Content uploaded by Chanida Suwanprasit
Author content
All content in this area was uploaded by Chanida Suwanprasit on Oct 26, 2018
Content may be subject to copyright.
39
Proceedings of the Asia-Pacific Advanced Network 2012 v. 33, p. 39-47.
http://dx.doi.org/10.7125/APAN.33.4
ISSN 2227-3026
Impacts of spatial resolution on land cover classification
Chanida Suwanprasit* and Naiyana Srichai
ANdaman Environment and natural Disaster research center (ANED)
Faculty of Technology and Environment, Prince of Songkla University, Phuket Campus /
80 Moo 1 Vichit-Songkram Rd., Kathu, Phuket, THAILAND ; Tel.: +6-676-276-140;
Fax: +6-676-276-102
E-Mails: chanida.s@phuket.psu.ac.th*; naiyana.s@phuket.psu.ac.th
Abstract: Regularly updated land cover information is a requirement for various land
management application. Remote sensing scenes can provide information highly
useful for real-time modeling of the earth environment. However, the spatial
resolution is also a very important factor to acquire the information on satellite
imagery. This paper summarizes the basic conclusions of work in which the spatial
resolution of satellite imagery, related to the factor of scale for land cover
classification, was investigated. Optical data collected by two different sensors
(THEOS with 15-m resolution and Landsat 5-TM with resolution 30-m) in 2010 were
tested against the ability to correctly classify specific land cover classes at different
scales of observation. Support Vector Machines (SVMs) classifier was used and
Kathu district, Phuket, Thailand was the study area. The land cover was classified into
7 groups as forest, built-up, road, water, agriculture, grassland and bare land. The
result indicated that the overall accuracy of THEOS with 15 m was slightly higher
than Landsat-5 TM with 30 m resolution (90.65% and 89.00%, respectively). The
outcome of the study can be discussed further to assess the suitable spatial resolution
for land cover classification mapping of Kathu district. Understanding the role of scale
on the spectral signatures of satellite data will help the correct interpretation of any
classification results.
Keywords: Land cover mapping; spatial resolution; Support Vector Machines
(SVMs); THEOS; Phuket.
40
1. Introduction
One of the fundamental characteristics of a remotely sensed image is its spatial resolution, or
the characteristic size on the ground associated with the radiance measurement of a pixel [1].
Unseemly choice of different spatial resolution can lead to misleading interpretation. Selection
of appropriate spatial resolution becomes more complex as resolutions increase. Since the basic
information contained in a remotely sensed image is strongly dependent on spatial resolution and
the spatial resolution of an image extensively affects every stage of image classification.
Markham and Townshend [2] found that image classification accuracy is affected by two factors.
The first factor is the influence of boundary pixels on classification results. The second factor
which influences classification accuracy is that finer spatial resolution increases the spectral-
radiometric variation of land cover types. The objective of this study is to determine spatial
resolution effects on land use/ land cover classification. THEOS with 15 m resolution and
Landsat-TM with 30 m resolution were compared. Each image is classified into 7 land use/land
cover types using support vector machines (SVMs) classifier. The classification accuracy at each
resolution is reported.
2. Study area and Methods
2.1. Study area
This research focused on the Kathu district, Phuket, Thailand (Figure 1). In this area, there are
many land use/land cover types such as urban areas, tourist areas, forest, agriculture, water
reservoir, golf courses, etc. The study area is approximately 31.8 km2
and located in the west of
Phuket Island. It neighbors Thalang to the north, Mueang Phuket to the east and south, and the
Andaman sea to the west. Kathu is the district which covers the most attractive beach of Phuket
including Patong and Kamala and the north is much less developed.
41
Figure 1. Study area (Kathu district, Phuket, Thailand)
2.2. Data sets
THEOS imagery with 15 m, and Landsat-5 TM with 30 m resolution were used to classify
land use/land cover. The datasets were acquired in January 2011 and December 2009,
respectively. The characteristics of both data sets are shown in Table 1 and Figure 2.
Table 1. Characteristics of the two satellites data sets used in this study
Imagery Source Resolution (m) Band Spectral Type
LANDSAT 5 TM
30
1 (Blue)
0.45 – 0.52 µm
30
2 (Green)
0.52 – 0.60 µm
30
3(Red)
0.63 – 0.69 µm
30
4 (NIR)
0.78 – 0.90 µm
30
5 (NIR)
1.55 – 1.75 µm
60
6 (TIR)
10.40 – 12.5 µm
30
7(MIR)
2.80 – 2.35 µm
THEOS
15
1 (Blue)
0.45 -0.52 µm
15 2 (Green)
0.53 – 0.60 µm
15
3 (Red)
0.62 – 0.69 µm
15
4 (NIR)
0.77 – 0.90 µm
Source: [3]
42
(a) (b)
Figure 2. Data sets of the study area in true color composite (a) THEOS (b) Landsat-5 TM
2.3 Classification methods
Support Vector Machines (SVMs) are normally a supervised classifier, which requires
training samples. Mountrakis et al. [4] reviewed that SVMs are not relatively sensitive to
training sample size and scientists have improved SVMs to successfully work with limited
quantity and quality of training samples. SVMs have been used in many Remote sensing-based
applications, for example, land use/ land cover, forest and agriculture tasks. SVMs classifier
turned out to be an effective method at handling not only the complex distributions of the
heterogeneous land cover classes that characterized the study area but also in various spatial
resolution scales [4].
Supervised training was adopted in this study. Groups of contiguous pixels were selected as
training samples in the class signatures as forest, built-up, road, water, agriculture, grassland and
bare land. For each class at each data set, the overall and individual Kappa coefficient are
calculated for each confusion matrix to evaluate the agreement between the classification results
and reference data [5]. Figure 3 shows the study workflow.
43
Figure 3. Study workflow
3. Results
After collecting area of interests (AOIs), THEOS and Landsat-5 TM images were classified
into the 7 classes using SVMs classifier.
The comparison of classification results from THEOS and Landsat-5 TM images are shown in
Figure 4 and the represented colors in Figure 4 are as follows; green as forest, magenta as built-
up, red as road, blue as water, maroon as agriculture, cyan as grassland, and yellow as bare
land.
44
(a) (b)
Figure 4. SVMs classification results of (a) THEOS (b) Landsat-5 TM images
The results of SVMs classifying land use/land cover in Kathu district which used THEOS
with 15 m and Landsat-5 TM with 30 m resolution were shown in Table 2. The accuracy and
overall Kappa coefficient for classification products from THEOS and Landsat-5 TM were
90.65% and 89.00%, respectively, with Kappa statistics of 0.88 and 0.87. User’s accuracy of
individual classes from THEOS and Landsat-5 TM ranged from 64.62% to 99.29% and 66.78%
to 100.00% in that order.
The overall accuracy and Kappa values were very similar for both data sets. However, the
overall accuracy and Kappa value from THEOS with 15 m resolution was slightly higher than
Landsat-5 TM with 30 m resolution. The classification accuracy in forest class from Landsat-5
TM was higher than THEOS data set.
45
Table 2. Accuracy assessment for reference classifications
Class
THEOS
Landsat-5 TM
Prod. Acc. (%)
User Acc. (%)
Prod. Acc. (%)
User Acc. (%)
Forest
97.47
96.81
100.00
100.00
Built-up
62.37
71.18
97.02
97.57
Road
74.89
64.62
90.15
90.59
Water
99.87
99.29
83.25
78.71
Agriculture
92.21
84.22
76.69
75.37
Grassland
89.49
95.23
96.02
91.85
Bare land
76.78
91.31
60.88
66.78
Overall Accuracy
90.65% (Kappa Co.= 0.88)
89.00% (Kappa Co.=0.87)
The Landsat-5 TM bands with 30 m resolution were not appropriate for clearly selecting
training data set than THEOS with 15 m resolution. However, the two near-infrared bands (band
4 and 5) are helpful to extract man-made features (i.e. Road and Built up areas) than THEOS
which has only one near infrared band.
Table 3. Accuracy assessment for reference classifications
Class
THEOS
Landsat-5 TM
Area (km
2
Area (%)
)
Area (km
2
Area (%)
)
Forest
64.41
46.85
62.80
45.45
Built-up
3.83
2.79
6.00
4.34
Road
13.86
10.08
15.47
11.19
Water
29.10
21.17
28.43
20.58
Agriculture
20.23
14.71
14.48
10.48
Grassland
2.90
2.11
6.28
4.55
Bare land
3.13
2.28
4.71
3.41
Total
137.46
100
138.17
100
4. Conclusions
In this study, land use/land cover was identified into 7 groups as forest, built-up, road, water,
agriculture, grassland and bare land. The resolution effects and classification accuracy were
reported using multispectral images data. The overall and individual Kappa values obtained at
different resolutions were also discussed.
46
Some of the major findings from this study are as follows.
1. The finer spatial resolution has a significant influence on the classification results.
Therefore, it plays a critical role in the quality of the variability of land use/land cover
classification.
2. More spectral bands should be concerned as it also will be very helpful for identifying
land use/land cover.
3. The higher resolution image greatly reduces the mixed-pixel problem, and there is the
potential to extract much more detailed information on land use/land cover structures.
Woodcock and Strahler [1] mentioned that observed classification accuracies were the result
of a tradeoff of two factors. The first factor is the influence of boundary pixels on classification
results. As spatial resolution becomes finer, the proportion of pixels falling on the boundary of
objects in the scene will decrease. The second factor is the increased spectral variance of land
cover types associated with finer spatial resolution.
Acknowledgements
The authors wish to thank Geo-informatics and Space Technology Development Agency;
GISTDA (Public Organization) and ASEA UNINet for funding support. As well, support was
provided by Faculty of Technology and Environment, Prince of Songkla University, Phuket
Campus.
References
1. Woodcock, C.E. and A.H. Strahler, The factor of scale in remote sensing. Remote
Sensing of Environment, 1987. 21(3): p. 311-332.
2. Markham, B.L. and J.R.G. Townshend. Land cover classification accuracy as a function
of sensor spatial resolution. in XV International Symposium on Remote Sensing of
Environment, Ann Arbor. 1981.
3. GISTDA. THEOS Characteristics. 2011 [cited 2012; Available
from: http://www.gistda.or.th/].
47
4. Mountrakis, G., J. Im, and C. Ogole, Support vector machines in remote sensing: A
review. ISPRS Journal of Photogrammetry and Remote Sensing, 2011. 66(3): p. 247-259.
5. Chen, D., D.A. Stow, and P. Gong, Examining the effect of spatial resolution and texture
window size on classification accuracy: an urban environment case. International Journal
of Remote Sensing, 2004. 25(11): p. 2177-2192.
© 2012 by the authors; licensee Asia Pacific Advanced Network. This article is an open-access
article distributed under the terms and conditions of the Creative Commons Attribution license
(http://creativecommons.org/licenses/by/3.0/).
