Global Advanced Research Journal of Agricultural Science (ISSN: 2315-5094) Vol. 4(11) pp. 810-818, November, 2015.

Available online http://garj.org/garjas/home

Copyright 2015 Global Advanced Research Journals

Full Length Research Paper

A Comparison of Classification Techniques for the Land

Use/ Land Cover Classification
R. S. Morgan*a, I. S. Rahim*a, M. Abd El-Hady*b
*a *b
Soils and W ater Use dept., W ater Relations and Field irrigation Dept.,
National Research Centre, El Behouth St., Dokki, Cairo, Egypt
Accepted 29 November, 2015

Land use/land cover classification plays an important role in the sustainable management of the natural
resources. Remote sensing techniques have been intensively used for land use /land cover
classification. Over the last decade new technologies had emerged that increased the accuracy of the
produced maps. The aim of this work is to investigate for an effective image classifier by comparing
three classifiers i.e. Maximum Likelihood classifier (MLC), Artificial Neural Network (ANN) and Support
Vector Machine (SVM) for land use/ land cover classification for a selected area in the north of the Delta
+
of Egypt. In this work, the surface reflectance data of Landsat 7 ETM was used. To reduce the
redundancy in the data, the Principal Component Analysis (PCA) was used and the first three PCAs
were selected for further analysis. The accuracy of the three classification method was compared using
the confusion matrix. In this work the overall accuracy of ANN was satisfactory using both the ENVI's
Neural Net and the neural network Pattern Recognition tool of Matlab and therefore was not
recommended in this study area. The overall accuracy for MLC and SVM was 80.28 and80.64 %,
respectively. Both MLC and SVM showed similar overall accuracy with an advantage of MLC when
considering the producer and user accuracy. To enhance the classification results, the Normalized
Difference Water Index (NDWI) was included and the MLC and SVM were re-examined. This addition
enhanced the overall accuracy of MLC to 83.95 % while the SVM decreased to 79.03%. Thus, it is
recommended using the MLC when classifying the land use/ land cover in this study area using the
same approach.

Keywords: Maximum likelihood classification, artificial neural networks, support vector machine, land use/ land
cover, image classification

INTRODUCTION
Land use/land cover classification is essential in
developing sustainable land management strategies
especially those aiming to improve water quality and
*Corresponding Author's Email: randa_sgmm@yahoo.com
prevent land degradation. Remote sensing is considered a
rapid and lower cost method compared to ground surveys.
Image classification is one of the important parts in remote
sensing applications for the land use and land cover
mapping. However, it is not easy to generate a satisfactory
result because of the limitation of data and complexity of
the land use and cover types (Wong and Sarker, 2014).

The Maximum Likelihood classifier (MLC)is one of the
most popular classification method that was used for land
use/ cover classification (Shalaby and Tateishi, 2007;
Manandhar et al., 2009; Ahmad and Quegan, 2012;and
Abdikan et al., 2015). MLC is a parametric classifier which
assumes normal or near normal spectral distribution for
each feature of interest. Jensen (2005) reported that it is
based on the probability that a pixel belongs to a particular
class and takes the variability of classes into account by
using the covariance matrix. Nevertheless, it requires
sufficient ground truth data and it cannot be applied if the
distribution of the population does not follow the normal
distribution (Japan Association of Remote Sensing, 1999).
In real life applications, the nature of the distribution is
hardly known and therefore, it is preferable to use
nonparametric classifiers that are free from assumptions
(Moughal, 2013).
Artificial Neural Networks (ANN)are nonlinear analytic
techniques modeled after the hypothesized processes of
learning in the cognitive system and the neurological
functions of the brain and capable of predicting new
observations from other observations after executing a
process of so-called learning from existing data(Dell Inc.,
2013).They offer the potential to classify data based upon
a rapid match to overall patterns using previously
calculated weighting factors, rather than point by point
comparisons involving algorithmic logic applied to
individual (Miller et al., 1995). Furthermore, Neural
networks make no assumption about data distribution (Qiu
and Jensen, 2004).ANN are applicable in virtually every
situation in which a relationship between the predictor
variables and predicted variables exists, even when that
relationship is very complex and not easy to articulate in
the usual terms of "correlations" or "differences between
groups" (Dell Inc., 2013).Qiu and Jensen, (2004) reported
that improved image classification results compared to the
traditional methods and various studies documented such
improvements in land use/ land cover classification
(Ashish, 2002;Pacificiet al., 2009; Yuan et al., 2009 and
Ndehedehe et al., 2013).
Support vector machine (SVM) is one of the latest
developed image classification techniques, it is a
supervised classification method derived from statistical
learning theory, which describes the properties of learning
machines that allow them to give reliable predictions and
represent a major development in machine learning
algorithms (Ivanciuc, 2007). Also, it could separate the
classes with a decision surface that maximizes the margin
between the classes. That surface is called the optimal
hyper plane and the data points closest to the hyper plane
are called support vectors. That hyper plane represents the
optimal separation of linearly separable classes in decision
boundary space (Szuster et al., 2011; Thai et al., 2012 and
Moughal, 2013).SVM is a non-parametric binary classifier
that does not require the assumption of normality and is
Morgan et al. 811
insensitive to dimensionality. SVMs have often been found
to provide higher classification accuracies than other
widely used pattern recognition techniques, such as the
maximum likelihood. Furthermore, SVMs appear to be
especially advantageous in the presence of heterogeneous
classes for which only few training samples are available
(Moughal, 2013). Therefore, SVM is considered one of the
most favorable classification techniques (Carrao et al.,
2008; Szuster et al., 2011;Thai et al 2012 and Moughal,
2013).
The aim of this work is to investigate for an effective
image classifier by comparing three classifiers i.e.
Maximum Likelihood classifier (MLC), Artificial Neural
Network (ANN) and Support Vector Machine (SVM) for
land use/ land cover classification for a selected area in the
north of the Nile Delta, Egypt.
MATERIALS AND METHODS
Study Area
The study area is located in the northern part of the Nile
Delta, between longitudes 30 32' 52" and 31 8' 31" E and
latitudes 31 16' 26" and 31 36' 50" N, and covers an area
2
of about 1650 km with the Mediterranean Sea representing
its northern border (Map 1).The study area included El-
Burullus Lake, the sandy wetlands, the old clayey deltaic
soils, the relatively new reclaimed soils clayey soils(which
were dried from the clay flats and swamps that surrounded
the lake) and the new reclaimed sandy bare soils. The
main land use in the study area is agriculture; mostly field
crops in the north delta area are rice, barely, beans and
clover. Vegetables including tomatoes, potatoes,
watermelon and others are cultivated in small scattered
areas. Fish ponds are the second dominant land use in the
study area. El-Burullus Lake is a shallow brackish lake
characterized by a few species of abundant aquatic
vegetation.
The study area is subjected to numerous ecological threats
such as drying parts of the lake, the fast growth of fish
ponds mainly around the southern fringes of the lake, lake
water pollution due to the excessive inflow of agriculture
drainage water, and the fast growing new human
settlements. Furthermore, the investigated area is suffering
from water shortage, thus farmers frequently use drainage
water for supplementary irrigation officially or unofficially.
Therefore, its vital to have a consistent monitoring strategy
to keep track of the changes that is consistently occurring
in this area.
812. Glo. Adv. Res. J. Agric. Sci.
Map 1: Location map of the study area (Landsat 7, 742)
MATERIALS AND METHODS
+
The Landsat 7 ETM data acquired on 05/03/2005 was
employed in this study. The ENVI software was used to
pre- and process the data using three classifiers namely;
MLC, SVM and ANN. Furthermore, ANN was processed
using the pattern recognition tool of the Neural Network
application of Matlab software as well due to the poor
results obtained after various attempts to adjust the
network using the ENVI's Neural Net tool.
Data pre-processing
The surface reflectance bands were available separately
and stacked to produce a single multiband image and
thereafter was subsetted to the study area. The principal
component analysis (PCA) was adopted to reduce the
dimensionality and computational requirements for the
further processing of the data. The PCA is a technique
designed to reduce the redundancy in multispectral data
due tothe high correlation between bands which gives
mixed pixels and complicate classification (Pemberty,
2002). It allows redundant data to be compacted into fewer
bands, by which the dimensionality of the data is reduced.
The bands of PCA data will be non-correlated and
independent, and are ordered in terms of the amount of
variance. The first two or three components will carry most
of the real information of the original data set
(Santiesteban, 2003).
Accuracy Assessment
The training samples were used to train MLC, SAM and
ANN classifiers to categorize the land cover/ land cover
classes of the study area using Landsat data. The
confusion matrix was used to assess the overall accuracy
measures for all the three classification procedures by
using the available ground truth pixels. Out of the sixty
ground truth data seventy percent of the data were used
for training and thirty percent for evaluation. Both the
overall accuracy and the kappa coefficient were considered
in the evaluation. The overall accuracy is calculated as the
percentage of all validation pixels correctly classified to the
total number of pixels in all the ground truth classes. The
pixels classified correctly are found along the diagonal of
the confusion matrix table. The kappa coefficient measures
the agreement between classification and ground truth
pixels. A kappa value of 1 represents perfect agreement
while a value of 0 represents no agreement.
RESULTS AND DISCUSSION
Utilizing the ENVI software the unsupervised classification
was applied to collect the required information about the
study area and to choose the training sets which represent
the different land use/land cover classes. Based on field
work they included the following seven classes: sandy
wetlands, sandy bare soils, field crops, water, aquatic
 Morgan et al. 813

vegetation, fish ponds and settlements. The separability of
the training sets, represented by region of interests (ROI),
was evaluated to examine spectral separability between
each two classes. The more separable training data is the
better classification is expected that relies on that data to
be. The separability between classes was always about
1.9(more 1.88) according to both Jeffries-Matusita and
Transformed Divergence measures thus indicating that
each class is properly separated from each other.
Three classification techniques were adopted in this
study to classify the land use/ land cover classes, namely
MLC,SVM and ANN, and tested using the confusion
matrix. The overall classification accuracy from MLC and
SVM classifiers were 80.28% and 80.64%, respectively.
Thus indicating that the classification ability of both
classifiers were the same when used in land use/ land
cover classification in this study area. In accordance, the
Kappa coefficient were almost the same (0.764for MLC
and 0.768 for SVM), indicating a similar agreement
between the classification and ground truth data in case of
MLC and SVM. The output maps from the MLC and SVM
classification techniques are shown in Map 2 and 3,
respectively.
The ANN classifier attended only an accuracy of 23.33%
after various attempts to enhance the accuracy by
alternating the neural network parameters. Trying to

Map 2. Classification results according to MLC classification

Map 3. Classification results according to SVM classification

814. Glo. Adv. Res. J. Agric. Sci.
increase the accuracy of the ANN classifier, the data were
processed using Pattern recognition tool of the Neural
Network application of the Matlab software but still the
results were not satisfactory. The overall accuracy did not
exceed 65 % and extremely misclassified two classes in
particular the fish ponds and the sandy soils classes (Map
4).Therefore, the results of the ANN were discarded in this
study and not discussed further.
Based on the results of the confusion matrix of MLC
(Table 1) and SVM (Table 2), it could be concluded that the
water and settlements are the most critical classes in MLC
and SVM respectively. Only 59.46% of the water class was
classified correctly using the MLC, while the rest of the
pixels in the class were miss-classified as fish ponds
(20.95%) and aquatic vegetation (17.23%). On the other
hand, the same class showed better results using the SVM
with a percentage of 95.27%. Contrary, the MLC showed
. Map 4. Classification results according to ANN classification
Table 1. Confusion Matrix for MLC
better results when classifying the settlements class with
percentage of 70.95%, while with the SVM this class was
only 39.59% and was mostly misclassified as sandy bare
soils (49.59%).
The producer and user accuracy of the MLC and SVM
are shown in figure 1 a and b, respectively. Based on these
results, the odd classification behavior for the settlements
class classified by the SVM clas ifier could be observed. It
showed a very low producer accuracy accompanied by a
very high users accuracy. This classifier behavior is not
desirable. Therefore, contrary to the previous studies
(Moughal, 2013 and Wong and Sarker,2014), it is
recommended to use the MLC for land use/ land cover
classification in the study area especially when considering
the producer and user accuracy and taking into account
that the settlements class is the most changeable class in
the study area that needed to be monitored continually.

Table 2. Confusion Matrix for SVM
a b
Figure 1. producer and user accuracy (a- MLC and b- SVM )
Furthermore, to enhance the performance of the classifiers
especially in case of the MLC, the recommended classifier
in this study, to identify the surface water we included the
normalized difference water index (NDWI), which is
designed to maximize the reflectance of a water body by
using green wavelengths, minimize the low reflectance in
NIR of water bodies, and take advantage of the high
reflectance in NIR of vegetation and soil features(Karsli et
al., 2011).In addition, the NDWI is defined as
where Green and NIR are the reflectance of the green
and NIR bands, respectively. Using this index the water
body information will be enhanced and the background
(vegetation and soil features) information will be restricted
(Mcfeeters, 1996).
The NDWI was added to the first three PCAs and this
image was used in evaluating the two classifiers. The
overall accuracy of the MLC improved to 83.95 %, while
the SVM over all accuracy slightly declined(79.3 %) and
Morgan et al. 815
the same trend was observed with Kappa coefficient which
was 0.8081 with MLC and 0.7486 with SVM. The output
maps from the MLC and SVM classification techniques
after including the NW DI are shown in Map 5 and 6
respectively above.
The results of the confusion matrix are shown in table
3 for MLC and 4 for SVM. According to these results it
could be concluded that the water and settlements which
are the most critical classes showed enhanced accuracy
using MLC. The percentage increased from 59.46 to
68.58% for the water class and from to 70.95 to 84.73% in
the settlements class. On the other hand, the SVM
percentage decreased in case of the settlements class
from 39.00 to 30.27% and remained the same (95.27%) in
case of the water class. The rest of the classes were
almost the same for both classifiers.
The producer and user accuracy of the MLC and SVM
are shown in figure 2 a and b, respectively.
816. Glo. Adv. Res. J. Agric. Sci.

Map 5. Classification results according to MLC classification after including NDWI

Map 6. Classification results according to SVM classification after including NDWI

Table 3. Confusion Matrix for MLC after including the NDWI


Table 4. Confusion Matrix for SVM after including the NDWI
Figure 2:Producer and user accuracy after including the NDWI.(a. MLC and b.NDWI)
The SVM classifier showed a high producer accuracy and
low user accuracy and consequently, favoring the
classification of the MLC to be used in land use/ land cover
classification in the study area.
CONCLUSION
In this work the Land Use/ Land Cover classification
accuracy of remote sensing data using Maximum
Likelihood (MLC), Artificial Neural Network (ANN) and
Spectral Angle Mapper (SAM) have been compared.
Attaining this objective also included the selection of an
appropriate extracting features method which in this work
included using Principal Component Analysis (PCA) and
the use of the Normalized Difference Water Index (NDWI).
The resulted data from this study recommended the MLC is
more effective than the SVM in terms of classification
overall accuracy and producer-user accuracy.
Morgan et al. 817
This work also indicated that even so the ANN was not
appropriate for this study but a considerable improvement
in the accuracy could be attained by using Matlab instead
of ENVI and therefore it is recommended that further
studies should be applied is this aspect.
REFERENCES
Abdikan S, Bilgin G, Sanli FB, Uslu E, Ustuner M (2015).
"Enhancing land use classification with fusing dual-polarized Terra
SAR-X and multispectral Rapid Eye data", J. Appl. Remote Sens.
9(1):096054 (May 08, 2015); http://dx.doi.org/10.1117/1.JRS.9.096054.
Ahmad A, Quegan S (2012). Analysis of Maximum Likelihood
Classification on Multispectral Data. Applied Mathematical Sciences, 6
(129), 6425 - 6436.
Ashish D (2002). Land-Use Classification of Aerial Images Using Artificial
Neural Networks, Master Thesis, Graduate Faculty of The University of
Georgia, Athens, Georgia.
Carrao H, Goncalves P, Caetano M (2008). Contribution of multispectral
and multitemporal information from MODIS images to land cover
classification. Remote Sensing of Environment, 112 (3): 986997.
818. Glo. Adv. Res. J. Agric. Sci.

Dell Inc., (2013). Statistics: Methods and Applications. Tulsa, OK: Dell. Netherlands.Qiu F, Jensen JR (2004). Opening the black box of neural
(Electronic Version). http://documents.software.dell.com/statistics/current/textbook. networks for remote sensing image classification , INT. J. Remote
Ivanciuc O (2007). Applications of Support Vector Machines in Chemistry Sensing, 25 (9),17491768
Reviews. Computational Chemistry, vol.23, 291-400. Edited by K. B. Santiesteban IA (2003). Multi-sensor image fusion applied to
Lipkowitz and T. R. Cundari. Wiley-VCH2007, John Wiley & Sons, Inc. morphodynamics of coastal environments in Belize. M.Sc. Thesis,
Japan Association of Remote Sensing (JARS), (1999). Remote Sensing International Institute for Aerospace Survey and Earth Sciences (ITC),
Notes, Chapter 11, Image Processing - Classification. (Electronic Enschede, Netherlands.
version).http://wtlab.iis.u-tokyo.ac.jp/ ~wataru/lecture/rsgis/rsnote/contents.htm Shalaby A, Tateishi R (2007). Remote sensing and GIS for mapping and
Jensen JR (2005). Introductory Digital Image Processing: A Remote monitoring land cover and land-use changes in the Northwestern
Sensing Perspective, Third ed. Prentice Hall, Upper Saddle River, New coastal zone of Egypt. Applied Geography, 27 (1): 2841.
Jersey. Szuster BW, Chen Q, Borger M (2011). A comparison of classification
Karsli F, Gunerogluand A, Dihkan M (2001). Spatio-temporal shoreline techniques to support land cover and land use analysis in tropical
changes along the southern Black Sea coastal zone. J. Appl. Remote coastal zones. Applied Geography 31 (2): 525-532
Sens. 2011, 5, 053545. Thai LH, So T, Hai, Thuy NT (2012). Image Classification using Support
Manandhar R, Odeh OAI, Tiho A (2009). Improving the Accuracy of Land Vector Machine and Artificial Neural Network. I. J. Information
Use and Land Cover Classification of Landsat Data Using Post- Technology and Computer Science, 2012, 5, 32-38
Classification Enhancement. Remote Sensing, 1, 330-344.
Wong SN, Sarker MLR (2014). Land use/ land cover mapping using multi-
Mcfeeters SK (1996). The use of the normalized difference water index
scale texture processing of high resolution data. The 8th International
(NDWI) in the delineation of open water features. Int. J. Remote Sens.,
Symposium of the Digital Earth (ISDE8), IOP Conf., Series: Earth and
17, 14251432.
Environmental Science 18 (2014) 012185.
Miller DM, Kaminsky EJ, Rana S (1995). Neural Networks Classification of
Yuan H, Van Der Wiele CF, Khorram S (2009). An Automated Artificial
Remote Sensing Data. Computers and Geosciences Volume 21 (3):
Neural Network System for Land Use/Land Cover Classification from
377386.
Landsat TM Imager. Remote Sensing, 1, 243-265.
Moughal TA (2013). Hyperspectral image classification using Support
Vector Machine. 6th Vacuum and Surface Sciences Conference of
Asia and Australia (VASSCAA-6). Journal of Physics: Conference
Series 439 (2013) 012042.
Ndehedehe C, Ekpa A, Simeon O, Nse O (2013). Understanding the Neural Network
Technique for Classification of Remote Sensing Data Sets. New York Science
Journal, 6 (8) : 26-33.
Pacifici F, Chini M, Emery WJ (2009). A neural network approach using
multi-scale textural metrics from very high-resolution panchromatic
imagery for urban land-use classification. Remote Sensing of
Environment, 113 (6): 12761292.
Pemberty JFP (2002). Assessment of Morphodynamic Processes and Soil
Pollution as Indicators of Land Degradation. A Case Study in the
Tabernas Area, Southeastern Spain. M.Sc. Thesis, International
Institute for Aerospace Survey and Earth Sciences (ITC), Enschede,