Chapter 10

Crop Classification and Mapping for

Agricultural Land from Satellite Images

A. Kalaivani and Rashmita Khilar

Abstract Agriculture is the backbone of Indian production which is the vital sector
for food production. It is very important for national government to know what type
of crops are being grown in which region for budget planning to import and export
food products. Traditional ground survey method is laborious, time-consuming, and
expensive. Along with this, continuous monitoring of crops is highly difficult. Crop
area estimation is a key element in crop production forecasting and estimation. Crop
classification and mapping are the most challenging tasks among the land use/land
cover classification problems.
In agriculture domain, the common approach used by the government (farmers)
for crop monitoring is to go to the field and acquire the images using cameras for
estimation of the crop yield. So in this context, a fast, reliable, and automated system
is required which provides the exact crop mapping using satellite images. In recent
years, crop identification and area monitoring from satellite images are given more
and more attention.
The stages are image acquisition, image preprocessing, feature extraction, and
image classification. Satellite images are preprocessed and features are extracted
from input images. Based on the features extracted, images are classified based
on the extracted features. The proposed automated system should provide better
accuracy than the existing in the literature.

Keywords Crop mapping · Crop classification · Satellite images · Feature

subset selection · Supervised classifier

A. Kalaivani
Department of CSE, Saveetha School of Engineering, Saveetha Institute of Medical and Technical
Sciences, Kuthambakkam, Tamil Nadu, India
e-mail: [email protected]
R. Khilar ()
Department of IT, Panimalar Engineering College, Chennai, Tamil Nadu, India

© Springer Nature Switzerland AG 2020 213

D. J. Hemanth (ed.), Artificial Intelligence Techniques for Satellite Image Analysis,
Remote Sensing and Digital Image Processing 24,
https://doi.org/10.1007/978-3-030-24178-0_10
214 A. Kalaivani and R. Khilar

10.1 Introduction

Agriculture is a boom of Indian economy, which accounted for 10% of the total
export earnings and serves as an important raw material for many food production
industries. Agriculture is the major source of food production and serves as a
backbone for reduction of poverty. National growth across global market depends on
the agricultural growth which depends on agricultural production. The agriculture
production can be improved by introduction of new varieties of technology in
cultivation and production of crops. Changing of crop pattern can better improve
agricultural production and makes a position rise in global market.
Remote sensing refers to the detection, identification, and classification of the
objects by observing and recording their signals which are collected through aerial-
or satellite-based sensors. The RS technology started with an oblique photograph of
a village near Paris from a balloon. It expanded from black and white photographs
to color photographs based on sensors. The sensors used to record the signal are
divided into two classes of active and passive remote sensing. Active remote sensors
use a transmitter and receiver to record the reflected or scattered electromagnetic
radiation. The remote sensor systems which measure energy emitted or scattered
from the objects are called as passive sensors. A schematic diagram showing a solar
radiation-based passive remote sensors system is shown in Fig. 10.1.
In India, remote sensors are used in crop area estimation, and a number of studies
are also explored for crop inventory, acreage estimation, and yield prediction. Later,
the Ministry of Agriculture and Department of Space collaborated for large-scale
projects for demonstrating the applications of optical and microwave RS data for
crop studies. The operational project FASAL supports India in preharvest estimation
of crops and predicts multiple crop production on different scales. CAPE was the
first large-scale project in India that used remotely sensed data for crop studies to
identify crops and estimate production. The satellites launched by India from the
source ISRO 2016 are listed in Table 10.1.
India is one of the major food grain-producing countries in the world similar to
China, USA, Russia, Canada, etc. To withstand in global food grain market, India
should be aware of the demand and production of other countries. The agricultural
statistics at the state level are done by State Agricultural Statistics Authorities
(SASAs). At the all-India level, the agricultural statistics are done by the directorate

Fig. 10.1 Passive remote sensor system

10 Crop Classification and Mapping for Agricultural Land from Satellite Images 215

Table 10.1 Satellites launched by India

Satellite Launch date Launch vehicle Major payloads
CARTOSAT – 2C Jun 22, 2016 PSLV-C34 PAN
SARAL Feb 25, 2013 PSLV-C20 AltiKa, ARGOS, SCBT
RISAT-1 Apr 26, 2012 PSLV-C19 SAR (C-band)
Megha-Tropiques Oct 12, 2011 PSLV-C18 MADRAS, SAPHIR, ScaRaB, ROSA
RESOURCESAT-2 Apr 20, 2011 PSLV-C16 LISS-3, LISS-4, AWiFS, AIS
CARTOSAT-2B Jul 12, 2010 PSLV-C15 PAN
Oceansat-2 Sep 23, 2009 PSLV-C14 OCM, SCAT, ROSA
RISAT-2 Apr 20, 2009 PSLV-C12 SAR (X-band)
IMS-1 Apr 28, 2008 PSLV-C9 Mx, HySI
CARTOSAT-2A Apr 28, 2008 PSLV-C9 PAN
CARTOSAT-2 Jan 10, 2007 PSLV-C7 PAN
CARTOSAT-1 May 05, 2005 PSLV-C6 PAN (Fore), PAN (Aft)
RESOURCESAT-1 Oct 17, 2003 PSLV-C5 LISS-3, LISS-4, AWiFS
TES Oct 22, 2001 PSLV-C3 PAN
Oceansat (IRS-P4) May 26,1999 PSLV-C2 OCM, MSMR
IRS-1D Sep 29,1997 PSLV-C1 PAN, LISS-3, WiFS
IRS-P3 Mar 21,1996 PSLV-D3 WiFS, MOS, IXAE, CBT
1RS-1C Dec 28,1995 Molniya PAN, LISS-3, WiFS
IRS-P2 Oct 15,1994 PSLV-D2 LISS-2A, LISS-2B
IRS-1E Sep 20,1993 PSLV-D1 LISS-1, MEOSS
IRS-1B Aug 29,1991 Vostok LISS-1, LISS-2A, LISS-2B
SROSS-2 Jul 13,1988 ASLV GRB, MEOSS
IRS-1A Mar 17,1988 Vostok LISS-1, LISS-2A, LISS-2B
RS-D2 Apr 17,1983 SLV-3 Smart sensor, L-band beacon
Bhaskara-II Nov 20,1981 C-l Intercosmos TV cameras, SAMIR
Source: ISRO, 2016

of economics and statistics. Although this system is well recognized, the major
shortcomings are subjectivity in the crop acreage estimation and delays in crop
forecasts.
Moorthi et al. [15] developed FASALSoft, an ISRO software framework for
crop production forecast using primarily remote sensor data analysis. In India,
the FASALSoft is accepted as operational for making multiple in-season crop
production forecasts and implemented by MNCFC, New Delhi. During the year
1880–2010, economic growth increased due to the significant land use and land
cover (LULC). The interactions of human activities, climate systems, and ecosystem
in the country are not specified by accurate database of LULC. Tiana et al. [12]
used high-resolution RS datasets from Resourcesat-1 and historical archives at
district and state levels to generate LULC datasets during 1880–2010 in India. The
experimental results identified that there is a significant loss of forests, and cropland
area increased significantly during 1880–2010. Due to this farm mechanization,
216 A. Kalaivani and R. Khilar

electrification and introduction of high-yielding crop varieties are introduced in

government policies to achieve self-sufficiency in food production.
A methodology for monitoring progress of Rabi crop area at country scale was
developed by Nigam et al. [16]. He used temporal vegetation index derived or
captured from Indian geostationary satellite (INSAT 3A). The 10-day maximum
NDVI composite products were generated and used over six crop dominant states
of India. The estimates showed a deviation of 18.1% to 14.6% deviations, and the
inter-seasonal variability in the estimate was consistent with the reported statistics
for Rabi crop. The authors recommend NDVI product with finer spatial resolution
satellite data for crop area monitoring for their country. Singla et al. [21] have
justified the role of geo informatics to be used for discriminating different crops
at various levels of classification, monitoring crop growth, and prediction of the
crop yield. The author justified that in addition to RS technology usage of ground
observations, reviews, GIS, and soil analysis.

10.2 Land Utilization and Crop Pattern in Tamil Nadu

In Tamil Nadu, agricultural crops are broadly classified into food crops and non-
food crops. Human-consumed crops are food crops which are further classified
as cereals and pulses. The cereals crops produced all over our state includes rice,
wheat, jowar, bajra, maize, ragi, and other millets, and pluses include gram, arhar,
moong, peas, and masoor. The non-food crops are divided into commercial crops
which are cultivated for sale. The other category of non-food crops is a plantation
crop which covers a large estate. The final category of horticultural crops are crops
with rich source of carbohydrates, proteins, organic acids, vitamins, and minerals
for human nutrition.
Land utilization depends on environmental factors such as soil characteristics,
climate, topography, and vegetation. Land can be properly utilized by humans
for agriculture, industry, forestry, energy production, settlement, recreation, water
catchments, and storage. For better utilization of the land, identify the natural
characteristics, extent and location, and its quality, productivity, suitability, and
limitations. The high utilization of land can be achieved through application of bio
fertilizers, hybrid seeds, double cropping, and modern methods (Fig. 10.2).
In Tamil Nadu, the population is very high and land resources are not utilized
due to the mal-utilization of land. So, if proper planning of land utilization yields
to higher agricultural production, our states can concentrate on the major crops
such as rice, paddy, cholam, cumbu, ragi, sugarcane, cotton, groundnut, and million
hectares.
In Tamil Nadu, the districts are classified into 13 districts which are (1)
Chengalpattu (Chennai, Kancheepuram, Thiruvallur), (2) South Arcot (Cuddalore,
Villupuram), (3) North Arcot (Thiruvannamallai, Vellore), (4) Salem (Namakkal),
(5) Dharmapuri (Krishnagiri), (6) Coimbatore, (7) Erode, (8) Tiruchirapalli (Karur,
Perambalur), (9) Pudukkottai, (10) Thanjavur (Thiruvarur, Nagapattinam), (11)
10 Crop Classification and Mapping for Agricultural Land from Satellite Images 217

LAND UTILISATION IN TAMIL NADU 2014-15

(Area in Hectares)
Barren and
Forests & Land under
Unculturable land,
misc, tree crops and
Other fallow lands 488460, (4%)
groves not included
1717831, (13%) in the net area sown,
2370153, (18%)
Current fallow lands Land put to Non-
1115041, (9%) agricultural uses,
2189876, (17%)
Culturable Waste
328326, (3%)

Permanent Pastures
and other grazing
lands, 109567, (1%)
Net area sown,
4713862, (35%)

Fig. 10.2 Land utilization in Tamil Nadu

Madurai (Dindigul, Theni), (12) Ramanathapuram (Kamarajar, Pasumpon Muthu-

ramalingam, Sivagangai), and (13) Tirunelveli (Thoothukudi).
The data for cropping pattern of each district are extracted from Season and Crop
Reports of Tamil Nadu during the period from 1981–1982 to 2005–2006 specified
crops average yield of 25 years of data for thirteen districts and state as a whole
which are shown in Tables 10.2 and 10.3.

10.3 Literature Survey

Smith et al. [22] and Rachel Barrett et al. [1] reported how the current technology
used to recognize and differentiate between horticultural crops in Tasmania,
Australia. A methodology for the systematic recognition of individual crops was
developed with a better utilization of the technology in a commercial environment.
The focus is tailored individually to five commercially significant horticultural
crops such as poppies, pyrethrum, potatoes, peas, and onions. These crops are
selected due to the predictive accuracy higher in comparison to 17 different crops
investigated. The logical conclusion from both the perspective of agronomic
and computational level is that the probability of achieving a correct outcome
is enhanced by incorporating horticultural and environmental conditions used in
both the training and testing data. The predictive accuracy can be improved by
specifically tailoring interpretative images to cropping sequences.
Computer-assisted photo interpretation (CAPI) tool funded by National Control
and Paying Agencies (NCPA) checks the proper spending of funds by identifying
crops via remotely sensed imagery. The major issue of the proposed method is
time-consuming and requires skilled interpreters. The objective of May et al. [14]
 218

Table 10.2 District-wise crop data in Tamil Nadu – sugarcane, cotton, groundnut, tapioca
Sugarcane Cotton Groundnut Tapioca
Extent of Average Extent of Average Extent of Average Extent of Average
cultivation yield (kg cultivation yield (kg cultivation yield (kg cultivation yield (kg
Districts (1) (percent) (10) ha) (11) (percent) (12) ha) (13) (percent) (14) ha) (15) (percent) (16) ha) (17)
Chengalpattu 2.93 108 0.04 351 16.58 2190 0.05 32156
South Arcot 0.96 113 1.37 362 17.66 1601 2.03 32608
North Arcot 7.09 82 1.01 361 33.10 1333 0.08 32222
Salem 3.43 125 3.37 321 21.27 1417 6.01 37199
Dharmapuri 3.23 86 2.59 336 12.59 1392 2.24 32726
Coimbatore 3.66 105 4.41 383 10.87 1327 0.25 40915
Erode 6.50 119 2.29 385 17.09 1552 0.76 41670
Trichy 3.70 109 3.11 253 11.40 1398 1.01 38985
Pudukkottai 1.70 108 0.31 339 21.60 1232 0.01 31012
Thanjavur∗ 2.29 101 0.63 397 3.23 1639 0.05 34124
Madurai 3.18 111 5.70 322 9.25 1524 0.05 35721
Ramanathapuram 1.07 102 7.71 209 5.41 1029 0.01 27575
Tirunelveli 0.77 104 10.71 216 2.61 1546 0.04 31286
Tamil Nadu 3.77 105 3.22 269 13.38 1455 1.21 33627
A. Kalaivani and R. Khilar
Table 10.3 District-wise crop data in Tamil Nadu – paddy, cholam, cumbu, ragi
Crops in Tamil Nadu
Paddy Cholam Cumbu Ragi
Extent of Average Extent of Average Extent of Average Extent of Average
cultivation yield (kg cultivation yield (kg cultivation yield (kg cultivation yield (kg
Districts (1) (per cent) (2) ha) (3) (percent) (4) ha) (5) (percent) (6) ha) (7) (percent) (8) ha) (9)
Chengalpattu 66.18 2896 0.09 1259 0.50 1882 1.37 1799
South Arcot 35.61 3052 1.75 985 10.00 1323 0.89 1961
North Arcot 27.09 2865 4.54 1143 2.51 1079 2.43 1976
Salem 10.01 2900 10.38 1025 3.43 1375 3.49 2131
Dharmapuri 9.83 2935 7.91 1221 1.18 1510 15.5 1610
Coimbatore 5.96 3382 27.35 508 0.53 1565 0.28 1998
Erode 16.93 3859 10.03 500 2.38 1022 4.03 1195
Trichy 22.62 3068 20.33 777 7.54 635 0.21 1827
Pudukkottai 51.92 2515 0.99 1073 0.39 1423 1.10 1473
Thanjavur∗ 67.36 2559 - - 0.04 1597 0.03 2370
Madurai 21.36 3549 16.17 1254 2.69 1341 0.34 1970
Ramanathapuram 51.04 1940 2.21 945 2.71 1130 1.79 1728
Tirunelveli 26.96 3671 3.80 1510 7.26 1340 0.64 2416
Tamil Nadu 31.94 2855 7.72 910 3.31 1113 2.27 1750
10 Crop Classification and Mapping for Agricultural Land from Satellite Images
219
220 A. Kalaivani and R. Khilar

developed a reliable control system to replace CAPI for crop identification. The
proposed method reduced control costs and completion time. The authors proposed
an automatic control system (ACS) which classifies crops based on a reliability
requirement. The main concern of this method require proper decision-maker to set
a high reliability level which restricts automatic crop identification which classify
crops at high certainty.
The authors Jonas Schmedtmann and Manuel L. Campagnolo [20] quantified
the accuracy of the proposed approach and analyze the trade-off between the
reliability level and the proportion of parcels that can be automatically controlled.
The automatic classification of agricultural land parcels is reliable, and it provides
a good performance even when crops are difficult to discriminate.
Foster and Kesselman [9] used spectral mixture modeling to produce the land
cover maps for Tumkur District, Karnataka. Spectral Mixture Analysis (SMA)
was performed and evaluated on Landsat-8 ETM (Enhanced Thematic Mapper
Plus) data, and the results are compared with the ground truth data. The specified
techniques give more accurate results by taking the absolute difference between
actual and modeled estimations with homogenous coconut land cover. SMA is easy
and low computational method used successfully to classify different vegetation
covers in intensive agricultural areas. The methodology is used in Czajkowski
et al. [4] satellite data selection and suitable method for classification and checking
the accuracy. The most important challenges by Beeresh et al. [2] are multiple
crops identification and differentiation of crops of same. Good soft computing and
analysis skills are required to classify and identify the class from multispectral
and hyper-spectral images. Researchers focused on supervised and unsupervised
classification methods along with hard classifiers and also with soft computing
techniques like fuzzy C mean and support vector machine.
Remote sensing images are good source for decision-making related to crops
monitoring and mapping in optical region, Rajesh K Dhumal et al. [7] uses a
multispectral images which gives much detail for overall vegetation mapping in
large area. Whereas it is having limitation due to broad wavelength and spatial
resolution this paper lacks in differentiating crops of similar type, this problem
overcomes by hyper-spectral images. Selection of spectral bands in hyper-spectral
images is also a quite challenging task. Some limitation of optical remote sensing
can be overcome by fusing optical remote sensing images with microwave remote
sensing images.
Foster et al. [11] carried out a study on pixel-based cropland classification by
fusion of data from satellite images with different resolutions. Methodology is
based on various multispectral images acquired at different resolutions by different
imaging instruments, Landsat-8 and Rapid Eye. The proposed method overcomes
the shortcomings of different instruments in the particular cropland classification
scenario situated on the plains of Vojvodina in northern Serbia. Researchers
proposed a data fusion method with a robust random forest classifier. It improves
the overall classification performance with a coarser spatial resolution in the given
specific cropland classification task. The method developed by Predrag Lugonja
10 Crop Classification and Mapping for Agricultural Land from Satellite Images 221

et al. [10] provides an improvement over the existing pixel-based classification

approaches through the combination of different data sources. Another contribution
of this paper is the employment of crowdsourcing by Bahram Salehi et al. [19]
in the process of reference data collection via dedicated smartphone application.
NIR band contributed a significant influence in classifiers that use the proposed data
fusion method.
Nataliia Kussul et al. [13] used unsupervised neural network (NN) for optical
imagery segmentation which helps in restoring missing data due to clouds and shad-
ows and an ensemble of supervised NNs. A fully connected multilayer perceptron
(MLP) and RS community random forest approaches are used and compared with
convolution NNs (CNNs). The proposed method 2-D CNN provides better perfor-
mance than 1-D CNNs, and small objects in the final classification were smoothed
and misclassified which leads to further modifications. The methodology proposed
by Daniel M. Howard et al. [5] is to classify major crops in the Greater Platte River
Basin (GPRB) in addition to existing crop classification products. The input for
the system includes Moderate Resolution Imaging Spectroradiometer-normalized
differential vegetation index, average long-term growing season temperature, aver-
age long-term growing season precipitation, and yearly start of growing season.
The overall accuracy of 78% is achieved for a test sample of roughly 215,000
independent points that were withheld from model training. Ten 250 m resolution
annual crop classification maps were produced and evaluated for the GPRB region,
one for each year from 2000 to 2009. The proposed method validates and focuses
on spatial distribution and county-level crop area and later compared with NASS
CDL and county statistics. Crops are classified based on the spatial distribution and
exhibited a close linear agreement from USDA datasets.
The methodology developed by Yetkin Özüm Durgun et al. [8] is based on
phenological characteristics of different crop types applied using 100-m Proba-V
NDVI data for the season 2014–2015. The postclassification rules were applied to
aggregate the crop type at the plot level. The methodology is applicable to 100-m
Proba-V and used in crop area mapping across the world. The accuracy ranges from
65% to 86%, and the kappa coefficient varied between 0.43 and 0.84 depending on
the site and the temporal window used. The main objectives of researchers Bahram
Salehi et al. [19] is to evaluate how the number of multi-temporal images used in an
OBIA framework affects the classification results for both RE and RS-2 data. The
above technique combines the object- and pixel-based image classification method
for classification of accurate crop type. The above methodology can be improved,
and accurate classification at discrete locations (approximately 25 × 25 km frames)
can be applied in a separate procedure to increase the accuracy of crop area estima-
tion at the regional to provincial scale by linking these local, very accurate spatially
discrete results to national wall-to-wall continuous crop classification maps.
A separate quantitative accuracy assessment for segmentation needs to be
conducted. In spite of the considerable progress in segmentation algorithms, there is
still lack of a global segmentation quality measure [6], and visual assessment is still
222 A. Kalaivani and R. Khilar

widely used and required. Moreover, since only spectral information of segments
(objects) was utilized in the proposed framework and not shape (or geometry), the
geometry of the segment does not have much impact on the classification results. In
other words, as long as there is oversegmentation for crop fields (i.e., segments
are smaller than crop fields), the segmentation results are valid. Therefore, it is
important to, visually, ensure that there is no undersegmentation (one segment
covering more than one crop field) in image segmentation. In multi-resolution
segmentation, this can be controlled by selecting a scale parameter that results in
segments that are smaller than actual crop fields. With several SAR polarimetric
indicators and features extracted from the decomposition and scattering matrix
in the context of OBIA, it is expected to achieve improved classification results
compared to what was achieved solely based on the four intensity channels. This is
recommended to be explored in a future research.
Cankut Ormeci et al. [17] are the first who use two pixel-based classification
algorithm unsupervised classification and maximum likelihood supervised classifi-
cation on the ISODATA and object-based classification algorithm. The accuracies
of classifications achieved by the confusion are overall accuracy, user’s accuracy,
producer’s accuracy, and kappa coefficient. An error matrix is obtained as a simple
cross tabulation of mapped class label and data for a sample of cases at specified
locations. The methodology searches the efficiency of satellite images in crop type
and area is determined. The performance of crop areas identification was acceptable
both by pixel-based classification and object-based classification techniques, but
the object-based classification gives superior results compared to pixel-based
classification techniques. Rose M. Rustowicz et al. [18] provides a comparison of
mono- vs. multi-temporal results and justifies that temporal information is helpful
for successful crop classification. The test accuracies are 92% for the multi-temporal
case and 85% for the mono-temporal case. When input data features are compared,
the current data points are normalized spectral irradiance values at satellite sensor.
The system is adaptable for more classes and a more sporadic placement of crop
types. The areas need to be extend and individual needs to get involved in order to
improve the above methodology.
The approach explained by Gong Cheng et al. [3] are a comprehensive review
of the recent progress in this field and proposed a large-scale, publicly available
benchmark dataset by analyzing the limitations of existing datasets. The current
state-of-the-art scene classification methods performed on NWPU-RESISC4 are
investigated. Users performed knowledge discovery by crowdsourcing of infor-
mation through these location-based social media data. The methodology maps
“what-is-where” easily on the surface of the Earth using the “what” and “where”
aspects of the information. A comparison is carried out that with remote sensing
images, the ground photos uploaded by user hold higher resolution and are quite
different from satellite remote sensing in the observation direction, which can well
capture the detail and vertical characteristics of ground objects. The additional
information are very useful for the classification and recognition of remote sensing
images.
10 Crop Classification and Mapping for Agricultural Land from Satellite Images 223

New methods need to be proposed which combine remote sensing data and
information coming from social media and spatial technology which can be
deployed to promote the state-of-the-art remote sensing image scene classification.

10.4 Remote Sensing System Data Classification

Remote sensing system classification techniques are classified into three categories,
i.e., manual, automated, and hybrid classifications. Manual satellite image classi-
fication approaches are robust and effective methods. Most of the time, manual
methods consume more time. In manual methods, the analyst must have high degree
of familiarity with the area covered by the satellite image. Efficiency and accuracy of
the classification largely depend on analyst subject knowledge and familiarity of the
theme under study. Automated remote sensing classification methods use algorithms
that are applied systematically to the entire image for grouping meaningful pixels
into various categories. Majority of the classification methods falls under this
category. Hybrid image classification methods combine the advantages of both the
methods. These methods use automated satellite image classification methods to do
initial classification; further manual methods are used to refine classification output
by adding value to it based on the analyst’s subject knowledge. Automated data
classification methods are classified into supervised and unsupervised classification
methods. Supervised classifiers can be parametric or nonparametric. Parametric
classification is based on the observed measurement vectors obtained for each
class in each spectral band during the training phase and is normally distributed
Gaussian vectors. In nonparametric classification, no such assumption is made and
the maximum likelihood classification is the popular and widely used supervised
classifier. The most important data classification is discussed in detail.

10.4.1 Minimum Distance from Mean (MDM) Classification

MDM is the simple supervised classification technique which determines the

spectral distance between the measurement vector for the pixel and the mean
vector for each training class signature. It is used to classify unknown image pixels
into classes which minimizes the distance between the image pixel and the class
in multifeatured space. To compute minimum distance between the data and the
class having separable class variances, Euclidean distance or normalized euclidean
distance is used. It requires the least computational time among other supervised
methods, and class variability is not taken into consideration by this method. The
spectral feature space formed by the multisensory data needs to be normalized with
respect to different radiometric ranges present in the datasets to produce the desired
results.
224 A. Kalaivani and R. Khilar

10.4.2 Mahalanobis Distance Classification (MDC)

The Mahalanobis distance is calculated as group in m-dimensional space defined by

m variables and their covariance. This method is used if there is a correlation among
the axes in feature space. This method takes the variability of classes which is more
useful than minimum distance which considers weighting factors. However, this
method tends to overclassify signatures with relatively large values in the covariance
matrix. The proposed method is slower in computation and depends on a normal
distribution of the data in each input band.

10.4.3 Maximum Likelihood Classification

MLC is one of the most popular supervised classification techniques used in the
analysis of RS data. The proposed methods define the posterior probability of a pixel
belonging to a particular class. The MLC for each class in each band follows normal
distribution and calculates the probability accordingly. The pixels are classified into
the class with the highest probability, and the other pixel remains unclassified.
The proposed method needs long time of computation and depends on a normal
distribution of the data in each input band. Sufficient ground truth/training sites
have to be selected for computing the variance-covariance matrices of population.

10.4.4 K-Means Classification

The K-means algorithm is one of the simplest unsupervised learning algorithms

that solve the well-known clustering problem. The K-means clustering algorithm is
a partition-based cluster analysis method which is the most popular unsupervised
clustering classification techniques. In clustering techniques, the pixels are grouped
into different clusters and the clusters exhibit property of homogeneity among the
same cluster and heterogeneity between different clusters. In K-means, K stands for
number of clusters to be formed which may be defined by user or from any automatic
techniques or initializes K randomly. The algorithm is iterative in nature and each
pixel is assigned to an exclusive cluster. In the iteration process, it minimizes within
cluster scatter and the process repeats until the scatter is less than a threshold
value or reaches the maximum number of iterations. The image is divided into
K clusters and the mean of each cluster is computed. In the next iteration, the
pixels are assigned to the nearest classes and new class means are computed. The
final partitioning of the clusters is based on the final K means obtained after the
completion of iteration, hence the name K-means.
10 Crop Classification and Mapping for Agricultural Land from Satellite Images 225

10.4.5 ISODATA

Iterative Self-Organizing Data Analysis (ISODATA) is the most popular unsuper-

vised clustering method. It is used to partition the spectral image into number
of spectral classes based on the statistical information inherent in the image.
ISODATA creates predefined number of clusters in a satellite image. Cluster centers
are randomly placed and pixels are assigned based on the shortest distance to
center method. The standard deviation within each cluster and the distance between
cluster centers are calculated. Clusters are further divided if standard deviation is
greater than the user-defined threshold otherwise they are merged. Further iterations
continue until either the average intercenter distance falls below the user-defined
threshold, the average change in the intercenter distance between iterations is less
than a threshold, or the maximum number of iterations is reached. Unsupervised
classification can produce more accurate results than supervised classification
which are ideal to large, complex, and heterogeneous areas with lack of intimate
familiarization and field information.

10.4.6 Fuzzy Classification

In fuzzy classification, the image pixels are grouped into a fuzzy set whose
membership function truth is the value of a fuzzy propositional function. A fuzzy
set allows its members to have different grades of membership in the interval [0, 1].
Fuzzy classification is a soft classification which takes heterogeneous and imprecise
nature (mix pixels) of the real-world scenario. Proportions of the multiple classes
within a pixel (e.g., 30% bare soil, 30% forest, and 40% crop) can be obtained. In the
case of a hard classification (ISODATA, K-means, MLC, etc.), each pixel belongs
to the class, and in the case of soft classification, each pixel can belong to more
than one class and has membership grades for each class. If needed, defuzzification
can produce a crisp result (one pixel to only one class) from fuzzy membership
grades. In fuzzy classification, the boundary between two neighboring classes is
assumed as a continuous, overlapping area within which an object has partial
membership in each class. This viewpoint reflects the reality of many applications
in which categories have fuzzy boundaries. The fuzzy classifier is computationally
complicated and is not recommendable, if conventional approach yields a satisfying
result.

10.4.7 Artificial Neural Networks (ANN)

Artificial neural network is a parallel distributed processor made up of simple pro-

cessing units called neurons. Artificial neural network or ANN is a computational
226 A. Kalaivani and R. Khilar

system that resembles the organizational principles present in biological nervous

systems. It has a normal tendency for storing experiential knowledge like a human
brain and uses this for pattern recognition. In ANN, the basic computational element
is known as the neuron or node. The neuron processes data in stages. First, the
messages received are aggregated by an internal activation function. Then, the
information is sent to transfer functions, which determine whether the neurons will
send the output message or not. Multiple neurons are connected together in layers.
These layers called the input layers are set up to receive input information, process
the data through one or more hidden layers, and produce a corresponding output
pattern through the output layer.
ANN have a distinct advantage over statistical classification methods in
achieving higher training accuracy—the capabilities of ANN for nonlinear
function approximation, data classification, nonparametric regression, and nonlinear
decision-making. ANN approach are ideally suitable to crucial applications such
as land use/cover-related feature classification from a satellite imagery. They are
nonparametric and require little or no a priori knowledge of the distribution model
of input data. ANNs have high processing speed, robustness, and capability to deal
with high-dimensional data spaces.

10.4.8 Support Vector Machines (SVM)

Support vector machine (SVM) is a supervised nonparametric learning technique

independent of any assumptions for the underlying data distribution. It finds a
hyperplane which separates the dataset into a discrete predefined number of classes.
SVMs are linear binary classifiers and assume that the multispectral feature data are
linearly separable in the input space. SVMs are not widely used by remote sensing
community as SVM performs similar to other established method.

10.4.9 Decision Rule-Based Tree Classification

Decision tree classifier is one of the most popular supervised classification methods,
which is nonparametric and does not require the data in normal distribution.
Decision tree classifiers are easy to train and they learn quickly from examples. It
follows a tree-structured graph or model of decision rules and their possible result.
Following a tree architecture, it is composed of a root node and a series of internal
nodes and leaf nodes. Every node can have only one father node and two or more
child nodes. Nodes are connected with each other by branches. Each node is passed
through certain test properties. Similarly, each leaf node corresponds to a class
property. Not only decisions rules are designed following a tree architecture, but
also a group of IF-THEN rules is also used.
10 Crop Classification and Mapping for Agricultural Land from Satellite Images 227

Classification rules in turn are easy to interpret and can serve as a knowledge
base for further classification of satellite image. It is easy to insert additional layers
of ancillary data with decision trees due to its nonparametric nature. The basic
scheme in decision tree classification is to mask every target as an image layer,
so that the influence of one target on the other is minimum. In comparison to
decision properties, rules are more popular, because of their simplicity, flexibility,
and convenience to use to build up the base of an expert system. The decision tree
algorithms are ID3, CD4.5, CART, etc. Decision tree classification is widely used
for the classification of remote sensing images for the extraction of information and
utilization of land use coverage.

10.5 System Methodology

The Landsat datasets are taken from UCI machine repository generated by the
Australian Centre for Remote Sensing. A sample database includes 82 rows and 100
columns from the original data. The binary values were converted to their present
ASCII and the classification for each pixel was performed. This data is of numerical
and at a single resolution and is highly suitable for standard maximum likelihood
classification (Fig. 10.3).
The Landsat MSS imagery consists of four digital images in which two of them
are in the visible region mostly in green and red regions and two are in the infrared
regions. Each pixel is an 8-bit binary word, with 0 corresponding to black and 255
to white, and the pixel spatial resolution is about 80m x 80m. Each image contains

Training Phase
Landsatellite Feature Supervised
Images Extraction Model

Testing Phase

Landsatellite Feature
Test Image Extraction

Land
Classification

Fig. 10.3 Proposed system methodologies

228 A. Kalaivani and R. Khilar

2340 × 3380 such pixels. The database is a (tiny) subarea of a scene, consisting
of 82 × 100 pixels. Each line of data corresponds to a 3 × 3 square neighborhood
of pixels completely contained within the 82 × 100 subarea. Each line contains
the pixel values in the four spectral bands (converted to ASCII) of each of the 9
pixels in the 3 × 3 neighborhood and a number indicating the classification label
of the central pixel. The classes used in the datasets are red soil, cotton crop, gray
soil, damp gray soil, soil with vegetation stubble, and very damp gray soil. The four
spectral values for the top-left pixel are followed by the four spectral values for the
top-middle pixel and then those for the top-right pixel and so on with the pixels
read out in sequence from left to right and top to bottom. The spectral values for the
central pixel are given by attributes 17, 18, 19, and 20.

10.6 Results and Discussion

For the experimentation, Weka toolkit is used for land coverage data and the
performance and effectiveness of the data are identified through prediction models.
The various Feature Selection Methods chosen are cfssubset evaluation, consistency
subset evaluation, and filtered subset evaluation. The features selected and the
features subjects are listed in Table 10.4.
The classifier model is built for the chosen feature subset such as J48 classifier,
BF Tree, J48 Graft, NB Tree, Random Forest, and LAD Tree supervised classifier.
The performance of an identified classifier is validated based on error rate and
computation time. The classification accuracies are measured in terms of precision,
recall, and F-score.
The evaluation parameters are the correctly classified instance (CCI) and incor-
rectly classified instance (ICCI). The error measures recorded for the classifier
performance are mean absolute error (MAE) and root mean square error (RMSE).
The classifier performance results are shown from Tables 10.5, 10.6, 10.7, 10.8,
10.9, and 10.10.

Table 10.4 Optimal feature subset

Feature selection methods Search methods Features selected Total features subset
Cfssubseteval Best first 1,2,4,5,6,9,10,12,13, 26
14,16,17,18,20,21,
22,24,25,26,28,29,
32,33,34,36,class
Consistency subseteval Best first 4,5,9,11,13,16,17,18, 13
19,23,27,33,36,class
Filtered subseteval Best first 1,2,4,5,6,9,10,12,13, 24
14,16,17,18,20,21,
22,24,25,28,29,
32,33,34,36, class
10 Crop Classification and Mapping for Agricultural Land from Satellite Images 229

Table 10.5 J48 classifier results

Performance Accuracy Error measure
Feature selection methods CCI ICCI Prec Recall F-Score MAE RMSE
AD features 81.67 18.32 81.11 81.17 81.20 5.6 21.47
Cfssubseteval 82.45 17.50 81.5 82.4 81.7 5.63 21.17
Consistency subeval 82.26 17.73 81.4 82.3 81.6 5.73 20.99
Filtered subseteval 82.44 17.55 81.5 82.4 81.7 5.63 21.17

Table 10.6 BF classifier results

Performance Accuracy Error measure
Feature selection methods CCI ICCI Prec Recall F-score MAE RMSE
All features 85.18 16.81 82.1 83.6 82.2 6.8 19.85
Cfssubseteval 83.59 16.40 82 83.6 82.2 6.83 19.8
Consistency subeval 82.90 17.09 82.1 82.9 82.3 5.86 20.5
Filtered subseteval 83.59 16.40 82.0 83.6 82.2 6.83 19.8

Table 10.7 J48 graft classifier results

Performance Accuracy Error measure
Feature selection methods CCI ICCI Prec Recall F-score MAE RMSE
All features 85.57 14.12 81.4 83.0 82.0 5.26 20.86
Cfssubseteval 83.45 16.54 81.7 83.5 82.3 5.37 20.68
Consistency subeval 82.63 17.36 81.4 82.6 8l.8 5.61 20.8
Filtered subseteval 83.45 17.36 81.7 83.5 82.3 5.37 20.68

Table 10.8 NB tree classifier results

Performance Accuracy Error measure
Feature selection methods CCI ICCI Prec Recall F-score MAE RMSE
All features 76.06 23.03 81.0 78.9 79.3 6.12 23.78
Cfssubseteval 78.56 21.43 80.5 78.6 78.9 6.11 23.27
Consistency subeval 77.60 22.39 78.8 77.6 77.4 6.52 23.4
Filtered subseteval 78.56 21.43 80.5 78.6 78.9 6.11 23.27

Table 10.9 Random forest classifier results

Performance Accuracy Error measure
Feature selection methods CCI ICCI Prec Recall F-score MAE RMSE
All features 83.11 16.81 87.1 88.3 87.0 5.81 15.88
Cfssubseteval 87.75 12.24 87.0 87.8 86.5 5.82 15.99
Consistency subeval 79.15 20.81 74.8 79.2 76.6 8.3 20.3
Filtered subseteval 87.75 12.24 87.0 87.8 86.5 5.82 15.99

The performance metrics of CCI and ICCI for different supervised tree classifier
chosen for efficient features subsets are listed in Table 10.11. The comparative
performance measure graph is shown in Fig. 10.4.
230 A. Kalaivani and R. Khilar

Table 10.10 LAD classifier results

Performance Accuracy Error measure
Feature selection methods CCI ICCI Prec Recall F-score MAE RMSE
All features 85.75 14.12 77.2 80.4 78.7 80.25 20.26
Cfssubseteval 87.75 12.24 87.0 87.8 86.5 5.82 15.99
Consistency subeval 79.15 20.81 74. 8 79.2 76.6 8.3 20.3
Filtered subseteval 80.89 19.10 77.8 80.9 99.2 8.05 19.98

Table 10.11 Comparative Supervised classifiers Performance

classifier performance
CCI ICCI
analysis
All features 81.67 18.32
J48 classifier 82.45 17.5
B-Tree classifier 83.59 16.40
J48 graft 83.45 16.54
NB tree 78.56 21.43
Random forest 87.75 12.24
LAD tree 87.75 12.24

90
80
70
Performance

60
50
40
30
20
10
0
s

er

ft

ee

st

ee
fie
re

ra

re
fi

Tr

Tr
tu

si

si

Fo
s

s
ea

D
8
la

la

N
J4

LA
m
lF

do
8

ee
Al

J4

an
Tr

R
BF

Classifiers
Performance CCI Performance ICCI

Fig. 10.4 Comparative classifier performance analysis

The classifier accuracy in terms of precision, recall, and F-score is listed in Table
10.12. The comparative performance measure graph is shown in Fig. 10.5.
The classifier error measures in MAE and RMSE are listed in Table 10.13. The
comparative performance measure graph is shown in Fig. 10.6.
From the experimental results analyzed, cfs subset feature subset selection
method produces the optimal features of 26 features from the total features of 37
features. The optimal subset features produces the better classifier accuracy results
10 Crop Classification and Mapping for Agricultural Land from Satellite Images 231

Table 10.12 Comparative Supervised classifiers Precision Recall F-score

classifier accuracy analysis
All features 81.11 81.17 81.20
J48 classifier 81.11 81.17 81.20
B-tree classifier 82.1 83.6 82.2
J48 graft 81.4 83.0 82.2
NB tree 81.0 78.9 79.3
Random forest 87.1 88.3 87.0
LAD tree 77.2 80.4 78.7

90
88
86
84
Accuracy

82
80
78
76
74
72
70
s

er

er

ft

ee

st

ee
re

ra

re
ifi

ifi

Tr

Tr
tu

G
ss

ss

Fo
ea

D
8
la

la

N
J4

LA
m
lF

do
8

ee
Al

J4

an
Tr

R
BF

Supervised Classifiers
Accuracy Precision Accuracy Recall Accuracy F-Score

Fig. 10.5 Comparative classifier performance analysis

Table 10.13 Comparative Supervised classifiers MAE RMSE

classifier error measure
analysis All features 5.4 20.12
J48 classifier 5.6 21.47
B-tree classifier 6.8 19.85
J48 graft 5.26 20.86
NB tree 6.12 23.78
Random forest 5.81 15.28
LAD tree 5.82 15.99

with supervised classifiers. The supervised tree classifier such as J48 classifier, BF
Tree, J48 Graft, NB Tree, Random Forest, and LAD Tree classifier are compared
on the feature subsets. The comparative metrics chosen for assessment of better
classifier are on the performance, accuracy, and error measures. Random Forest
classifier produces outperforming results with other supervised classifier. So it
is concluded that Random Forest classifier seems to be the better model for
identification of crop category. The Random Forest classifier identifies the crop
classification correctly for 88% and 12% only produces misclassification.
232 A. Kalaivani and R. Khilar

25

20

15

10

0
es

er

er

ft

ee

st

ee
ra

re
ifi

ifi
ur

Tr

Tr
G
ss

ss

Fo
at

D
8
la

la
Fe

N
J4

LA
m
C

do
l

ee
Al

J4

an
Tr

R
BF

Error Measures MAE Error Measures RMSE

Fig. 10.6 Comparative classifier error measure analysis

10.7 Conclusion

In this chapter, crop classification and mapping technique have been devised for crop
classification and mapping, and their performance is also measured and compared
with other supervised classifier model. The produced classification results are very
much promising with 88% accuracy of correct classification, and F-score is of 87%
with reduced error measures of 12%. The proposed method may provide an adequate
support to the farmers in differentiating crops suitable for the given land cover with
high accuracy and of low error measures. The research can be further enhanced to
develop better preprocessing techniques with an improved classifier performance to
reduce both false positives and false negatives, employing high-resolution real-time
satellite images of Indian agriculture system.

References

1. Barrett R, Crowther P, Laurence R, Lincolne R (2000) Agricultural crop identification using

spot and landsat images in Tasmania. In: International archives of photogrammetry and remote
sensing. vol XXXIII, Part B7. Amsterdam
2. Beeresh HV, Latha BM, Thimmaraja Yadava G, Dandur N (2014) An approach for identifi-
cation and classification of crops using multispectral images. Int J Eng Res Technol (IJERT)
3(5). https://doi.org/10.1007/978-3-319-11933-5_65. ISSN: 2278-0181
3. Cheng G, Han J, Lu X (2017) Remote sensing image scene classification: benchmark and state
of the art. Accepted by proceedings of the IEEE
4. Czajkowski K, Fitzgerald S, Foster I, Kesselman C (2001) Grid information services for
distributed resource sharing. In: 10th IEEE international symposium on high performance
distributed computing. IEEE Press, New York, pp 181–184
10 Crop Classification and Mapping for Agricultural Land from Satellite Images 233

5. Daniel M, Howard A, Bruce K, Wylie B, Larry L, Tieszen B (2012) Crop classification

modeling using remote sensing and environmental data in The Greater Platte River Basin,
USA. Int J Remote Sens 33(19):6094–6108
6. Dey A, Dwivedi BS, Bhattacharyya R, Datta SP, Meena MC, Das TK, Singh VK (2016)
Conservation agriculture in a rice-wheat cropping system on an alluvial soil of north-western
Indo-gangetic plains: effect on soil carbon and nitrogen pools. J Indian Soc Soil Sci 64(3):246–
254. https://doi.org/10.5958/0974-0228.2016.00034.7
7. Dhumal RK, Rajendra Y, Kale KV, Mehrotra SC (2013) Classification of crops from remotely
sensed images: an overview. Int J Eng Res Appl (IJERA) 3(3):758–761. ISSN: 2248-9622
www.Ijera.com
8. Durgun YÖ, Gobin A, Van De Kerchove R, Tychon B (2016) Crop area mapping using 100-M
Proba-V time series. Remote Sens 8:585. https://doi.org/10.3390/Rs8070585
9. Foster I, Kesselman C (1999) The grid: blueprint for a new computing infrastructure. Morgan
Kaufmann, San Francisco
10. Lugonja P, Brkljac B, Brunet B (2014) Classification of small agricultural fields using
combined landsat-8 and rapideye imagery: case study of Northern Serbia Vladimir Crnojevic.
J Appl Remote Sens 8(1):083512
11. Foster I, Kesselman C, Nick J, Tuecke S (2002) The physiology of the grid: an open grid
services architecture for distributed systems integration. Technical report, Global Grid Forum
12. Tiana H, Bangera K, Bo T, Dadhwal VK (2014) History of land use in India during 1880–2010:
large-scale land transformations reconstructed from satellite data and historical archives. Glob
Planet Change 121:78–88. https://doi.org/10.1016/j.gloplacha.2014.07.005
13. Kussul N, Lavreniuk M, Skakun S, Shelestov A (2017) Deep learning classification of land
cover and crop types using remote sensing data. IEEE Geoscı Remote Sens Lett 14(5):1–60
14. May P, Ehrlich HC, Steinke T (2006) ZIB structure prediction pipeline: composing a complex
biological workflow through web services. In: Nagel WE, Walter WV, Lehner W (eds) Euro-
Par 2006. LNCS, vol 4128. Springer, Heidelberg, pp 1148–1158
15. Moorthi MS, Oza MP, Misra I, Gambhir RK, Darji NP, Sharma S, Jain DK, Dhar D, Padia K,
Ramakrishnan R, Chowdhury S, Parihar JS (2014) FASALSoft – An ISRO software framework
for crop production forecast using remote sensing data analysis. J Geomatics 8(1):27–33
16. Nigam R, Bhagia N, Vyas S, Manjunath R (2015) Estimation of rabi crop area progression
from INSAT 3A CCD, SAC/EPSA/BPSG/CAD/FASAL-R&D/SR/03/2015
17. Ormeci C, Alganci U, Sertel E (2010) Turkey identification of crop areas using spot – 5 data,
an article
18. Rustowicz RM (2017) Crop classification with multi-temporal satellite imagery, an article
19. Salehi B, Daneshfar B, Davidson AM (2017) Accurate crop-type classification using multi-
temporal optical and multi-polarization SAR data in an objectbased image analysis framework.
Int J Remote Sens 38(14):4130–4155. https://doi.org/10.1080/01431161.2017.1317933
20. Schmedtmann J, Campagnolo ML (2015) Reliable crop identification with satellite imagery in
the context of common agriculture policy subsidy control. Campagnolo Remote Sens 7:9325–
9346
21. Singla N, Babbar B (2015) Rodent damage and infestation in wheat and rice crop fields: district
wise analysis in Punjab State. Indian J Ecol 37(2):184–188
22. Smith TF, Waterman MS (1981) Identification of common molecular subsequences. J Mol Biol
147:195–197