HIGH RESOLUTION MONITORING OF WOODY VEGETATION

CHANGE OVER NEW SOUTH WALES

Tim Danaher1, Richard Hicks1, Tony Gill2, Geoff Horn1, Arndt Meier1, Stuart
Smith1, Andy Taylor1

1
Information Sciences Branch
Department of Environment, Climate Change and Water
PO Box 856, Alstonville, NSW, Australia, 2477
Tel: +61 2 6627 0224, Fax: +61 2 6628 3937
Email: tim.danaher@environment.nsw.gov.au
2
Joint Remote Sensing Research Program, University of Queensland, School
of Geography, Planning and Enviromental Management, St Lucia, QLD, 4072
Queensland, Australia

Abstract

Remote sensing is used by the Department of Environment, Climate change

and Water (DECCW) for mapping and monitoring the extent of woody
vegetation in the State of New South Wales (NSW). In the past DECCW has
relied solely on Landsat imagery to provide for these requirements but the need
to map woody vegetation change in landscapes such as open woodlands with
scattered trees, grasslands, and highly modified areas, has demanded the use
of higher resolution imagery. Since 2007, DECCW has been developing a
program for statewide monitoring of woody vegetation using SPOT 5 imagery.

The methods being used to map woody vegetation change have been adapted
from the SLATS Landsat approach. They includes radiometric correction of
SPOT 5 imagery, cross-calibration of SPOT and Landsat Foliage Projective
Cover (FPC) products, and development of a change index based on existing
training data from Landsat monitoring. Vegetation change for the 2008–2009
period is currently being mapped using SPOT 5 imagery and these methods.
The methods will be further refined by incorporating improved radiometric
correction techniques and additional validation data as they become available.
When a time series of SPOT 5 imagery is available higher resolution maps of
woody vegetation extent and FPC will also be produced.

Introduction
DECCW uses remote sensing for monitoring and reporting on changes to
vegetation extent across NSW and on-ground compliance activities related to
the NSW Native Vegetation Act. Since 2006, the reduction of woody vegetation
over NSW has been mapped using Landsat TM and ETM+ imagery. Woody

1
vegetation change in NSW has been mapped using methods developed for the
Queensland, Statewide Landcover and Trees Study1 (SLATS) and reports on
woody vegetation change covering time periods from 1988–2009 have been
published (DECCW 2009). However, the woody vegetation change mapped
with Landsat does not include all vegetation change. Ability to detect woody
vegetation change is influenced by both the 30 m resolution of the Landsat
imagery, and the pattern of vegetation on the ground. These factors combine
to reduce the ability to detect woody vegetation change in landscapes such as
open woodlands with scattered trees, grasslands, and highly modified areas.
In February 2007, the NSW Premier pledged $24m over 4 years for enhanced
vegetation change monitoring and compliance, using High Resolution Satellite
Imagery, to develop monitor and report annually on woody vegetation across
NSW. The Vegetation Change Monitoring Project was commenced in July 2007
following the 2007 election. This program aims to provide annual monitoring
based on SPOT 5 (10 and 2.5m) imagery for the whole state using summer
image epochs.
It is expected that the higher resolution SPOT 5 imagery will enable the
detection of vegetation change that cannot be mapped with Landsat TM
imagery. This includes areas of scattered trees and highly modified landscapes.
Since there were established methods to map woody extent and monitor
change in woody extent using Landat imagery, it was expected that these
existing methods would be adapted to enable analysis of SPOT imagery.
This paper describes the methods that are being used for the analysis of woody
vegetation change using SPOT 5 imagery and the results so far.

Data
The CRC for Spatial Information was commissioned to review the remote
sensing technologies suitable for mapping and monitoring the extent of woody
vegetation in NSW. In particular this included a review of the suitability of
remote sensing sensors with a spatial resolution of 10m or greater.
Subsequently, following a tender process, the supply of satellite imagery was
let to SPOT Image.
SPOT 5 imagery has been acquired for 2007/08, 2008/09 and 2009/10 epochs.
Imagery for 2004/05 was acquired through an earlier project and is also
available for use. A single statewide SPOT image coverage of NSW requires
the acquisition of approximately 340 SPOT scenes. Most of the images have
been acquired during the dry season, which typically coincides with summer for
most of the state. However, due to the time required for SPOT data acquisition
over NSW, cloud cover and off nadir look angle limitations, there are some
images that were not acquired at the optimum date. As a result some images
either have some cloud cover or were acquired when there was green
herbaceous cover, which make it difficult to separate woody from non-woody
vegetation.

1
http://www.derm.qld.gov.au/slats

2
The acquisition of this imagery involves the review of imagery acquired and
proposed by SPOT Image. The data is provided by SPOT Image in Dimap
format. These data are passed onto Geoimage who process the imagery to
create rectified images and mosaic products for internal distribution by the
DECCW image web server.

Methods
Pre-processing
The raw (Dimap) SPOT imagery and rectified products from Geoimage are
loaded on the DECCW Linux computing cluster. An open source computing
environment using Python2, GDAL3 and SciPy4 is used for image analysis.
Images are imported using a script that renames these files according to a
standard file naming convention. During the load, PostgreSQL database tables
containing scene footprints and processing metadata are populated.
The calibration provided with the SPOT imagery is applied to the data to
convert the data to radiance. Some preliminary validation using existing
pseudo-invariant calibration sites was done to confirm that the SPOT calibration
was accurate and being correctly applied to the imagery. While it showed a high
level of agreement with the published SPOT gain coefficients it was based on a
very small data sample and further validation is required.
Atmospheric correction was performed on the SPOT imagery to retrieve
estimates of surface reflectance. The radiative transfer modelling software, 6S
(Vermote et al., 1997), was used. Table 1 lists the parameters required by 6S
and the source of those parameters used for this study. The aerosol optical
depth is one of the most sensitive and also difficult parameters to obtain. After
testing many approaches (Gillingham, 2009) it was decided to run with a
constant AOD of 0.05 as in most cases the atmosphere in NSW is quite clear
and all techniques for estimation of AOD from the image data were problematic.
Table 1 lists the parameters required by 6S and the source of those parameters
used for this study.

Removal of cloud is an important step if automated processing is to be run on

the data. A cloud masking method has been developed and is being applied.
The method uses morphological image processing to identify cloud marker
pixels from which segments are grown using the watershed transformation
(Fisher, 2010). It can generate a semi automated cloud mask based on a single
image in which clouds are bright features. Further refinement of the method is
also being done through a Joint Remote Sensing Research Program 5(JRSRP)
project.

2
http://www.python.org/
3
http://www.gdal.org/
4
http://www.scipy.org/
5
http://www.gpem.uq.edu.au/jrsrp

3
Table 1: Parameters of the 6S correction and references
Parameter
Continental aerosol model
Ozone
Aersol Optical Depth (AOD) at
550 nm
Precipitable water
Solar zenith and azimuth angles
View zenith and azimuth
Source
6S (Vermote et al., 1997)
TOMS climatology 1996-2003
(Ziemke et al., 2006)
Fixed at 0.05
Daily interpolated vapour pressure
point observations (Jeffrey et al.,
2001)
6S (Vermote et al., 1997)
Calculated from the image data

SPOT 5 / Landsat FPC cross-calibration approach

The metric of vegetation cover used in many Australian vegetation classification
frameworks is Foliage Projective Cover (FPC). Regression approaches and
other statistical techniques (Armston et al., 2009) have been used to predict
FPC from Landsat imagery and these Landsat FPC data are then used for
mapping vegetation extent and monitoring woody vegetation change.
Operational mapping of overstorey FPC requires an efficient and automated
method due to the large volume of satellite imagery that require processing and
interpretation. This can only be achieved if the imagery has been
radiometrically corrected to remove systematic spatial and temporal
differences.
As the current radiometric correction for SPOT imagery is only an atmospheric
correction and doesn’t include correction for BRDF effects, significant scene to
scene differences exist in some cases. Hence, it was not possible to develop a
general FPC model that could be applied statewide to the SPOT imagery. To
overcome this problem a cross calibration approach combining atmospherically
corrected SPOT 5 imagery and the existing calibrated Landsat woody extent
and FPC product is being used to generate SPOT FPC products.
The steps in calculating the SPOT FPC using the Landsat multi-temporal FPC
for calibration are:
1. The satellite image data of which the FPC is to be calculated (e.g. SPOT
5) is brought to the same spatial resolution than the original FPC product
(25m).
2. A check for spatial variability is performed on a per pixel basis: pixels are
excluded if the spatial variability in a surrounding 7x7 pixel matrix is
extending a certain threshold in the coefficient of variation.
3. The remaining pixels are sub-sampled and data values are exported as
ASCII values of which a regression model predicts the FPC values using
all four input bands of the satellite image.
4. A colour table is applied and the data are exported as Erdas Imagine
.img files.

4
Where SPOT images in the same path were acquired on the same date, they
are combined for cross-calibration. This step helps increase the area of SPOT
imagery which also increases the range of FPC values used in the cross-
calibration of an image.
The cross calibration method was run to provide SPOT 5 resolution FPC data
for the vegetation extent and change mapping. It has been run on the 10m
resolution SPOT imagery and in future will be tested on the 2.5m pan-merged
products.

Vegetation change mapping

The method used to map change in woody vegetation using Landsat imagery
was developed in the SLATS program and is described by Scarth et al. (2008).
The difficulties in consistently and reliably detecting woody change across a
wide range of vegetation types over a large geographical area required the
development of a change method that combined both the spectral changes
from scene to scene and the historical variability in foliage cover. This
combination of change detection methodologies was based on combining a
modified form of image differencing with a time series model using the FPC
time series to detect true change. Calibration data, collected from an analysis of
previous operator-interpreted and field checked woody change data sets from
previous years, were used to develop the spectral and temporal change
indices. The final classifier sits within the SLATS operational processing
framework and is run automatically across the state. However, in order to use
this classifier in an operational reporting environment, three levels of
classification at the 2%, 5% and 15% omission level are produced which are
then further interpreted, edited and field checked by an operator.
The approach being used to map woody vegetation change using the SPOT 5
imagery is based on a similar change index that combines spectral and FPC
indices from the earlier and later images. The change index being used was
developed using Landsat data and existing vegetation change data for training.
To simulate SPOT data, only Landsat TM bands that are similar in wavelength
to the SPOT 5 bands e.g. TM 2,3,4,5 were used. When applying this index to
the SPOT 5 imagery there is an adjustment applied to correct for differences in
the radiometric correction applied to the SPOT 5 imagery and the empirical
correction (Danaher, 2002) applied to Landsat TM imagery.
As the radiometric correction applied to the SPOT imagery is only an interim
correction at this stage and does not remove BRDF effects, it was not possible
to develop a change mapping method that included automated thresholding of
the change index, as is done with Landsat. Thresholding of the woody change
index is done on a scene by scene basis or multi-scene block, where the image
dates are the same. Four possible change class thresholds are set to assist in
visual interpretation of the change images. This thresholding task is being done
by two experienced image analysts to ensure consistency in the approach.
Similar to the Landsat change method, visual editing is used to check the
output of the classifier, and further improve the accuracy of the final product.

5
Initially only the multi-spectral (10m) SPOT data is being used in the automated
analysis. When a suitable pan-merging method is tested and implemented,
processing based in the 2.5m data will commence. Polygons used in this
editing process (Erdas Imagine AOI’s) are being kept to enable the re-coding to
be repeated with 2.5m resolution change data when available.

Vegetation extent mapping

The major limitation of predictions made from any single date model of FPC is
that overstorey and understorey FPC are not decoupled. The use of overstorey
FPC products derived from TM or ETM+ data requires the assumption of a
senescent or absent herbaceous understorey at the time of image acquisition
(Armston et al., 2009). This has serious implications for trend analysis of woody
vegetation cover because photosynthetic herbaceous FPC will increase the
variance of the sparse Landsat time series.
A multi-temporal rule based approach has used to map the woody extent and
FPC over the States of Queensland and New South Wales (Kitchen et al.,
2010). The woody extent and FPC product uses statistics based on a time
series of images to distinguish woody from non-woody vegetation and calculate
the best possible prediction of FPC for each pixel within a scene, for a specified
time period e.g. 1988–2008. It is planned to apply a similar approach to SPOT
5 imagery once there is sufficient multi-temporal coverage of NSW.
To demonstrate this approach using SPOT 5 imagery, a small test area was
chosen in an image overlap area to provide a sequence of four different image
dates. The four SPOT FPC images were used to calculate time series mean,
minimum and standard deviation. These layers were then used to classify the
woody extent and mean FPC. The SPOT and Landsat products were visually
compared.

Results and Discussion

SPOT Cross-Calibrated FPC
The FPC cross calibration approach was applied to 627 SPOT images acquired
during the 2008–2009 period. For each scene the SPOT FPC was validated
against the Landsat FPC product using a minimum of 1100 points and an
average of 44000 per run. The average slope, intercept and RSME were 0.66,
6.7 and 4.5 respectively.
As the cross calibration approach predicts FPC from a single date SPOT image
using the multi-date Landsat FPC product for calibration, it has some
limitations. The single date SPOT images are likely to over-predict FPC where
there is green herbaceous cover present. This limitation was most obvious in
images where non-woody vegetation is the dominant cover. There was a spatial
trend evident in the validation statistics with the slope varying from 0.75 in the
eastern part of NSW to 0.4 in the western areas. These differences are most
likely due to the distribution of woody vegetation where the eastern areas are
dominated by woody vegetation and western areas contain predominately non-

6
woody vegetation. The cross calibration approach performs best for images
where woody vegetation cover dominates. Examples of the SPOT FPC images
are shown in Figure 1(b) and 1(e). The range of FPC is represented by shades
of green from dark to light corresponding to high to low FPC. The light brown
cover is zero FPC.

Vegetation change mapping

The method for mapping woody vegetation change has been applied to SPOT
5 imagery over NSW for the 2008–2009 period. This has involved the
processing of 325 pairs of images on the DECCW Linux cluster using open
source software. While the automated change analysis process is complete the
visual interpretation and recoding of change rasters is still in progress. When
this is completed for the 2008–2009 period the results will be combined to form
statewide data sets and vegetation change rates calculated, in a similar way to
the Landsat processing.
Figure 1 shows an example of the woody change products based on SPOT 5
and Landsat imagery. It is comprised of six images. The top row from left to
right are (a) the 2008 SPOT 5 image, (b) the cross-calibrated FPC and (c) the
2008–2009 SPOT to SPOT change raster. The second row comprises (d) the
2009 SPOT5 image, (e) the 2009 cross-calibrated FPC and (f) the Landsat
2008–2009 change raster.
The 2008–2009 SPOT to SPOT change raster presented in Figure 1(c) shows
the four probability bands from highest to lowest probability of change as red,
pink, light and dark grey respectively. The green areas indicate that these areas
have been interpreted as actual change and coded into a particular change
category. For the Landsat change raster presented in Figure 1(f) red, pink and
light grey show the highest to lowest probability of change respectively.
Figure 1 shows an area of clearing along a fence line in the upper third of the
image, as well as a few scattered trees in the lower right hand corner. Long
linear clearing patterns are typically difficult to accurately identify in mis-
registered imagery. This is due to the relative displacement of the woody areas,
resulting in a large commission component in the change raster. However with
precisely registered imagery from push-broom sensors such as SPOT 5, this
type of clearing is more readily identified. This can be seen in both Figures 1(c)
and 1(f) at the top of the images.
The precision of the image to image registration is important in per pixel change
detection. For the examples presented in Figure 1, the SPOT to SPOT
registration is more precise than the Landsat to Landsat registration yielding a
more readily identifiable change pattern. Comparing the two images on the right
hand side of the figure it is apparent that the SPOT5 change raster identifies
the area of change in a more precise manner than the Landsat change raster.
Further, the scattered tree clearing in the lower right of the image is not clearly
identified in the Landsat Change raster. This type of clearing can be difficult to
capture in Landsat Imagery due to the pixel resolution and greater registration
variance.

7
 (a) (b) (c)

(d) (e) (f)

Figure 1. Comparison of SPOT 5, FPC and change raster images. a) 2008 SPOT5
Image showing bands 4,3,1 as RGB. (b) 2008 cross-calibrated FPC image. (c) 2008–
2009 SPOT5 change raster image. (d) 2009 SPOT5 Image showing bands 4,3,1 as
RGB. (e) 2009 cross-calibrated FPC image. (f) 2008–2009 Landsat change raster
image.

8
Vegetation Extent Mapping
There has only been one small test area processed but it suggests that the time
series woody extent and FPC approach (Kitchen et al., 2010) used with
Landsat imagery will also provide a useful product when applied to the SPOT 5
imagery. Figure 2 shows the resultant SPOT product compared to the Landsat
product. The most recent date of time series imagery used in the SPOT and
Landsat examples is different so they are not directly comparable, however,
they can be compared in some parts of the imagery.
The most notable difference between the FPC products is the mapping of linear
features. The SPOT product is clearly better at mapping narrow areas of
riparian vegetation and non-woody areas associated with roads, power lines
and fences. This difference is most likely due to a combination of better image
resolution and more precise image to image registration.
The SPOT product still contains some areas of commission where green
pasture has been incorrectly classified as woody vegetation. This is due to the
relatively short time series of SPOT imagery (four images) compared to the
Landsat product (12 images). When developing an operational product this will
need to be addressed either by using a longer time series of SPOT imagery or
developing an image fusion approach, integrating lower resolution but higher
frequency information such as Landsat.

(a) (b) (c)

Figure 2. Comparison of woody extent and FPC products from multi-date Landat and
SPOT imagery. (a) 2009 SPOT 5 image showing bands 4,3,1 as RGB (b) woody
extent and FPC based on SPOT images 2007–2009 (c) woody extent and FPC based
on Landsat images 1988–2008.

9
Conclusions
The FPC cross-calibration from Landsat to SPOT has provided a product to
enable woody extent and change mapping using SPOT imagery, without the
need for additional site calibration data. Similar to Landsat, the FPC images
discriminate woody from non-woody vegetation best when imagery is acquired
in a dry season. Radiometric correction improvements, due later this year, will
provide better scene to scene matching, which will enable FPC cross-calibration
based on larger SPOT regions. The SPOT cross-calibrated FPC method will be
refined when the more complete radiometric correction of SPOT 5 imagery is
implemented.
The existing Landsat change mapping method was adapted for analysis of
SPOT imagery and is being used to map woody vegetation change for the
2008–2009 change period. It has enabled delivery of first products but can be
refined to improve efficiency. An improved change index will be developed for
the 2009–10 period using existing validated 2008–2009 change rasters to
optimise a change mapping index. It will be based on SPOT imagery with more
complete radiometric correction.
SPOT5 data and woody vegetation change method is being applied
operationally and detecting change not visible in Landsat imagery. Therefore
the SPOT woody vegetation monitoring is likely to provide different figures to
the Landsat based methods due to spatial and spectral differences. The
magnitude and reasons for these differences will be evaluated and explained
when reporting the results from the SPOT program.
When there are at least four statewide coverages of SPOT imagery available it
will be possible to develop a higher resolution SPOT woody extent and FPC
product based on the time series of imagery.

References
Armston, John D., Robert J. Denham, Tim J. Danaher, Peter F. Scarth, and
Trevor N. Moffiet. 2009. Prediction and validation of foliage projective cover
from Landsat-5 TM and Landsat-7 ETM+ imagery. Journal of Applied Remote
Sensing 3 (1):033540-28.
Danaher, T. 2002. An empirical BRDF correction for Landsat TM and ETM+ im
agery. In: Proceedings of the 11th Australasian Remote Sensing and Photogra
mmetry Conference, Brisbane, September 2 - 6, 2002.
Department of Environment Climate Change and Water, 2009. NSW Annual
Report on Native Vegetation 2008, DECC 2009/523 ISBN 978 1 74232 390 9.
14 pp.
Fisher, A. 2010. Joint Remote Sensing Research Program, University of
Queensland.
Gillingham, S. 2009. Department of Environment and Resource Management,
personal communication.

10
Jeffrey S. J., Carter, J. O., Moodie K. B. and Beswick A. R. 2001. Using spatial
interpolation to construct a comprehensive archive of Australian climate data.
Environmental Modelling & Software, 16:309-330.
Kitchen J. , Clarke A., Danaher T., Armston J. and Scarth P. 2010. Operational
Use of Annual Landsat-5 TM and Landsat -7 ETM+ Image Time-Series for
Mapping Wooded Extent and Foliage Projective cover in North-Eastern
Australia, In: Proceedings of the 15th Australasian Remote Sensing
Conference, Alice Springs, September 2010.
Scarth, P, S Gillingham, and J Muir. 2008. Assimilation of spectral information
and temporal history into a statewide woody cover change classification Paper
read at 14th Australian Remote Sensing and Photogrammetry Conference, at
Darwin.
Vermote E. F., Tanre D., Deuze J., Herman M. and Morcrette J. 1997. Second
simulation of the satellite signal in the solar spectrum, 6S: an overview. IEEE
Transactions on Geoscience and Remote Sensing, 35:675-686.
Ziemke J. R., Chandra S., Duncan B. N., Froidevaux L., Bhartia P. K., Levelt P.
F. and Waters J. W. 2006. Tropospheric ozone determined from Aura OMI and
MLS: Evaluation of measurements and comparison with the Global Modeling
Initiative's Chemical Transport Model. Journal of Geophysical Research, 111:
D19303.

11