Professional Documents
Culture Documents
15arspc Submission 20
15arspc Submission 20
1
Geodesy and Earth Observing Systems Group (GEOS)
School of Surveying and Spatial Information Systems (SSIS)
University of New South Wales (UNSW), Sydney, NSW 2052, Australia
Phone: +61 2 9385 4201
Fax: +61 2 9313 7493
Email: anngoodluck2008@gmail.com; bl.ge@unsw.edu.au; cxj.li@unsw.edu.au
a
Abstract
The 2009 Victorian bushfires, also called the Black Saturday bushfires, ignited
across the Australian state of Victoria on Saturday 7th February 2009, resulting
in Australia’s highest ever loss of life from a bushfire. According to the Victorian
Police, the bushfires caused at least 173 known deaths of people and 414
people injured. The use of multispectral Landsat Thematic Mapper (TM) data
was able to detect and analyse the land cover changes caused by the bushfires
in a rapid and cost-effective way. Many digital change detection techniques,
such as image differencing and post-classification have been widely used for
this purpose. The objective of this research was to quantify land cover changes
by using two Landsat TM data, one pre-fire and one post-fire, acquired on 1st
March 2008 and 21st April 2009, respectively, and to examine the strength and
weakness of different methods. Change detection results derived from each
method were assessed for accuracy against ground survey.
Keywords: Bushfires, land cover change, Landsat TM, Image differencing,
Post-classification comparison
1. Introduction
Australia was frequently ravaged by bushfires during the warmer months of the
year. This is because that the warm and dry conditions intensify the probability
of fire. Especially, the southeast state of Victoria was fire prone owing to the
dense population, rugged terrain and well suited fire conditions in comparison of
other states (Beringer, 2000), as evident with the most notorious bushfire of
2009 Victorian bushfires. According to the Victoria Police (2009), the bushfires
caused at least 173 known deaths of people and 414 people were listed as
injured at 14th February 2009. And many towns which located in the north-east
of the state capital Melbourne, such as Kinglake and Marysville, were badly
damaged or almost completely destroyed (ABC News, 2009). These enormous
1
figures led to a heated discussion about how to alleviate the devastating
consequences of the bushfires.
Due to the hazardous nature of bushfires, fire fighters ideally require enough
information about bushfires to ensure to control them in time and minimise the
potential risk to communities and properties. However in bushfire cases, this
task has proven to be difficult due to the factors such as lack of accessibility and
large scales assessment (Graml & Wigley, 2008). For instance, traditional field
survey techniques such as using total station and GPS were accurate, it seems
to be high risk for field survey in fire case. Furthermore, it needed to use more
working hours to collect the data, so traditional regional mapping was basically
time-consuming to be applied for the large size assessment of bushfires
(Paradzayi et. al, 2008). But the online inventory was demanded for the maps of
damaged area. Therefore, satellite remote sensing was introduced to overcome
the problems above.
In addition, the Landsat TM program for Earth observation has provided the
valuable information about the Earth’s surface characteristics over the past
three decades. Before the 2009 Victorian bushfires, a few studies using Landsat
TM data have been already reported to efficiently evaluate the devastated
destruction in the bushfire case. Tupper (2000) analysed that two consecutive
overpass Landsat TM imagery can be used to quantify agricultural loss in 2007
Southern NSW bushfires. Turner et. al (1994) and White et. al (1996) also used
the Landsat TM imagery to assess the burned patterns in 1988 Yellowstone
National Park and Glacier National Park, respectively. Thus, the 2009 Victorian
bushfires detection will be examined by means of multispectral Landsat TM
data.
2
Figure 1 Selected bushfire detection area and the subset of Landsat TM false color
image.
Table 1 Scene information of Landsat TM images acquired on 1st March 2008 and 21st
April 2009.
Acquisition Path/Row Sun’s Datum Map Projection
Date elevation/Sun’s
azimuth angles
3
3. Methods
Figure 2 illustrates the detail analysis of bushfire detection methods used in this
study, and describes in the following sections.
4
3.2 Change detection approaches
Two change detection techniques including image differencing and post-
classification were used in this research to determine the changes have
occurred by the 2009 Victorian bushfires.
3.2.1 Image differencing
Image differencing is a technique that refers to subtracting a pixel’s Digital
Number (DN) on one date from corresponding pixel’s DN on the second date,
and produce a residual image which represents the change between two dates
(Mas, 1999). The subtraction results in positive and negative values reflect the
areas of change and zero value reflects no change (Sohl, 1999). This method is
the most widely used digital change detection algorithms, because it is
straightforwardness and easily to implement (Sunar, 1998). But this approach
cannot provide a detailed change matrix, and requires selection of thresholds.
In addition, bushfires may make the changes of land surfaces by reducing the
greenness of the vegetation and altering both aboveground and belowground
moisture and exposing soil. Patterson and Yool (1998) claimed that the
vegetation removal, soil exposure and the moisture content change of the
Earth’s surface can be detected by different remotely sensed indices. Thus,
remotely sensed indices such as Normalise Burn Ratio (NBR), NDVI
(Normalised Difference Vegetation Index) and MNDWI (Modified Normalised
Difference Water Index) were integrated in the image differencing approach for
discriminating between the bushfire areas and non-bushfire areas in this study.
5
multivariate statistics, respectively, in the remote sensing field (Congalton,
2004).
3.3.1 Error Matrix
Error matrix also can be called as confusion matrix. It is a square array of
numbers defined in rows and columns. These rows and columns express the
number of sample units such as pixels, clusters of pixels or polygons, which are
assigned to a particular category relative to the actual category as indicated by
the reference data (Congalton, 2004). Generally, the columns represent the
reference data and the rows indicate the results generated from the different
classification approaches. In addition, several statistical accuracies can be
generated from this matrix such as user accuracy, producer accuracy, overall
accuracy and Kappa coefficient (Foody, 2002).
6
Figure 3 Remotely sensed index of: 1)1stMarch2008NBR; 2)21stApril 2009NBR;
3)1stMarch2008NDVI; 4)21stApril2009NDVI; 5)1stMarch2008MNDWI;
6)21stApril2009MNDWI.
7
Figure 4 Index differencing of: 1)NBR; 2)NDVI; 3)MNDWI.
8
Figure 5 Color composite of NBR differencing, NDVI differencing and MNDWI
differencing.
9
4.2 Post-classification comparison
Land cover classification of each Landsat TM image was implemented by using
Maximum Likelihood Classification, in which six classes were selected including
forest, grassland, bared land, water body, urban area and residential area (Fig.
7). Then, the land cover change caused by the 2009 Victorian bushfires was
derived by comparing two individual classified images pre- and post- bushfires
(Fig. 8). Since all of the vegetation samples were classified as the forest or
grassland before the 2009 Victorian bushfires. Therefore, if a pixel changes
from the forest or grassland to the bared land, it would be classified as the
burned area. On the other hand, it would be defined as the unburned areas if it
is unchanged.
Figure 7 Image classification result of: 1)1st March 2008 Maximum Likelihood
Classification; 2)21st April 2009 Maximum Likelihood Classification.
10
Figure 8 Result of post-classification comparison
The image differencing method integrated the remotely sensed indices including
NBR, NDVI and MNDWI offers the best bushfire detection result, which can be
found in the following Table 2. It generates the KHAT value about 84.44%,
indicating a strong agreement. Furthermore, it’s omission error, such as the
actual bushfire pixels are not detected, and commission error, such as the
change pixels but not belong to bushfires, are 14.07% and 13.38%, respectively
(Tab. 4). On the other hand, the approach of post-classification comparison
provides the lowest KHAT value (i.e. 59.84%), which represents moderate
agreement (Tab. 3). And the omission error and commission error are 15.96%
and 46.15%, respectively (Tab. 4).
11
Table 2 Error matrix for image differencing method derived from the NBR, NDVI and
MNDWI (Unit: pixel).
Burned areas Unburned Row total User’s
areas accuracy
Burned areas 1606236 248191 1854427 86.62%
Table 4 Bushfire detection performance acquired from error matrices using different
change detection techniques.
Change Detection Omission Error Commission KHAT(%)
Techniques (%) Error(%)
Image 14.07 13.38 84.44
differencing
Post- 15.96 46.15 59.84
classification
Comparison
12
4.4 Comparison of the results of two different change detection
techniques
Although each of these change detection techniques could individually
contribute to the bushfire detection, it can be found from this study that the
image differencing approach integrated the three remotely sensed indices
including NBR, NDVI and MNDWI, represents the better result than the post-
classification comparison technique. This is because that the accuracy of the
post-classification comparison highly relays on the accuracy of the initial
classification, which is especially controversial in that it always pose a serious
error to the final result of post-classification comparison (Teng et. al, 2008). And
the mis-classification errors in the original images always make the results,
which are obtained using post-classification comparison, are judged
unsatisfactory (Coppin et. al, 2004). For instance, most of bared lands were
misclassified as either residential areas or grasslands. The reason is that the
low spatial resolution of Landsat TM imagery results in low capability of spectral
separation between different classes.
5. Conclusions
The profound and serious consequences caused by the 2009 Victorian
bushfires result in the assessment of bushfires become critically significant. In
order to minimise the bushfires negative impacts on society, an efficient and
reliable bushfire detection system was proposed to assess the devastated
effects of the 2009 Victorian bushfires. It is possible to utilise the repetitive
capability of satellite remote sensing imagery to identify the location of change
to the Earth’s surface and integrate the different remotely sensed indices. The
results confirm that the procedure can offer essential spatial information for
bushfire assessment.
Acknowledgements
This research was strongly supported by the Geodesy and Earth Observing
Systems Group (GEOS) of the School of Surveying and Spatial Information
Systems (SSIS) in the University of New South Wales (UNSW), and thanks the
Victorian Country Fire Authority (CFA) for providing the ground truth data.
13
References
ABC News, 2009, Absolute devastation: Victoria gutted by deadly bushfires,
Australian Broadcasting Corporation, accessed on 03 December 2009,
<http://www.abc.net.au/news/stories/2009/02/08/2485299.htm>
Beringer. J., 2000, Community fire safety at the urban/ rural interface: The
bushfire risk, Fire Safety Journal, vol.35, pp.1-23.
Chander. G., Helder. D. L., Markham. B. L., Dewald. J.D., Kaita.E., Thome. K.
J., Micijevic. E. and Ruggles. T. A., 2004, Landsat-5 TM reflective-band
absolute radiometric calibration, IEEE Transactions on Geoscience and
Remote Sensing, vol. 42, no.12, pp.2747- 2760.
Cheng. K. S., Wei. C. and Chang. S. C., 2004, Locating landslides using multi-
temporal satellite images, Advances in Space Research, vol. 33, pp. 296-
301.
Congalton. R. G., 2004, Remote Sensing and GIS Accuracy Assessment, CRC
Press, United States of America.
Coppin. P., Jonckheere. I., Nackaerts. K. and Muys. B., 2004, Digital change
detection method in ecosystem monitoring: a review, International Journal
of Remote Sensing, vol. 25, no.9, pp.1565-1596.
Graml. R. and Wigley. G., 2008, Bushfire hotspot detection through uninhabited
aerial vehicles and reconfigurable computing, IEEE Aerospace Conference
Proceedings.
Landis. J. and Koch. G., 1977, The measurement of observer agreement for
categorical data, Biometrics, vol.33, pp.159-174.
Paradzayi. C., Annegarn. H. J., Matsika. R. and Erasmus. B., 2008, Field
surveys for biomass assessment in African Savanna woodlands,
International Geoscience and Remote Sensing Symposium, vol. 3, no.1, pp.
III632-III635.
14
Patterson. M. W. and Yool. S. R., 1998, Mapping fire-induced vegetation
mortality using Landsat thematic mapper data: a comparison of linear
transformation techniques, Remote Sensing of Environment, vol. 65,
pp.132-142.
Richards. J. A. and Jia. X. P., 1998, Remote Sensing Digital Image Analysis,
Springer-Verlag, Germany.
Singh. A., 1989, Review article: digital change detection techniques using
remotely sensed data, International Journal of Remote Sensing, vol. 10,
no.6, pp.989-1003.
Sohl. T., 1999, Change analysis in the United Arab emirates: an investigation of
techniques, Photogrammetric Engineering and Remote Sensing, no.65,
pp.475-484.
Sunar. F., 1998, An analysis of change in a multi-date data set: a case study in
the Ikitelli area, Istanbul, Turkey, International Journal of Remote Sensing,
vol.19, pp.225-235.
Suzanchi. K., Sahoo. R. N., Kalra. N. and Pandey. S., 2006, Land use/ land
cover change analysis with multi-temporal remote sensing data, Proc. of
SPIE, vol.6405.
Teng. S.P., Chen. Y. K., Cheng. K. S. and Lo. H. C., 2008, Hypothesis-test-
based landcover change detection using multi-temporal satellite images – a
comparative study, Advances in Space Research, vol.41, no. 11, pp.1744-
1754.
Tupper. G. J., Worsley. P. M., Bowler. J. K., Freckelton. D. A., McGown. I. J.,
Pradhan. U. C., Roger. R. E. and Worsley. M. A., 2000, The use of remote
sensing and GIS technologies by New South Wales agriculture for
emergency management, IGASS 2000. IEEE 2000 International
Geoscience and Remote Sensing Symposium, vol.4, pp. 1489-1491.
White. J. D., Ryan. K. C., Key. C. C. and Running. S. W., 1996, Remote
sensing of forest fire severity and vegetation recovery, International Journal
of Wildland Fire, vol.6, no.3, pp.125-136.
15