Content uploaded by Ankit Mahule
Author content
All content in this area was uploaded by Ankit Mahule on Mar 30, 2024
Content may be subject to copyright.
Ankit.Arun.Mahule et al., International Journal of Advanced Trends in Computer Science and Engineering, 9(2), March - April 2020, 2284 – 2288
2284
Hybrid Method for Improving Accuracy of Crop-Type Detection
using Machine Learning
Ankit.Arun.Mahule1, Dr.A.J.Agrawal2
1Department of Computer Science and Engineering
Shri Ramdeobaba Kamla Nehru Engineering College, Nagpur-440013, Maharashtra, India
ankitmahule2@gmail.com
2Department of Computer Science and Engineering
Shri Ramdeobaba Kamla Nehru Engineering College, Nagpur-440013, Maharashtra, India
agrawalaj@rknec.edu
ABSTRACT
Crop type prediction using sensors requires a large amount
of training data from different sensors. These sensors
include but are not limited to temperature, humidity, wind
direction, wind speed, gas sensors, etc. Gathering accurate
data from these sensors is a relatively easy task, but the
prediction of crop type from the gathered data requires the
knowledge and implementation of high-end classification
algorithms. In this work, we analyze different classification
algorithms and compare their performances to evaluate the
best algorithms suited for the task of crop classification.
Furthermore, we also analyze the effect of these algorithms
on different crop-prediction applications and recommend
which techniques are best suited for which kind of crop.
Finally, we suggest some recommendations to the existing
algorithms to make them more effective in terms of
prediction and response time. This work also suggests a
novel algorithm for crop-type detection, and fuses the
reviewed algorithms for obtaining a better accuracy in
crop-prediction systems.
Key words: Accuracy, Classification, Crop, Prediction,
Recommend
1. INTRODUCTION
Government and agricultural managers require statistics on
the spatial distribution and location of cultivated vegetation
for making plans functions. Groups can more accurately plan
the import and export of food products based on such
records. Although some ministries of agriculture and meals
protection yearly commission their workforce to map
exceptional crop types, these ground surveys are costly and
but cover only a pattern of farms. Crop identification and
classification is a multi-domain image processing problem, in
which image segmentation or division is one of the most
useful tasks while performing classification. The division
procedure permits us to isolate the picture into noteworthy
parts agreeing on a specific paradigm. Grouping calculations
like k-implies fluffy c-implies and in this way the Iterative
Self-Organizing Information Analysis Technique
(ISODATA) calculation are utilized effectively for division
issues, anyway, these strategies characteristically, don't think
about the logical data for a pixel, what is important to get a
genuine division. An extremely viable methodology for
including the highlights of the pixel neighbourhood is that
the Bayesian estimation nears the Markov Random Field
(MRF) [2-9]. With this methodology, a name field is
registered accepting that reliance exists between all
likelihood circulations of the pixels having a place with an
equal neighbourhood. This supposition is dictated by
considering Markov Chain prior dissemination. Gauss
Markov Measure Field (GMMF) [4] is one of the models that
mix Bayesian estimation in with Markov Random Field and
it's utilized in various game plan endeavours [5-8, 11-13].
One of the most difficulties for GMMF, concerning all
procedures reinforced the mix of Bayesian estimation and
MRF for picture division, is that the likelihood count. On
account of 1D or 3D include spaces the probability is
frequently registered upheld the relating standardized
histograms. Nonetheless, the calculation of the probability
turns into an extremely difficult issue at the point when the
measure of highlights increments. Inside the instance of
harvests grouping for satellite pictures the number of
highlights is high, which might impede the immediate
utilization of the GMMF [4]. On the contrary hand, crop
arrangement might be a perplexing assignment because of the
comparability of the unearthly marks among various crops.
Consequently, the decision of the element space, i.e., data
sources, might be a key advance during this exploration, with
the goal that we will utilize the GMMF as a classifier that
licenses us to incorporate logical data and to arrive at great
order results. In creators considered a pixel-based picture
approach to section 5 diverse land spread types in Russia.
The trial work incorporated the base Euclidean separation,
the case classifier, Mahalanobis separation, the most extreme
probability classifier, and grouping strategies. The
component space was made out of blue, green, and red
groups. The least difficult presentation was accomplished by
the most extreme probability. Creators in utilized three
diverse vegetation lists: the Normalized Difference
ISSN 2278-
3091
Volume 9 No.2, March - April 2020
International Journal of Advanced Trends in Computer Science and Engineering
Available Online at http://www.warse.org/IJATCSE/static/pdf/file/ijatcse209922020.pdf
https://doi.org/10.30534/ijatcse/2020/209922020
Ankit.Arun.Mahule et al., International Journal of Advanced Trends in Computer Science and Engineering, 9(2), March - April 2020, 2284 – 2288
2285
Vegetation Index (NDVI), the Green Normalized Difference
Vegetation Index (GNDVI), and the Normalized Difference
Red Edge Index (NDRE) for crop order inside the district
found in Turkey. All lists were registered mulling over the
ghostly groups acquired from the Rapid Eye satellite, which
is that the principal high-goals multispectral satellite
framework joining the red-edge band which is touchy to
vegetation chlorophyll. Four diverse element spaces were
concentrated inside the examination in: the essential ghostly
space contains NDVI, GNDVI, and NDRE lists, the other
three components ghostly spaces are made out of just two of
the three lists included inside the first space. For crop order,
the help vector machine technique was utilized. Thus, to
solve this issue, the next section describes some of the recent
algorithms used for this purpose, and statistically analyses
them to find the advantages and drawbacks of each of them.
The section also proposes some improvements in them and
suggests which algorithms must be used for what kind of
application.
2. LITERATURE REVIEW
Calculations that order a given arrangement of readings into
a gathering of yield types are called crop characterization
calculations. These calculations take a shot at a harvest based
preparing set to return up with a model or a gathering of
decides that arrange a given information sensor information
into one among a gathering of discrete yield types. The k-
implies [5] calculation has been utilized for grouping
purposes. There are many grouping calculations, seemingly
superior to this, and consequently the method of reasoning
behind picking this is because the code overall has been
composed during a multi-strung, bunch prepared way, and k-
implies is that the main calculation in Microsoft R supporting
such a component.
Conventional enhancement calculations, as stochastic slope
plunge (SGD), upgrade the experimental misfortune work
straightforwardly. The SDCA [6] picks an exceptional
methodology that upgrades the twin issue. The twin
misfortune work is parameterized by per-model loads. In
each cycle, when a preparation model from the preparation
informational index is examined, the relating model weight is
balanced all together that the twin misfortune work is
advanced concerning this model. No learning rate is required
by SDCA to settle on a choice advance size as is required by
different inclination drop techniques.
Meanwhile neural systems are commonly known to be
utilized in profound learning and demonstrating complex
issues like picture acknowledgment, they're additionally
handily adjusted to re lapse issues. Any class of measurable
models is regularly seen as a neural system if they utilize
versatile loads and may inexact non-direct elements of their
data sources. Neural system relapse is especially fit for issues
where a progressively customary relapse model can't fit an
answer.
Quick Logistic Regression [7] is taken care of by the
rxLogisticRegression () calculation which is utilized to
anticipate the value of an all-out factor from its relationship
to at any rate at least one free factors accepted to have a
strategic circulation. The contrast between this model and
subsequently the first (Logistic Regression) is this is
frequently commonly quicker and takes into account a more
extensive customization.
The advancement system used for rxLogisticRegression is
that the constrained memory Broyden-Fletcher-Goldfarb-
Shanno (L-BFGS). Both the L-BFGS and standard BFGS
calculations utilize semi Newtonian techniques to assess the
computationally concentrated Hessian framework inside the
condition utilized by Newton's strategy to ascertain steps. In
any case, the L-BFGS [8] guess utilizes just a constrained
measure of memory to process ensuing advance heading, so
as that it's particularly fitted to issues with an outsized
number of factors. The memory Size parameter determines
the measure of past positions and slopes to store to be
utilized inside the calculation of the resulting step.
This student can utilize flexible net regularization: a straight
blend of L1 (tether) and L2 (edge) regularizations.
Regularization might be a strategy that will render a not well-
presented issue progressively tractable by forcing imperatives
that give data to enhance the information which forestalls
over fitting by punishing models with extraordinary
coefficient esteems [9]. This will improve the speculation of
the model learned by choosing the ideal unpredictability
inside the predisposition fluctuation exchange off.
Regularization works by adding the punishment that is
identified with coefficient esteems to the mistake of the
speculation. An exact model with outrageous coefficient
esteems would be punished more, yet a less precise model
with increasingly traditionalist qualities would be punished
less. L1 and L2 regularization have various impacts and uses
that are corresponding in specific regards. l1Weight: are
regularly applied to inadequate models, when working with
high-dimensional information. It pulls little loads related
highlights that are generally irrelevant towards 0. l2Weight:
is ideal for information that is not inadequate. It pulls huge
loads towards zero. The resulting segment portrays the
proposed work.
3. PROPOSED WORK
The block diagram for the proposed system is shown in
Figure 1. From the diagram we can observe that the input
goes into different classification layers, and an output is
obtained from each of the layers. A simple machine learning-
based selection system is then used to find out the most
matching class. This class is shown at the output as the
predicted crop class.
Ankit.Arun.Mahule et al., International Journal of Advanced Trends in Computer Science and Engineering, 9(2), March - April 2020, 2284 – 2288
2286
Figure 1: Block diagram of the proposed hybrid classifier
Crop type classification is a very complicated process, due to
the number of interdependent steps which are needed to be
followed in a proper order. Due to these complexities, the
selection of standard algorithms in each stage is very crucial
so that the final outputs can be accurate. The following
general steps are followed for any hyper spectral image
classification system; these steps can be modified as per the
application, in which the classifier is applied,
Placement of the image acquisition device
Acquisition of the image
Image fusion
Segmentation or clustering of the fused image
Feature extraction and selection
Classification and post-processing
In the proposed crop type classification application requires
these steps to be followed in a proper sequence. In our work
we will be utilizing the concept of classification to evaluate
the Type of crop present in a particular region, which will
help industries to use automatic drone imagery data for
analysis of different areas. The layers which are required for
processing & classifying the leaf diseases are,
• Image Capture Layer
This is the primary layer within the processing set. This layer
is liable for capturing leaf images and arranging them for
correct processing. This layer must confirm that each one the
captured leaf images are taken at a correct angle, at a correct
distance, and at a correct lightning condition to urge high
accuracy during processing.
• Pre-processing Layer
Images captured by the primary layer are given for pre-
processing. The pre-processing layer will remove any noise
(if present) within the image, smoothen the image, and make
it ready for processing. The only responsibility of this layer is
to convert the image into a form that's consistent in terms of
size, brightness, amplitude, etc. This layer is typically not
concentrated upon during the design of the classification
network, but if properly designed it can improve the system
performance by leaps and bounds.
• Segmentation Layer
The segmentation layer is that the first point of processing for
the image. The pre-processed image is given to the
presentation layer to extract the foremost important
components from the image. The components include the
regions of interest, just like the leaf regions, the infected
regions, and therefore the background regions. Usually the
background regions are detected with utmost accuracy by
algorithms like Otsu, Saliency maps, etc. then subtracted
from the input image to urge the ultimate segmented image.
• Feature Extraction Layer
The segmented regions are given for feature extraction. This
step defines the extent of accuracy of the system, and must be
most carefully selected. Input image pixels can't be given
directly for classification because they will be very high in
number, and may need inconsistent leaf positioning. Thereby
this layer converts the input image into a numerical array of
features that will be used for representing the image
uniquely. An accurate feature extraction layer will produce
similar features for images in the same class, while it'll
produce dissimilar feature sets for images of various classes.
Methods like grey level co-occurrence matrix, colour maps,
edge maps, etc. are used for this purpose.
• Feature Selection Layer
While the feature extraction layer will try its best to supply
different feature sets for various class of images, but
redundancy will always persist. This redundancy affects the
classification performance just in case of huge datasets.
Thanks to this, there's a requirement to get rid of the
redundancy from the feature sets. Algorithms like feature set
variation, differential evolution, etc. are used for this
purpose.
• Classification Layer
A major part of research in plant disease detection is directed
towards this layer. This layer is liable for comparing the
features with one another, and evaluating the simplest
possible category for the query image. Algorithms like
convolution neural networks, deep nets, support vector
machines, etc. are used for this purpose. The proposed ML
selector works by finding out the most recurring class which
is identified by ANN, Random Forest, Naïve Bayes, SVM,
and CNN, and provides that particular class at the output.
4. RESULTS AND ANALYSIS
In this section, we define a set of criteria for analyzing the
algorithms. The set of algorithms are categorized as per the
length of the input, the type of features extracted, the
accuracy, and the probable application of the system under
various real-time scenarios. Table 1 showcases this
comparison in detail.
Ankit.Arun.Mahule et al., International Journal of Advanced Trends in Computer Science and Engineering, 9(2), March - April 2020, 2284 – 2288
2287
Table 1: Performance and analysis of different classifiers
Algorithm Crop Types Accuracy (%)
ID3[1] Wheat, Bajra 85.00%
Naïve Bayes [2] Wheat, C
otton,
Bajra 86.00%
Bayesian Belief [3]
Various 79.50%
K2 [4]
Corn, Cotton,
Orange 85.00%
Gradient Ascent
Training [5] Various 89.00%
Linear Regression
[6] Various 86.00%
Multi-Layer
Perceptron [7,8,9]
Various 91.00%
RBF [10,11] Various 93.00%
Conjunctive Rule
Algorithm[12] Various 30.00%
Proposed Various 95.00%
From the above table we can statistically summarize that the
performance of the proposed classifier which is better than
MLP and RBF systems will produce better accuracy than any
other classifier combination.
5. CONCLUSION AND FUTURE WORK
From the outcome we can see that the accuracy of the
proposed classifier which is better than MLP and RBF is
better than the usual classification systems. Moreover, the
proposed classifier, MLP, and RBF have been proven to be
the most effective in terms of classification, but their
effectiveness can be further improved with the help of a bag
of features like a bag of sensors and a bag of visual features.
These features when combined with a deep learning-based
classification system can produce very high accuracy.
As an extension of this research, we propose the use of
artificial intelligence techniques like q-learning and
reinforcement learning to further improve the performance of
the system. Moreover, the usage of feature selection
techniques will further improve the system accuracy.
REFERENCES
[1] M.V.R. Vivek, D.V.V.S.S. Sri Harsha, P. Sardar Maran,
“A Survey on Crop Recommendation Using Machine
Learning”, International Journal of Recent Technology
and Engineering (IJRTE) ISSN: 2277-3878, Volume-7,
Issue-5C, February 2019.
[2] S.Pudumalar, E.Ramanujam, "Crop Recommendation
System for Precision Agriculture", 2016 IEEE Eighth
International Conference on Advanced Computing
(ICoAC).
[3] Melchizedek I. Alipio, Allen Earl M. Dela Cruz, Jess
David A. Doria and Rowena Maria S. Fruto, "A Smart
Hydroponics Farming System Using Exact Inference
in Bayesian Network", 2017 IEEE 6th Global
Conference on Consumer Electronics (GCCE 2017).
[4] Ranjith Bose, Ranjith, Suraj Prakash, Subham Kumar
Singh, Dr Vishwanath, "Intelligent Approach for
Classification of Grain Crop Seeds Using Machine
Learning”, International Research Journal of
Engineering and Technology (IRJET) Volume: 05 Issue:
05 May-2018.
[5] P. Kanaga Priya, Dr. N. Yuvaraj, "An IoT Based
Gradient Descent Approach for Precision Crop
Suggestion using MLP", International Conference on
Physics and Photonics Processes in Nano Sciences
[6] Iwan Syarif, Dito Hafizh Indiarto, Ira Prasetyaningrum,
Tessy Badriyah, Edi Satriyanto, "Corn Pests and
Diseases Prediction using linear Regression and
Natural Spline Methods", 2018 International
Conference on Applied Science and Technology
(ICAST).
https://doi.org/10.1109/iCAST1.2018.8751583
[7] Nataliia Kussul, Mykola Lavreniuk, Sergii Skakun, and
Andrii Shelestov, "Deep Learning Classification of
Land Cover and Crop Types Using Remote Sensing
Data", IEEE Geosciences and Remote Sensing Letters.
[8] Muhd Khairulzaman Abdul Kadir, Mohd Zaki Ayob,
Nadaraj Miniappan, "Wheat Yield Prediction:
Artificial Neural Network based Approach", 2014 4th
International Conference on Engineering Technology
and Technopreneuship (ICE2T).
https://doi.org/10.1109/ICE2T.2014.7006239
[9] Andreas Kamilaris, Francesc X. Prenafeta-Boldu, "Deep
learning in agriculture: A survey", Computers and
Electronics in Agriculture 147 (2018) 70–90.
https://doi.org/10.1016/j.compag.2018.02.016
[10] Abhishek Pandeya, Sunil Kr. Jhab, and R.Prasada,
"Retrieval of Crop Parameters of Spinach by Radial
Basis Neural Network Approach Using Xband
Scatterometer Data”, ISSN 10683674, Russian
Agricultural Sciences, 2010, Vol. 36, No. 4, pp. 312–
315. Allerton Press, Inc., 2010.
https://doi.org/10.3103/S1068367410040245
[11] G. Camps-Valls, A. J. Serrano-Lopez, L. Gomez-Chova,
J. D.Martın-Guerrero, J. Calpe-Maravilla, and J.
Moreno, "Regularized RBF Networks for Hyper
Spectral Data Classification", Conference Paper in
Lecture Notes in Computer Science September 2004.
https://doi.org/10.1007/978-3-540-30126-4_53
[12] Jose M. Pena, Pedro A. Gutierrez, Cesar Hervas-
Martinez, Johan Six, Richard E. Plant and Francisca
Lopez-Granados, "Object-Based Image Classification
Ankit.Arun.Mahule et al., International Journal of Advanced Trends in Computer Science and Engineering, 9(2), March - April 2020, 2284 – 2288
2288
of Summer Crops with Machine Learning Methods",
Remote Sens. 2014, 6, 5019-5041.
https://doi.org/10.3390/rs6065019
[13] Shubhangi Neware, ”Fruit Grading System using k
means clustering and Artificial Neural Network”,
International Journal of Advanced Trends in Computer
Science and Engineering, Volume 9, No.1, January –
February 2020
https://doi.org/10.30534/ijatcse/2020/95912020
[14] H. D. Gadade, Dr. D.K.Kirange, ”Machine Learning
Approach towards Tomato Leaf Disease
Classification”, International Journal of Advanced
Trends in Computer Science and Engineering, Volume
9, No.1, January – February 2020
https://doi.org/10.30534/ijatcse/2020/67912020
