Machine Translated by Google

International Journal of Geosciences, 2017, 8, 611-622

[Link]
ISSN Online: 2156-8367
ISSN Print: 2156-8359

Accuracy Assessment of Land Use/Land Cover

Classification Using Remote Sensing and GIS

Sophia S. Rwanga1,2*, JM Ndambuki3

1
Department of Civil Engineering, Tshwane University Technology, Pretoria, South Africa
2
Department of Civil Engineering, Vaal University of Technology, Vanderbijlpark, South Africa
3
Department of Civil Engineering, Tshwane University of Technology, Pretoria, South Africa

How to cite this paper: Rwanga, SS and Abstract

Ndambuki, JM (2017) Accuracy Assessment
of Land Use/Land Cover Classifica Remote sensing is one of the tools which is very important for the production
tion Using Remote Sensing and GIS. , Inter of land use and land cover maps through a process called image classification.
nationalJournal of Geosciences8, 611-622.
For the image classification process to be successful, several factors should
[Link]
be considered including the availability of quality Landsat imagery and
Received: February 14, 2017 secondary data, a precise classification process and the user's experience and
Accepted: April 27, 2017 expertise in the procedure. The objective of this research was to classify and
Published: April 30, 2017
map land-use/land-cover of the study area using remote sensing and Geospatial
Copyright © 2017 by authors and
Information System (GIS) techniques. This study includes two sections (1)
Scientific Research Publishing Inc. Landuse/Landcover (LULC) classification and (2) accuracy assessment. In this
This work is licensed under the Creative
study supervised classification was performed using Non Parametric Rule.
Commons Attribution International
License (CC BY 4.0). The major LULC classified were agriculture (65.0%), water bodies (4.0%), and
[Link] built up areas (18.3%), mixed forest (5.2%), shrubs (7.0%), and Barren/bare
OpenAccess land (0.5%). The study had an overall classification accuracy of 81.7% and a
kappa coefficient (K) of 0.722. The kappa coefficient is rated as substantial and
hence the classified image was found to be fit for further research. This study
presents an essential source of information whereby planners and decision
makers can use to sustainably plan the environment.

Keywords
Accuracy assessment, Geographic Information Systems (GIS), Land Use Land
Cover (LULC), Remote Sensing

[Link]
Land use and land cover information is required for policy making, business and
administrative purposes. With their spatial details, the data are also crucial for
environmental protection and spatial planning. Landuse classification is vital

DOI: 10.4236/ijg.2017.84033 April 30, 2017

Machine Translated by Google

SS Rwanga, JM Ndambuki

because it gives data which can be used as input for modeling, especially the one
dealing with the environment, for example models deal with climate change and
policy developments [1]. Hench the combined LULC grant a comprehensive
means of understanding the interaction of geo-biophysical, socioeconomic
systems behaviors and interactions [2]. To provide more useful information in
land cover, Remote Sensing is often paired with a Geographic Information System
(GIS) technique.
Remote sensing is the main source for several kinds of thematic data critical to
GIS analyses, including data on land use and land cover characteristics. Aerial
and Landsat satellite images are also frequently used to evaluate land cover
distributions and to update existing geospatial features. With the introduction of
remote sensing systems and image processing software, the importance of
remote sensing in Geospatial Information Systems (GIS) has expanded
significantly [3]. The accelerated use of remote sensing data and techniques has
made geospatial processes faster and powerful, although the increased complexity
also creates in creased possibilities for error [4]. Previously, accuracy assessment
was not a priority in image classification studies. However, because of the
accelerated chances for error presented by digital imagery, accurate assessment
has become a very vital process [5].
Accuracy assessment or validation is a significant step in the processing of
remote sensing data. It establishes the information value of the resulting data to
a user. Productive utilization of geodata is only possible if the quality of the data
is known. The overall accuracy of the classified image compares how each of the
pixels is classified versus the definite land cover conditions obtained from their
corresponding ground truth data. Producer's accuracy measures errors of
omission, which is a measure of how well real-world land cover types can be
classiÿed. User's accuracy measures errors of commission, which represents the
like liability of a classified pixel matching the land cover type of its corresponding
real-world location [5] [6] [7]. The error matrix and kappa coefficient have come a
standard means of assessing image classification accuracy. Moreover, Error
matrix has been used in numerous land classification studies and
were a crucial component of this research.
The objective of this research was to classify and map land-use/land-cover of
the study area using remote sensing and Geospatial Information System (GIS)
techniques and to carry out accuracy assessments in order to find out how well
the classification procedure was undertaken and also to understand how to
interpret the usefulness of the classification.

Study Areas

The study area map was prepared from the Limpopo province map. The area
falls under latitude 23ÿ0'31.0956"S, 29ÿ30'48.5697"E and longitude 24ÿ2'48.3007"S
and 29ÿ32'16.9088"E. The total study area is 7138 ranges
km2 . The
fromrainfall
290.565(average)
mm to
1410.24 mm. The study area is shown in Figure 1.

612
Machine Translated by Google

SS Rwanga, JM Ndambuki

Figure 1. Study area map.

2. Materials and Methods

This paper covers two sections: 1) Landuse/Landcover (LULC) classification and 2)
accuracy assessment. The land use/cover classification of the study area and
accuracy assessment were carried out as per the methodology presented in Fig
ure 2.

Landuse/Landcover Classification
Image Pre-Processing
Classification process and analysis of the different LULC classes were done using
two Landsat satellite images covering the Landsat 8 OLI/TIS acquired on 16
September 2015. These images include; L8 OLI/TIRS (path 170, rows 68) and L8
OLI/TIRS (path 170, rows 77) (Table 1). The Landsat images were down loaded
from the United States Geological (USGS) Earth Explorer (https://
[Link]/). The selection of the Landsat satellite image date was
influenced by the quality of the image, especially for those with limited or low cloud
cover. Each Landsat was georeferenced to the WGS_84 datum and Universal
Transverse Mercator Zone 35 North coordinate system.
An intensive pre-processing such as geo-referencing, mosaic, and layers

613
Machine Translated by Google

SS Rwanga, JM Ndambuki

Figure 2. Schematic of work flow for LULC and accuracy assessment.

Table 1. Details of Landsat 8 OLI/TIS used for classification.

Grid cells
Satellites _ID sensors Path/row Layers Date of acquisition
size(m)

LC81700762015259LGN00 170/77
Landsat 8 OLI/TIS 11 Sept. 16, 2015 30
LC81700762015259LGN00 170/68

stacking were carried out in order to Ortho-rectify the satellite images. The im
age was then processed in ERDAS IMAGINE 2015 software. The satellite image
of each band was stacked in ERDAS Hexagon within the interpreter main icon
utility ties with layers stacked function. Then, from the stacked satellite image the
study area image was extracted by clipping the study area using ArcGIS 10.3 software.

Landuse/Landcover (LULC) Classification: Supervised

For this study, only supervised classification was performed. Supervised
classification according to [8] is where “the user develops the spectral signatures
of known categories, such as urban and forest, and then the software assigns
each pixel in the image to the cover type to which its signature is most comparable”.
"Supervised classification is the process most frequently used for quantitative
analyzes of remote sensing image data" [9]. The supervised classification was
applied after defined area of interest (AOI) which is called training classes. More
than one training area was used to represent a particular class. The training sites
were selected in agreement with the Landsat Image, Google Earth and Google
Map (Figure 3). The basic sequence of operations followed on supervised
classification was;
• definition of Training Sites : The first step in undertaking a supervised classifi

614
Machine Translated by Google

SS Rwanga, JM Ndambuki

cation is to define the areas that will be used as training sites for each land
cover class. This is usually done by using the on-screen digitized features.
The features created are called Areas of Interest (AOI). The selection of the
training sites identified was based on those areas clearly in all sources of im
ages. In this study, one hundred training sites have been identified.
• Extraction of Signatures: After the training site (AOI) being digitized, the next
step was to create statistical characterizations of each information. These are
called Signatures editors in ERDAS Imagine 2015. In this step, the goal was to
create a signal (SIG) file for every informational class. The SIG files contain a
variety of information about the land cover classes described. After the entire
signature has been created, then the SIG file is saved as dialog (Table 2).

• Classification of the Image (Supervised classification): The supervised

classiÿcation has been applied after defined training classes. One or more than
one training area was used to represent a particular class. During the supervised
classification process, the entire Signature editor was selected in order to be
used in the classification process. Then the classify was selected from the

Figure 3. Identification of training sites using Landsat image (Erdas Imagine 2015), Google earth and Google map.

Table 2. Signature editor table for classified image.

Class # Signature name color red Green Blue Value Order Count Prob.

1 Mixed forest 0.000 0.392 0.000 8 166 1267 1,000

2 Barren/bare land 0.824 0.706 0.549 3 168 87 1,000

3 Shrubs 0.101 0.899 0.730 6 169 50 1,000

4 Agricultures 0.000 1,000 0.000 10 171 78377 1,000

5 Built-up area 0.698 0.528 0.581 2 173 4628 1,000

6 Water bodies 0.000 0.000 1,000 1 174 5524 1,000

615
Machine Translated by Google

SS Rwanga, JM Ndambuki

Figure 4. Classified map of study areas.

Table 3. Landcover classification scheme.

land covers Description

Water body Lakes, reservoirs, streams, rivers, swamps

Land covered by buildings and other man-made structures. Residential,

Built-up areas
commercial services, industrial areas, mixed urban or built up lands

Lands with exposed soil, sand or rocks, and never have more than 10% vegetated
Barren/bare
land cover during any time of the year. Bare ground, bare exposed rocks, strip mines,
quarries and gravel pits

Lands with woody vegetation less than 2 meters tall. The shrub foliage can be
Shrubs
either evergreen or deciduous

Mixed forest Lands dominated by trees with a percent cover >60% and height exceeding 2
meters, Deciduous forest land and evergreen forest land

Lands covered with temporary crops followed by harvest period, Crop fields and
Agriculture
pastures

Editor Menu bar, classify/supervised. Non Parametric Rule was used in this
classification. The Image was classified into six classes namely; Water
bodies, Built up areas, Barren/bare land, shrubs, Mixed forest and Agriculture
(Table 3).

Classification Results and Discussion

Supervised classification was carried out at study area. The area of each class
was calculated taking into account the pixel count and total area (study area).
Thus allocations of each classified area, (percentage) are tabulated in Table 4.
The percentage of areas as classified are; Agriculture (65.0%), water body
(4.0%), built up areas (18.3%), mixed forest (5.2%), shrubs (7.0%), and Barren/bare land

616
Machine Translated by Google

SS Rwanga, JM Ndambuki

(0.5%) (See Figure 5). Agriculture was found to be the dominant type of
classified Land use which covers about 65.0% of the total study area, followed
by Built-up areas while the least classified was Barren/bare land which accounts
for 0.5%. During the classification, among the water bodies classified were
rivers (sand river and Houtriver).

3. Classification Accuracy Assessment

One of the most important final steps in the classification process is accuracy
assessment. The aim of accuracy assessment is to quantitatively assess how
effectively the pixels were sampled into the correct land cover classes. Moreover,
the key emphasis for accurate assessment of pixel selection was on areas that
could be clearly identified on both Landsat high resolution image, Google Earth
and Google Map. A total of 307 points (locations) were created in the classified
image of the study area. The Accuracy Assessment Cell Array Reference
column was filled according to the best guess of each reference point. hydrogeological

Table 4. Classified area under different Landuse classes in study area.

Land cover Area in km2 Percentage (%)

Classes area

Waterbody 283 4.0

Built-up areas 1309 18.3

Barren/bare land 37 0.5

Shrubs 499 7.0

Mixed forest 372 5.2

Agriculture 4638 65.0

Total 7138

Figure 5. Pie chart showing the distribution of classified areas in percentage.

617
Machine Translated by Google

SS Rwanga, JM Ndambuki

Figure 6. Landsat (classified) image of the study area covered with 307 points from random sampling.

Table 5. Theoretical error matrix of LULC classification.

water Built up Barren/bare Mixed

S. No Classified Shrubs Agriculture Total Correct sampled
body areas land forest

1 Water bodies 20 3 3 0 0 1 27 20

2 Built-up areas 2 61 23 1 3 2 92 61

3 Barrens/bare lands 0 0 12 0 0 0 12 12

4 Shrubs 0 2 4 25 0 3 34 25

5 Mixed forest 0 0 3 2 31 1 37 31

6 Agriculture 1 1 0 0 1 102 105 102

Total 23 67 45 28 35 109 307 251

map series of the republic of South Africa, Topographic map, Google earth and
Google Map were used as reference sources to classify the selected points.
Table 5 shows the relationship between ground truth data and the corres
ponding classified data obtained through error matrix reports.
The overall classification accuracy = No. of correct points/total number of
251
points = = 81.7% .
307
Table 5 shows a theoretical confusion matrix (error matrix) of a LULC
classification. The columns of the confusion matrix show to which classes the
pixels are in the validation set belong (ground truth) and the rows show to which
classes the image pixels have been assigned to in the image. The diagonal shows
the pix els that are classified correctly. Pixels that are not assigned to the proper
class do not occur in the diagonal and give an indication of the confusion between
the different land-cover classes in the class assignment.
Furthermore, the off-diagonal elements in the rows of the confusion matrix,
divided by the total number of pixels assigned to the landsat image class cores

618
Machine Translated by Google

SS Rwanga, JM Ndambuki

pondering to the row, represent the commission errors and describe the confusion
between that image class and describe the other land-cover classes. The com mission
errors describe the chance that a pixel that has been assigned to a particular class actually
belongs to one of the other classes.
Moreover, this study considered other metrics derived from the error matrix to further
describe the accuracy of assessments including; commission and omission error,
sensitivity and specificity, positive and negative predictive power and Kappa statistics. For
thorough information on these concepts, refer to [10] and [11].

In this research, various statistics related to classification accuracy as well as overall

Kappa statistics are computed based on [12] formulation as indicated be low:

a '
=
Sensitivity ( ) equivalent to Producer's Accuracy
ab +
d
Specificity =
bd +

Commission error 1 Specificity

=ÿ

Omission error 1 Sensitivity

=ÿ

a
Positive Predictive Power = ( Equivalent
) to User's accuracy
ab +
d
Negative Predictive power =
cd +

where:
a = number of times a classification agreed with the observed value =
b number of times a point was classified as X when it was observed to not be X.
c = number of times a point was not classified as X when it was observed to be
d X. = the number of times a point was not classified as X when it was not
N = ( a=+) b +
observed to be X. Total points c + d

KAPPA analysis is a discrete multivariate technique used in accurate assessments

[13]. KAPPA analysis yields a Khat statistic (an estimate of KAPPA) that is a measure of
agreement or accuracy [5]. The Khat statistics are computed as;
r r
ÿ ÿi 1
Nx = ii
ÿ

i =1
(xi Xx + 1+ )
K =
r
N
2 ÿ

ÿ i =1
(xiiXx 1+ )

where;
r = number of rows and columns in error matrix , N = total number of observers
vations (pixels )

XII = observation in row and column ii ,

Xi+ = marginal total of row , and X+ii = marginal total of column i

A Kappa coefficient equal to 1 means perfect agreement where as a value close to

zero means that the agreement is no better than would be expected by chance.
As per [14] categorization of Kappa statistic is widely referenced which is re produced in
Table 6.

619
Machine Translated by Google

SS Rwanga, JM Ndambuki

Table 6. Rating criteria of Kappa statistics.

[Link] Kappa statistics Strength of agreement

1 <0.00 Poor

2 0.00 - 0.20 Slight

3 0.21 - 0.40 Fair

4 0.41 - 0.60 Moderate

5 0.61 - 0.80 Substantial

6 0.81 - 1.00 Almost perfect

Table 7. Category wise accuracy assessment statistical parameters.

Observed proportion of Expected proportion of

Kappa coefficient (K)
agreements (Po) agreement (Pe)

0.9674 0.850 0.782

0.8795 0.613 0.689

0.8632 0.805 0.298

0.9609 0.818 0.785

0.9674 0.793 0.843

0.9707 0.547 0.935

Table 8. Category wise accuracy assessment statistical parameters.

Parameters
Classified Data
Sensitivity Specificity Commission Error Omission Error UA PA

Water bodies 0.8696 0.97535 0.0246 0.1304 0.741 0.870

Built-up areas 0.9104 0.87083 0.1292 0.0896 0.663 0910

Barren/bare
0.2667 0.96565 0.0344 0.7333 0.571 0.267
land

Shrubs 0.8929 0.96774 0.0323 0.1071 0.735 0893

Mixed forest 0.8857 0.97794 0.0221 0.1143 0.838 0.886

Agriculture 0.9444 0.98492 0.0151 0.0556 0.971 0936

Results and Discussion on Accuracy Assessments

Using the formulae furnished in section 3.0, the various accuracy evaluations of
the meters were computed and tabulated in Table 7 and Table 8.
The results from the accuracy assessment showed an overall accuracy obtained
from the random sampling process for the image of 81.7%. User's accuracy ranged
from 57.1% to 97.1% while producer's accuracy ranged from 26.7% to 93.6%. The
broad range of accuracy indicates a severe confusion of Barren/bare land with
other land cover classes. Moreover, the measure of producer's accuracy (Sensitivity)
reflects the accuracy of prediction of the particular category. The User's accuracy
reflects the reliability of the classification to the user. User's accuracy is the more
relevant measure of the classification's actual utility in the

620
Machine Translated by Google

SS Rwanga, JM Ndambuki

fields. Agriculture was found to be more reliable with 97.1% user accuracy.
The commission error reflects the points which are included in the category while
they really do not belong to that category. For instance, the commission error is
highest in the case of built - up areas which means that more number of points (31)
which do not fall under this category are classified as built up areas.
Equally, the omission error reflects the number of points which are not included in
the category while they really belong to the category. The omission error in the case
of Barren/bare land is more (0.7333) with 33 points which actually belong to this
category not being categorized in this class. In this study an overall Kappa coefficient
of 0.722 was obtained which was rated as substantial. Apart from over all
classification accuracy, the above individualized parameters give a classifier a more
detailed description of the model performance of a particular class or category of his
field of interest or study.

4. Conclusions
Remote sensing is very important for the production of Land Use / Land Cover maps
which can be done through a method called image classification. This me thod had
made huge improvements over the past decades in the following four areas for
example; LULC maps production at any scale, improvement and use of advanced
classification processes such as pre field and sub pixel, classification procedures
using knowledge base process and incorporation of auxiliary data into classification
procedures; such data includes, digital elevation model (DEM), road, soil, land use
and census data. Moreover classifying landsat image ries in order to obtain accurate
and reliable LULC information still remains a
a challenge that depends on several factors for example the selected images,
landscape complexity, image processing techniques and classification processes
themselves.
The accelerated usage of remote sensing data and techniques has made
geospatial process faster and powerful, although the increased complexity also
creates increased possibilities for error. The objective of this paper was to classify
and map land use - land cover (LULC) of the study area using Remote Sensing and
GIS techniques and also to carry out accurate assessments in order to assess how
well a classification worked.
The supervised classification was performed using the Non Parametric Rule. The
image was classified into six classes; Agriculture (4638 km2 ), water bodies (283
km2 ), built up areas (1309 km2 ), mixed forest (372 km2 ), shrubs (499 km2 ), and
Barren/bare land (37 km2 ). Agriculture was the dominant type of Landuse classiÿed
which covers about 65.0% of the total study.
In addition, classified images need to be assessed for accuracy, before the same
could be used as input for any application. Individual accuracy assessment
parameters are useful for assessing model performance in respect of a particular
category/class of specific interest for the study. In this study, the accuracy
assessment was performed using an error matrix. The study had an overall
classification accuracy of 81.7% and a kappa coefficient of 0.722. The kappa coefficient is rated

621
Machine Translated by Google

SS Rwanga, JM Ndambuki

as substantial and hence the classified image found to be fit for further research.

References
[1] Disperati, L., Gonario S. and Virdis, P. (2015) Assessment of Land-Use and Land Cover
Changes from 1965 to 2014 in Tam Giang-Cau Hai Lagoon, Central Viet
name Applied Geography , 58, 48-64.

[2] Moran, EF, Skole, DL and Turner, BL (2004) The Development of the International Land-Use
and Land-Cover Change (LUCC) Research Program and Its Links to NASA's Landcover and
Land-Use Change (LCLUC) Initiatives. Kluwer Academic Publication, Netherlands.

[3] Merchant, JW and Narumalani, S. (2009) Integrating Remote Sensing and Geo , Paper 216.
graphic Information Systems. Papers in Natural Resources
[Link]

[4] Murty, PS and Tiwari, H. (2015) Accuracy Assessment of Land Use Classification —A Case
Study of Ken Basin. Journal of Civil Engineering and Architecture Re
search , 2, 1199-1206.

[5] Congalton, RG (1991) A Review of Assessing the Accuracy of Classifications of Remotely

Sensed Data. , 37, 35-46. Remote Sensing of Environment
[Link]

[6] Campbell, JB (2007) Introduction to Remote Sensing. 4th Edition, The Guilford Press, New York.

[7] Jensen, JR (2005) Introductory Digital Image Processing: A Remote Sensing Press pective. 3rd
Edition, Pearson Prentice Hall, Upper Saddle River, NJ.

[8] Eastman, JR (2003) Guide to GIS and Image Processing 14, 239-247. Clark University Manual,
USA.

[9] Richards, J. and Jia, X. (2006) Remote Sensing Digital Image Analysis: An Introduction.
Springer, Berlin.

[10] Fielding, AH and JF Bell. (1997) A Review of Methods for the Assessment of Pre
diction Errors in Conservation Presence/Absence Models. Environmental Concerts
vation , 24, 38-49. [Link]

[11] Lurz, PWW, Rushton, SP, Wauters, LA, Bertolino, S., Currado, I., Mazzoglio, P. and Shirley,
MDF (2001) Predicting Gray Squirrel Expansion in North Italy: A Spatially Explicit Modeling
Landscape Ecology , 16, 407-420.
Approach. [Link]

[12] Jenness, J. and Wynne, JJ (2007) Kappa Analysis (kappa_stats.avx) Extension for ArcView
3.x. Jenness Enterprises. [Link]

[13] Jensen, JR (1996) Introductory Digital Image Processing: A Remote Sensing Press pective.
2nd Edition, Prentice Hall, Inc., Upper Saddle River, NJ.

[14] Landis, JR and Koch, GG (1977) A One-Way Components of Variance Model for Categorical
Data. , 33, [Link]
[Link]

622
Machine Translated by Google

Submit or recommend the next manuscript to SCIRP and we will provide the best
service for you:
Accepting pre-submission inquiries through Email, Facebook, LinkedIn, Twitter, etc.
A wide selection of journals (inclusive of 9 subjects, more than 200 journals)
Providing 24-hour high-quality service
User-friendly online submission system
Fair and swift peer-review system
Efficient typesetting and proofreading procedure
Display of the result of downloads and visits, as well as the number of cited articles
Maximum dissemination of your research work

Submit your manuscript at: [Link]

Or contact ijg@[Link]