rilli
76,
00, H
International Journal of Applied Earth Observation and Geoinformation 12 (2010) 138–149
Vegetation mapping
us
d.
, to
y ch
cha
o map vegetation as an important indicator for environmental change. Three
Contents lists available at ScienceDirect
International Journal of Appl
Geoinfor
w.1. Introduction
1.1. Heard Island
Heard Island is a pristine sub-Antarctic island south of the
Antarctic Polar Frontal Zone in the Indian Ocean. This Australian
territory is a 2800 m high volcanic and glaciated island, and
because of its remoteness, human visits to the island are very
infrequent. Heard Island is unique in terms of its location, climatic
conditions, vegetation communities, geology, volcanic activity, and
glacial cover (Bergstrom and Chown, 1999; Bergstrom and Selkirk,
2000; Bergstrom et al., 2002; Scott and Bergstrom, 2006). Up-to-
date and accurate spatial information is of crucial importance for
sustainable management of the island. Because of the island’s
remoteness, satellite imagery provides advanced and cost-effec-
tive means to map its land cover and to quantify environmental
changes. This information is important for sustainable manage-
ment of this pristine island, to study the regional effects of climate
change, and to assess the effects of human impacts. The glaciers on
Heard Island have been receding since 1947 when glacial extent
was first estimated from aerial photographs (Thost and Truffer,
2008). This recession was most likely caused by a temperature rise
of þ0:9 deg C between 1947 and 2004. Glacial retreat has exposed
new land that has become available for colonisation of plant
species.
During previous expeditions to Heard Island in 1986/1987,
1987/1988, 2000/2001 and 2003/2004 terrestrial plant ecology has
been studied and vegetation maps have been produced. These
maps were produced manually, based on visual interpretation of
aerial photographs and satellite imagery, combined with GPS-
based field samples (Bergstrom and Selkirk, 2000; Bergstrom et al.,
2002; Scott, 1989; Australian Antarctic Division, 2009). Because of
the inaccessibility of Heard Island, field surveys are often
expensive and labour intensive, and expeditions can potentially
be intrusive. Satellite images have been successfully used in
vegetationmapping, monitoring, and ecological applications in the
past (Aplin, 2005; Coppin et al., 2004; Jensen, 2000; Xie et al.,
2008). Very high spatial resolution imagery (VHR) such as IKONOS
(1-4 m spatial resolution, 4 multispectral bands) provides a
valuable new source of information for remotely sensed vegetation
mapping. Given that the island is rarely visited, satellite image
classification could be a suitable technique to produce vegetation
maps regularly and accurately. Satellite sensors are also able to
capture imagery of large parts of the island or even the entire
island, so that complete vegetation maps can be produced. If we
can produce accurate vegetation maps from VHR satellite imagery,
Multispectral classification
Grey level co-occurrence matrix (GLCM)
Texture-based classification
Sub-Antarctic Heard Island
IKONOS imagery
classification techniques were compared: multispectral classification, texture based classification, and a
combination of both. Texture features were calculated using the Grey Level Co-occurrence Matrix
(GLCM). We investigated the effect of the texture window size on classification accuracy. The combined
approach produced a higher accuracy than using multispectral bands alone. It was also found that the
selection of GLCM texture features is critical. The highest accuracy (85%) was produced using all original
spectral bands and three uncorrelated texture features. Incorporating texture improved classification
accuracy by 6%.
 2010 Elsevier B.V. All rights reserved.
* Corresponding author.
E-mail address: Arko.Lucieer@utas.edu.au (A. Lucieer).
0303-2434/$ – see front matter  2010 Elsevier B.V. All rights reserved.
doi:10.1016/j.jag.2010.01.006Texture-based classification of sub-Anta
Heard Island
Humphrey Murray a, Arko Lucieer b,*, Raymond W
aDatalive Software, 6 George St, Launceston, Tasmania 7250, Australia
b School of Geography and Environmental Studies, University of Tasmania, Private Bag
c School of Computing and Information Systems, University of Tasmania, Private Bag 1
A R T I C L E I N F O
Article history:
Received 7 December 2009
Received in revised form 5 January 2010
Accepted 12 January 2010
Keywords:
A B S T R A C T
This study was the first to
sub-Antarctic Heard Islan
multispectral information
Island’s pristine and rapidl
regional effects of climate
and non-invasive means t
journal homepage: wwctic vegetation communities on
ams c,*
Hobart, Tasmania 7001, Australia
obart, Tasmania 7001, Australia
e high-resolution IKONOS imagery to classify vegetation communities on
We focused on the use of texture measures, in addition to standard
improve the classification of sub-Antarctic vegetation communities. Heard
anging environmentmakes it a relevant and exciting location to study the
nge. This study uses IKONOS imagery to provide automated, up-to-date,
ied Earth Observation and
mation
elsev ier .com/ locate / jag

texture analysis; (b) to determinewhich texture features should be
used for classification; (c) to perform a classification using texture
alone; (d) to combine multispectral and texture features in a
classification; and (e) to compare this approach to standard
multispectral classification.
1.4. Paper Structure
Section 2 of this paper describes the study area and the imagery
being used for this research. Section 3 presents the texture-based
methods that were used in the study and explains the GLCMwhich
was a fundamental component in all of the techniques described.
Section 4 discusses the feature reduction and classification
techniques used, as well as the techniques for validation. Section
5 presents and discusses the results obtained while Section 6
presents the conclusions that can be drawn from the methods and
results.
2. Study area, imagery and field data
H. Murray et al. / International Journal of Applied Earth Observation and Geoinformation 12 (2010) 138–149 139we can potentially map and quantify changes in vegetation cover,
which for a pristine area like Heard Island provides an important
indication of the regional effects of climate change. This study is
the first to use satellite imagery for semi-automated vegetation
classification on Heard Island.
1.2. Texture-based classification
The multispectral bands of satellite imagery are often
transformed into thematic classes using an appropriate classifica-
tion technique (Lu andWeng, 2007; Tso andMather, 2001).Most of
these techniques, however, only look at the spectral values of
individual pixels and do not take into consideration the spatial
context of pixels. With recent VHR imagery, real world objects or
regions that were previously represented by only one or two pixels
now consist of many pixels. Therefore, techniques that take into
account the spatial properties of an image region need to be
developed and applied. One approach for including the spatial
relationship of pixels is modelling texture. Texture can be defined
as the various measures of smoothness, coarseness, and regularity
of an image region (Gonzalez and Woods, 1992). Previous studies
have shown that combining both multispectral and texture data
together can lead to improved classification accuracy (Ruiz et al.,
2004; Zhu and Yang, 1998). The popular grey-level co-occurrence
matrix (GLCM) texturemodel (Haralick et al., 1973; Haralick, 1979)
has been widely used in remote sensing studies (Clausi, 2002;
Franklin et al., 2001; Ouma et al., 2008). Recently, texture-based
classification algorithms have been successfully applied to VHR
satellite imagery (Aguera et al., 2008; Ouma et al., 2008; Puissant
et al., 2005). Tsai et al. (2005) and Tsai and Chou (2006) applied the
GLCM toVHRQuickbird imagery to detect invasive plant species. In
this study, we apply GLCM texture-based classification to VHR
IKONOS imagery of Heard Island.
In addition, the issues of scale and complexity in texture
definitions have been raised in previous remote sensing studies.
The size of the texture window should ideally match the spatial
scale of the object or class under consideration, but this is not
always a trivial exercise. The window should be large enough to
capture the relevant patterns, but if the window becomes too large
edge effects could dominate the results (Puissant et al., 2005).
Several studies have looked at the influence of the window size on
classification accuracy (Aplin, 2006; Chen et al., 2004; Franklin
et al., 1996). The GLCM texture model generates a range of
correlated texture features that can be used in a classification
(Haralick et al., 1973; Hall-Beyer, 2007). In this study, we
systematically examine the effect of the GLCM window size and
texture feature selection on classification accuracy.
In addition to spatial scale, thematic scale or complexity is
another issue in classifying natural ecosystems (Aplin, 2006). In the
last decade or so, a range of studies have explored the use of
hierarchy theory in image classification (Akc.ay and Aksoy, 2008;
Blaschke and Strobl, 2001; Benz et al., 2004; Burnett and Blaschke,
2003; Franklin et al., 2001; Ju et al., 2005; Wu, 1999). A
classification hierarchy arranges thematic classes into a hierarchi-
cal tree with the most generic classes at the top and the more
detailed classes further down, inheriting characteristics from their
parent classes. In this study we apply a hierarchical classification
approach toward classifying Heard Island’s vegetation communi-
ties.
1.3. Aim and objectives
In summary, the main aim of this study is to investigate
whether incorporating texture improves vegetation classification
based on VHR IKONOS imagery of Heard Island. The objectives of
the study are (a) to determine an appropriate window size for2.1. Study Area
Heard Island is a sub-Antarctic island located in the Indian
Ocean at approximately 53:11S; 73:54E. Fig. 1 shows how the
island is located with respect to Australia and Antarctica. The
Territory of Heard Island and McDonald Islands (HIMI) was
inscribed on the World Heritage List in 1997 for its outstanding
universal natural values. In addition to being recognised interna-
tionally for their conservation values, the Heard Island and
McDonald Islands and Marine Reserve are significant at the
Australian national level for their contribution to the National
Representative System of Marine Protected Areas (NRSMPA), their
heritage values and their important wetlands (Bergstrom and
Selkirk, 2000; Bergstrom et al., 2002; Australian Antarctic Division,
2009; Kiernan and McConnell, 1999; Scott, 1989).
Fig. 2 shows the topography of Heard Island. The island contains
a large volcano, Big Ben, which is covered in ice, snow, and glaciers.
However, the low lying coastal regions (approximately 20% of the
island) are covered bymany different vegetation communities. It is
the only large sub-Antarctic island free of introduced predators
and, because of its remote location, human visits are rare. Any
observed changes to the island are likely to have resulted from
causes that are not directly related to local human intervention.
Therefore by observing changes in the vegetation on Heard Island,
a valuable insight into the regional effects of climate change can be
obtained.
Fig. 1. Location of Heard Island (source: Australian Antarctic Data Centre, 2009).