphorous,
Potassium and Nitrogen, crop rotation and

Crop
parameters
Database

Pre-
Gateway processing

Feature
selection
Crop Area

Prediction Engine Two tier machine

learning model
Data Data
Source Preparator Classification

User Evaluation Prediction

Prediction
Result Result

Notification Server
 Figure 8: Architecture of the proposed methodology

Figure 8 illustrates the architecture of the ML is a computational learning model that

proposed crop yield prediction model. works on prediction from statistical value.
Initially, Pre-processing is done to remove FS model is applied to identify necessary
the noise in data and the features are features which are powerful in correlation
selected on the basis of features selection with crop production. The main reason to
methods like CBFA and VIF. Finally, the employ FS is that it enables the ML
classification uses two tier systems. In the algorithm to train faster, minimizes the
first tier, the proposed aKNCN model is model complexity and makes it easy to
used to classify the soil quality based on interpret. It also increases the system
IoT system collected soil nutrients. Then accuracy when the proper subset is
in the second tier, ELM-mBOA is utilized selected and reduces overfitting. The
for crop yield prediction and the accuracy computation time of the algorithm is less
is improved by optimal weight selection necessary than its classification for normal
using mBOA. This model improves the size feature sets. But the feature selection
accuracy of the system with minimum is necessary for large datasets. Various
error values. statistical approaches can be employed in
FS like filter, embedded and wrapper
5.1 Pre-processing methods. Filter methods choose the
intrinsic characteristics of the features
The data is gathered from various sources computed by univariate statistics instead of
and pre-processing is done. Pre-processing performance of cross-validation. These
is a necessary phase in ML since it can’t methods are faster and less
handle noisy data. Noisy data means it has computationally expensive than wrapper
errors and outliers. Before applying the methods. When dealing with high-
data to classification it has to be pre- dimensional data, it is computationally
processed for inserting missing values, cheaper to use filter methods. Hence, in
eliminate unwanted data, functionality this work filter based FS methods like
extraction and maintain the appropriate CBFA algorithm and VIF algorithm are
data range. In this work isnull() approach used. CBFA chooses the best feature set
is used to check the null values then the which is mainly correlated with yield. VIF
label encoder() is used to convert verifies the multicollinearity among
categorical data (string format) into independent features. Therefore, it
numerical data (numeric format). Since eliminates all multicollinear features.
Python does not handle categorical data it
must be converted into numeric format. 5.2.1 Correlation based Feature
Once the data is converted to numeric selection Algorithm (CBFA)
format, it is applied for feature selection.
CBFS orders feature subset based on the
5.2 Feature selection(FS) correlation heuristic evaluation function.
This function is towards a subset that has
features which have high correlation
amongst class and uncorrelated with each completely orthogonal, R2 will be 0
other. The features which are not relevant resulting in VIF of 1.
must be removed since they have less
correlation amongst class and high 5.3Tier 1-Classification
correlation with other features. Feature In this work, aKNCN [34] is used to
acceptance is based on the extent to which classify the soil classes from the different
it identifies classes in areas which are parameters. The proposed aKNCN
already not identified by other features. overcomes the challenges of conventional
The CBFS is computed as: KNCN and enhanced the performance of
KNCN classification. KNCN is a non-
Nrc
M  (1) parametric classifier that depends on the
N  N ( N  1) rf centroid distance. This states the nearest
neighbours of the test samples should
Where N is the total number of features, satisfy the following criteria - it should be
rc is the average correlation, rf is average close to test samples and the nearest
feature to pair wise correlation. neighbour distribution should be
symmetrical in test samples. But it is
5.2.2 Variance Inflation Factor complex to determine neighbours in a
algorithm (VIF) feature that satisfy these properties.
Though KNCN achieves good accuracy it
VIF computes the strength of the multi
lags in classification time. Hence aKNCN
colinearity in the analysis of least squares
is developed for improving the
regression. It gives an index that computes
classification time by adaptively adjusting
how much the variance of an evaluated
the nearest centroid neighbour for every
regression coefficient is enhanced due to
input sample to enhance the classification
colinearity. VIF model is employed for
accuracy. Two properties of aKNCN are
removing correlated independent features.
given as follows:
This method is fast and it exploits one pass
search to the predictor. In addition, this Property 1:
method is computationally efficient in
testing every predictor to the model and it The aKNCN method satisfies a stable
avoids the overfitting issue. It is achieved searching phase only when j th distance of
by regressing each independent variable, nearest centroid is more than pre-defined
let Y on the remaining independent limit which is multiplier product of kl and
variables (W and Z) and checking how the first nearest centroid z ncn ,1 to the test
much of it (of Y) is explained by these
sample d ( y , z ncn ,1 ) . The size of
variables.VIF is measured by
1 neighbourhood is represented as
V (2)
1 R2

From the expression it is shown that the

d ( y, z is )  k1  d ( y, z ncn,1 ) (3)
higher the VIF, higher the R2 which means
the variable X is collinear with Y and Z
variables. If all the variables are
Where d ( y , zis ) is the nearest centroid convergence problems and provides
robustness ELM has learning speed and
distance among test samples, zis and y .
has a better generalization because there is
The multiplier product is higher or equal to
no need to tune the initial parameters of
1 and the 1st centroid distance is
the hidden layer. The hidden layer Feed
d ( y , z ncn ,1 ) . Forward Network (FFN) is converted into
the linear equation by minimum norm least
Property 2:
squares. The aim of the ELM is to reduce
The aKNCN method satisfies searching the output norm weight and training error
phase only when the entire sample class, at the same time. For the samples
M i , is found amongst j nearest neighbour {( Z i , Ti ) | X i  S m , Ti  S n , i  1, 2,... N } ,
and the whole samples per class to the P neurons hidden layer with the output
compete class is lesser than M i  1 . Then function is:
the property is defined as P
f P (Z )    i r (Z )  r (Z ) (5)
( MCM i (Vwi )  (Vwi ' ) i 1

(4) Where   [ 1 ,  2 .......... P ] is the output

( MCM i ( x )  {mcm xy | mcm xy  V } yj 1 is j weight vector between output neuron and
P . The hidden layer output vector to the
nearest neighbour of Vwi  { y iy | y iy  V }Myi1 input X is given by
is a subset of V from wi with training
r ( Z )  [r1 (Z ), r2 (Z )............rP ( Z )] (6)
samples, M i and Vwi '  { y iy | y iy  V }Myi1' is
a subset of V ' from wi with training For enhancing the generalization and to
samples, M i . Finally, the soil quality is reduce the training error of neural
classified and the crop yield is predicted networks, at the same time both output
based on the classified properties for weight and the training error must be
different classes. The yield prediction is minimized.
performed using ELM, which is discussed
min :|| h  T ||, ||  || (7)
in the next subsection.
According to (Karush–Kuhn–Tucker) the
5.4 Tier 2-Prediction
equation (19) can be written as
In this phase, ELM [36] is proposed to
1
predict the crop yield based on the 1 
  h T   hh T  T
classified soil properties of different R 
classes along with different parameters (8)
such as rainfall and temperature. In ELM,
a new metaheuristic algorithm called Where h is the output matrix of the
mBOA is hybridized to tune the optimal hidden layer, R is the coefficient of the
set of ELM parameters such as thresholds reflection and T is the expected samples
and weights that enhance the performance and the ELM algorithm output function is
accuracy with fast convergence. mBOA is
a novel approach, which solves the
 1 procedure. The s in the optimal solution
1 
f ( Z )  g ( Z )h T   hh T  T of the method is expressed as
R 
(9)
 0.025 
st 1  s t    (13)
When the feature mapping function r (Z ) is  ct .Tmax 
unknown, ELM kernel matrix on the basis
of Mercer’s condition is given by Where Tmax is the maximum iteration and
initial value of s is 0.01.
M  hh T , : mij  r ( Z i )r ( Z j )  L( Z i , Z j )
Further, there are two stages in the
(10)
process, global search and local search
The output function g (Z ) on the basis of space. The mathematical calculation of
KOELM is given by global search is calculated as

x tj1  x tj  (ri  g b  x tj ) f r (14)

2
1
1 
f ( Z )  L( Z , Z 1 )......L( Z , Z n )]   M  T
R 
(11) Here x tj is the solution vector x j of the
j th butterfly in iteration t and ri is a
Where L(Z , Z1 ) and M  hh are the T

random number and rages from [0,1]. Then

hidden neurons kernel function of single
gb is the present best solution identified
hidden layer FFN networks. The functions
like polynomial kernel, exponential kernel, among every stage in the present stage.
linear kernel and Gaussian kernel will Then the local search space is expressed as
satisfy the Mercer condition.
x tj1  x tj  (ri  xim  x tj ) f r (15)
2

BOA [35] is a nature based metaheuristic

approach which influences the behaviour Here xim and x tj are the m th and j th
of mating and foraging of butterfly. One of
butterflies selected randomly and when xim
the major properties of BOA varies from
t
other optimization approaches that are and x j is considered under same iteration,
every butterfly has its separate scent. The that means butterflies becomes a local
fragrance is expressed as: random walk. When xim and x tj is not

f r  sI b (12) considered under same iteration random

walk may diversify the solution.
Where f r represents the identified
Both local search and global search for
magnitude of fragrance, s is modality of mating and food partner via the butterflies
sensor and I b represents stimulus intensity in nature can happen. Hence a switch
with absorption of fragrance. probability is considered to transform the
intensive local search and normal global
The value of s ranges from [0, ] but the
search.
value is identified by a particularity of the
optimization issues in the BOA iterative It is seen from Equation (14) and (15) that
choosing randomly local and global search
will affect BOA is trapped by local optima. End for
Further the parameter ri capacity for
Find the best fragrance
adjusting local and global is limited.
Therefore some modification is needed. For every search agent
Hence the new optimal solution is
obtained by the following equations. Initialize the random number

Therefore the new equation for global If ri <p then

search is calculated as
Move towards best position using
Equation (16)
x tj1  w. x tj  (ri  g b  x tj ) f r (16)
2

Else
The new equation for local search space is
expressed as Move randomly by Equation (17)

x tj1  ( w  1) x tj  (ri  xim  x tj ) f r (17)

2 End if

End for
Where w is a weighting coefficient.
Comparing with Equation (14) and (15), Update the power exponent b
the updated equation (16) and (17) has the
features like the weighting coefficient is End while
able to adjust among local and global Return the best fitness solution
search when compared to the original
BOA. The best solution is updated either
by equation (16) or equation (17). These
6. Experimental results and discussion
two equation provides better convergence
speed because of the weighting coefficient. This section gives the performance
Further this model has better convergence analysis and discussion about the
speed and avoid local optima. Therefore developed scheme. The entire
this mBOA provides better results due to implementation has been processed on a
the optimal value. The Pseudo-code system with 8 GB RAM and Intel Core i5
modified BOA is algorithm 1. CPU with 3.0 GHz speed. To implement
the proposed scheme, PYTHON 3.8 is
Algorithm 1: Pseudo-code of mBOA
utilized. The dataset taken in this paper for
Initialize butterflies population the experimentation is soil dataset. The
developed approach performance is
Initialize stimulus intensity I b implemented with metrics like RMSE, R2,
MSE, MedAE, MAE, MSLE, MAPE and
Initialize switch probability and modality
EVS, error measures and accuracy are
of sensor s
utilized for the performance evaluation.
While termination criteria not met do
6.1 Evaluation metrics
Compute fragrance by Equation (12)
The error measures are estimated at every
iteration and the mean of each metric is
evaluated after the imputation of all the MedAE is a robust computation of the
values to obtain their overall performance.
variability of a univariate sample of
The mathematical expression for each
quantitative data and it is expressed as
metric is defined as:

MedAE  median p  pˆ ,...., p  pˆ 

MSE computes the mean of the squares of
the errors. That means mean squared  i i n n 
variation among actual and estimated (22)
values. MSE is represented as
MAPE also known as mean absolute
MSE 
1 n

 p  pˆ i
ni 1 i
2  percentage deviation (MAPD), it is a
calculation of prediction accuracy of a
(18) forecasting approach in statistics and it is
MSLE is a ratio among the actual and computed as

predicted values. Mean squared

1 n
MAPE   p
logarithmic error is as the name suggests, n i 1 i
it is a variation of the MSE. (23)
MAE is the absolute difference between an
observed value of a quantity and the true
2 value. That is the difference between true
MSLE 
1 n
  
 ln 1  pi  ln 1  pˆ i
ni 1
 and measured length and it is expressed as:

(19)
1 n
RMSE is the differences between values MAE   p  pˆ i
ni 1 i
predicted using an estimator and the (24)
2
observed values and it is represented as R is a statistical measure of fit that
represents how much difference of a
dependent variable is expressed by the
independent variable in a regression model
RMSE 
1 n

 pi  pˆ i
ni 1
2  and it is expressed as
1
  2
 


(20) 

  i
p  pˆ 2  
i
R 2  1  i 1 
 
EVS is the name suggested, it is metric  
 pi 
   
 i 1
   (25)
used to calculate the ratio among error
variance and true values variance and it is Accuracy is defined as the corrected

represented as prediction to the total number of prediction

and it is calculated by the below
1 n p i  pˆ i expression
EVS  
n i 1 pi

(21)
 Correct prediction the art techniques like ELM, artificial
Accuracy 
Total number of prediction neural network (ANN), support vector
machine (SVM), gradient boost (GB) and
(26)
random forest (RF). The simulation is
Where the expected outcome is denoted as
performed on aKNCN -ELM-mBOA with
pi , the predicted outcome for data is
these existing methods using the error
represented as p̂i , where i  1,2,....n . metrics to determine the prediction
6.2 Comparative analysis of aKNCN - efficiency of each method.
ELM-mBOA and Existing approaches
The performance of proposed aKNCN-
ELM-mBOA is compared with the state of
Table 1: Error measures of various approaches

Approaches Error measures

MAE MSE RMSE MSLE R2 EVS MedAE MAPE

aKNCN -ELM-mBOA 0.064 0.091 0.301 0.011 0.817 0.818 0.0504 3.932

aKNCN- ELM-BOA 0.067 0.095 0.309 0.011 0.806 0.808 0.0529 3.871

aKNCN -ELM 0.097 0.134 0.366 0.016 0.730 0.731 0.078 5.565

aKNC _ANN 0.130 0.181 0.426 0.022 0.636 0.636 0.105 7.809

aKNC-SVM 0.165 0.230 0.480 0.028 0.538 0.538 0.132 9.685

aKNC-GB 0.231 0.320 0.566 0.039 0.359 0.359 0.187 13.71

aKNC-RF 0.293 0.413 0.642 0.050 0.174 0.17 0.233 17.4

value. From the table, the MAE, MSE,

RMSE, MSLE, MedAE and MAPE value
Table 1 provides the analysis of error obtained by the proposed aKNCN -ELM-
values attained by various approaches in mBOA is lower than the existing methods.
terms of RMSE, MAE, MSLE, MAPE, While the R2 and EVS values attained by
MedAE, EVS and MSE. MSE is the mean the proposed aKNCN -ELM-mBOA are
of sum of the squared errors. MSLE is the
higher than other strategies.
squared logarithmic error’s predicted
 Figure 9: Comparative analysis on MAE, MSLE and MedAE

Figure 9 illustrates the resultant graph of and aKNC-RF are0.067, 0.097, 0.130,
MAE, MSLE and MedAE for various 0.165, 0.231 and 0.293 respectively.
approaches. From the graphical Similarly the proposed model achieves the
representation, the error values for MAE, better value of MSLE and MedAE. In
MSLE and MedAE obtained by the general, if the error occurred is less in the
proposed method is lower than the existing prediction system then it is considered as
methods. The MAE of the aKNCN-ELM- an effective model. Therefore, the
mBOA is 0.064 and the MAE of the proposed model is efficient for crop yield
aKNCN -ELM-BOA, aKNCN –ELM, prediction.
aKNC _ANN, aKNC-SVM, aKNC-GB
 Figure 10: Comparative analysis on RMSE, MSE and R2

aKNCN-ELM-BOA, aKNCN –ELM,

aKNC-ANN, aKNC-SVM, aKNC-GB and
Figure 10 shows the resultant graph of aKNC-RF are 0.806, 0.730, 0.636, 0.538,
RMSE, MSE and R2 for the proposed and 0.359 and 0.174. In addition the RMSE
the existing techniques. From the graphical value of the proposed model is 0.301
representation, the proposed model which is less error rate than the other
obtained less error values for MSE as well classification methods. Thus the proposed
as RMSE than other methods. However, model proved its performance in all the
R2 value attained by the proposed cases.
approach is higher than the existing
techniques. The R2 value of the proposed
model is 0.817, where the R2 value of
 Figure 11: Comparative analysis on EVS and MAPE

Figure 11 represents the resultant graph of ELM, aKNCN -ANN, aKNCN-SVM,

EVS and MAPE for the proposed and the aKNCN-GB and aKNCN-RF are 5.565,
existing techniques. From the graphical 7.809, 9.685, 13.71 and 17.1 respectively.
representation, the MAPE value obtained In all the metrics comparison, the proposed
by the proposed method is lesser than any model has generated better outcomes and
other methods. However, EVS value other methods achieve less accuracy due to
attained by the proposed approach is the computational complexity and
higher than the existing techniques. That overfitting problem. Our proposed model
is, MAPE value of aKNCN -ELM-mBOA attains higher outcomes due to the optimal
and aKNCN -ELM-mBOAare 3.932 and selection by mBOA.
3.871 where the MAPE value of aKNCN–
 Figure 12: Accuracy measure of proposed aKNCN -ELM-mBOA and existing methods

Figure 12 represents the accuracy measure This work focuses on predicting the yield
for the proposed aKNCN-ELM-mBOA of the crop based on two-tier ML approach
and the existing approaches. While named aKNCN and ELM-mBOA. In the
considering the accuracy measure, the first tier, the proposed aKNCN model is
actual and the predicted data are nearly used to estimate the soil quality based on
same then the system is said to be efficient IoT system collected soil nutrients. In the
for the crop yield prediction. It depicts the second tier, the soil quality score along
actual data and the predicted data of with other crop yield related parameters
different techniques. From the graph, the like temperature and rainfall are taken as
proposed aKNCN-ELM-mBOA predicted the input of ELM model to predict the crop
the result more accurately than the existing yield. The hyper parameter tuning of ELM
methods. The proposed aKNCN-ELM- prediction model is achieved by mBOA to
mBOA reached near to the actual data enhance the prediction performance of
whereas the other techniques did not attain ELM. PYTHON tool is used for the
a better accuracy. AKNCN-GB achieved implementation of proposed system. Soil
very low accuracy that other strategies. dataset is utilized for performance
Therefore, the proposed aKNCN-ELM- evaluation of the proposed prediction
mBOA is effective than the existing model. The proposed scheme attains better
techniques. results than the other classification models
on the basis of accuracy, RMSE, R2, MSE,
7. Conclusion MedAE, MAE, MSLE, MAPE and EVS.
The RMSE and MAE of the aKNCN-
ELM-mBOA is found to be 0.301 and [4] Savchenko, Olesya M., Maik Kecinski,
0.064 respectively. In future, analysis Tongzhe Li, and Kent D. Messer.
based on time-series will be done to
"Reclaimed water and food production:
predict the future values. The use of
different parameters like soil nutrients, soil Cautionary tales from consumer
quality, irrigated area and agricultural research." Environmental research 170
points can be used to extend the scope of
(2019): 320-331.
the research as well as improve the
accuracy of the system. In addition, deep
[5] Terence, Sebastian, and Geethanjali
learning based smart agriculture can be
used with the IoT system in order to Purushothaman. "Systematic review of
enhance the production quality. Internet of Things in smart

References farming." Transactions on Emerging

Telecommunications Technologies 31, no.
[1] Muangprathub, Jirapond, Nathaphon
6 (2020): e3958.
Boonnam, Siriwan Kajornkasirat,
Narongsak Lekbangpong, Apirat [6] Kalimuthu, M., P. Vaishnavi, and M.
Wanichsombat, and Pichetwut Nillaor. Kishore. "Crop Prediction using Machine
"IoT and agriculture data analysis for Learning." In 2020 Third International
smart farm." Computers and electronics in Conference on Smart Systems and
agriculture 156 (2019): 467-474. Inventive Technology (ICSSIT), pp. 926-
932. IEEE, 2020.
[2] Mekala, Mahammad Shareef, and P.
Viswanathan. "(t, n): Sensor Stipulation [7] Farooq, Muhammad Shoaib, Shamyla
with THAM index for smart agriculture Riaz, Adnan Abid, Kamran Abid, and
decision-making IoT system." Wireless Muhammad Azhar Naeem. "A Survey on
Personal Communications 111, no. 3 the Role of IoT in Agriculture for the
(2020): 1909-1940. Implementation of Smart Farming." IEEE
Access 7 (2019): 156237-156271.
[3] Ayaz, Muhammad, Mohammad
Ammad-Uddin, Zubair Sharif, Ali [8] Reddy, Kasara Sai Pratyush, Y.
Mansour, and El-Hadi M. Aggoune. Mohana Roopa, and Narra Sai Nandan.
"Internet-of-Things (IoT)-based smart "IoT based Smart Agriculture using
agriculture: Toward making the fields Machine Learning." In 2020 Second
talk." IEEE Access 7 (2019): 129551- International Conference on Inventive
129583. Research in Computing Applications
(ICIRCA), pp. 130-134. IEEE, 2020.
[9] Miranda, Jhonattan, Pedro Ponce, network." Neural Computing and
Arturo Molina, and Paul Wright. "Sensing, Applications (2020): 1-11.
smart and sustainable technologies for
[15] Gopal, PS Maya, and R. Bhargavi. "A
Agri-Food 4.0." Computers in
novel approach for efficient crop yield
Industry 108 (2019): 21-36.
prediction." Computers and Electronics in
[10] Shastry, K. Aditya, and H. A. Sanjay. Agriculture 165 (2019): 104968.
"Hybrid prediction strategy to predict
[16] PS, Maya Gopal. "Performance
agricultural information." Applied Soft
evaluation of best feature subsets for crop
Computing 98 (2021): 106811.
yield prediction using machine learning
[11] Elavarasan, Dhivya, and PM Durairaj algorithms." Applied Artificial
Vincent. "Crop yield prediction using deep Intelligence 33, no. 7 (2019): 621-642.
reinforcement learning model for
[17] Shook, Johnathon, Tryambak
sustainable agrarian applications." IEEE
Gangopadhyay, Linjiang Wu, Baskar
Access 8 (2020): 86886-86901.
Ganapathysubramanian, Soumik Sarkar,
[12] Dang, Chaoya, Ying Liu, Hui Yue, and Asheesh K. Singh. "Crop yield
JiaXin Qian, and Rong Zhu. "Autumn prediction integrating genotype and
Crop Yield Prediction using Data-Driven weather variables using deep
Approaches:-Support Vector Machines, learning." arXiv preprint
Random Forest, and Deep Neural Network arXiv:2006.13847 (2020).
Methods." Canadian Journal of Remote
[18] Nevavuori, Petteri, Nathaniel Narra,
Sensing (2020): 1-20.
Petri Linna, and Tarmo Lipping. "Crop
[13] van Klompenburg, Thomas, Ayalew Yield Prediction Using Multitemporal
Kassahun, and Cagatay Catal. "Crop yield UAV Data and Spatio-Temporal Deep
prediction using machine learning: A Learning Models." Remote Sensing 12, no.
systematic literature review." Computers 23 (2020): 4000.
and Electronics in Agriculture 177 (2020):
[19] Elavarasan, Dhivya, and PM Durai
105709.
Raj Vincent. "A reinforced random forest
[14] Bhojani, Shital H., and Nirav Bhatt. model for enhanced crop yield prediction
"Wheat crop yield prediction using new by integrating agrarian
activation functions in neural parameters." Journal of Ambient
Intelligence and Humanized [25] dos Santos, U.J.L., Pessin, G., da
Computing (2021): 1-14. Costa, C.A. and da Rosa Righi, R., 2019.
AgriPrediction: A proactive internet of
[20] Suchithra, M. S., and Maya L. Pai.
things model to anticipate problems and
"Improving the prediction accuracy of soil
improve production in agricultural
nutrient classification by optimizing
crops. Computers and electronics in
extreme learning machine
agriculture, 161, pp.202-213.
parameters." Information processing in
Agriculture 7, no. 1 (2020): 72-82. [26] Saranya, C. P., and N. Nagarajan.
"Efficient agricultural yield prediction
[21] Abbas, Farhat, Hassan Afzaal,
using metaheuristic optimized artificial
Aitazaz A. Farooque, and Skylar Tang.
neural network using Hadoop
"Crop yield prediction through proximal
framework." Soft Computing 24, no. 16
sensing and ML
(2020): 12659-12669.
algorithms." Agronomy 10, no. 7 (2020):
1046. [27] Filippi, Patrick, Edward J. Jones,
Niranjan S. Wimalathunge, Pallegedara
[22] Rezk, Nermeen Gamal, Ezz El-Din
DSN Somarathna, Liana E. Pozza,
Hemdan, Abdel-Fattah Attia, Ayman El-
Sabastine U. Ugbaje, Thomas G. Jephcott,
Sayed, and Mohamed A. El-Rashidy. "An
Stacey E. Paterson, Brett M. Whelan, and
efficient IoT based smart farming system
Thomas FA Bishop. "An approach to
using ML algorithms." Multimedia Tools
forecast grain crop yield using multi-
and Applications 80, no. 1 (2021): 773-
layered, multi-farm data sets and machine
797.
learning." Precision Agriculture 20, no. 5
[23] Bu, Fanyu, and Xin Wang. "A smart (2019): 1015-1029.
agriculture IoT system based on deep
[28] Sun, Jie, Liping Di, Ziheng Sun,
reinforcement learning." Future
Yonglin Shen, and Zulong Lai. "County-
Generation Computer Systems 99 (2019):
level soybean yield prediction using deep
500-507.
CNN-LSTM model." Sensors 19, no. 20
[24] Nevavuori, Petteri, Nathaniel Narra, (2019): 4363.
and Tarmo Lipping. "Crop yield prediction
[29] Sinwar, Deepak, Vijaypal Singh
with deep convolutional neural
Dhaka, Manoj Kumar Sharma, and Geeta
networks." Computers and electronics in
Rani. "AI-based yield prediction and smart
agriculture 163 (2019): 104859.
irrigation." In Internet of Things and
Analytics for Agriculture, Volume 2, pp. Using Principle Component Analysis and
155-180. Springer, Singapore, 2020. Adaptive k-Nearest Centroid Neighbor
Classifier. International Journal of
[30] Cao, Juan, Zhao Zhang, Fulu Tao,
Integrated Engineering, 13(1), pp.177-
Liangliang Zhang, Yuchuan Luo, Jing
187.."
Zhang, Jichong Han, and Jun Xie.
"Integrating Multi-Source Data for Rice [35] Arora, Sankalap, and Satvir Singh.
Yield Prediction across China using "Butterfly optimization algorithm: a novel
Machine Learning and Deep Learning approach for global optimization." Soft
Approaches." Agricultural and Forest Computing 23, no. 3 (2019): 715-734.
Meteorology 297 (2021): 108275.
[36] Li, B., Rong, X. and Li, Y., 2014. An
[31] Guo, Yahui, Yongshuo Fu, Fanghua improved kernel based extreme learning
machine for robot execution failures. The
Hao, Xuan Zhang, Wenxiang Wu,
Scientific World Journal, 2014.
Xiuliang Jin, Christopher Robin Bryant,
and J. Senthilnath. "Integrated phenology
AUTHORS PROFILE
and climate in rice yields prediction using
machine learning methods." Ecological
Indicators 120 (2021): 106935. Ms. Akanksha Gupta is
working as an assistant professor in the
[32] Wakhare, Prashant B., S. department of Computer Science,
Neduncheliyan, and Gaurav S. Sonawane. University of Delhi and currently pursuing
Ph.D in the area of Internet of Things, from
"Automatic Irrigation System Based on
Shri Venkateshwara University. She
Internet of Things for Crop Yield completed B.Sc honors in Computer
Prediction." In 2020 International science from University of Delhi in 2009
and MCA from Bharatiya Vidyapeeth
Conference on Emerging Smart
college, Indraprastha University, Delhi in
Computing and Informatics (ESCI), pp. 2012. She has also qualified UGC NET in
129-132. IEEE, 2020. computer science in 2013. She has a
teaching experience of over 7 years and
[33] Patil, Suhas M., and R. Sakkaravarthi. has earlier published 2 research papers in
"Internet of things based smart agriculture reputed international journals and 2
chapters in edited books and also
system using predictive analytics." Asian participated in many conferences.
J. Pharm. Clin. Res 10 (2017): 148-152.

[34] Rosdi, B.A., Mukahar, N. and Han,

Dr. Priyank Nahar pursued his
N.T., 2021. Finger Vein Recognition UG degree in Computer
Applications(BCA). He followed it with
MCA, M.Phil(CS), Ph.D(CS). He has over
14 years of teaching experience. He
worked as a corporate trainer in companies
like Hewlett Packard and Integer software,
Bangalore. He has attended more than 5
FDPs/seminars and conferences. He has
written 5 research papers in international
and national journals. He has also authored
13 books on paper solution series along
with a book on VB.Net for BCA students.
Currently, he is associated with MCA
department as an Associate Professor.