1.3 Trend Analysis
Trend analysis has been employed in this investigation as it has
proven useful in investigating vegetation cover dynamics in the
south west of Western Australia and in other parts of Australia
(Wallace and Thomas 1998, Curry et al 2008, Wallace et al
2006). The linear trend calculation requires a sequence of
rectified and calibrated imagery from which spectral indices
relating to variations in vegetation cover can be derived. Then
changes between image dates can be summarised by calculating
the linear regression (Furby ef al, 2008).
Previous trend analysis in the south west of Western Australia
on selected forest blocks containing mixed woodlands of
wandoo, marri, jarrah and mallet trees has been conducted
(Garkaklis and Behn, 2009). Long term trends in vegetation
cover using image dates between 1988 and 2005 were created
which were then validated in the field. Loss identified by the
trend analysis was confirmed in the field, but not all of the loss
was occurring in the wandoo trees but in co-occurring tree
species. The long term trends used in the Garkalis and Behn
study do not identify areas of recent recovery in canopy cover.
This study aims to interrogate the time series of imagery and use
shorter trend periods to inform the questions posed by the
Wandoo Recovery Group.
2. DATASETS
2.1 Landsat
The satellite imagery utilised is Landsat 5 Thematic Mapper
(TM) and Landsat 7 Enhanced Thematic Mapper Plus (ETM+)
supplied by the Land Monitor Project (Caccetta et al 2000). In
Western Australia the Land Monitor program used Landsat data
initially to monitor salt affected land and remnant vegetation,
and currently produces perennial woody vegetation monitoring
products (Land Monitor, 2012). The images are captured in the
dry season (summer) and are rectified and calibrated to base
images. This processing makes the imagery suitable for time
series and trend analysis. Image dates included in the analysis
are 1990, 1992, 1994, 1996, 1998, 2000 and annually 2002 -
2010.
2.2 Field Surveys
The Wandoo Recovery Group has developed a technique for
assessing wandoo crown decline (WCD). It categorises levels
of active crown foliage loss along a scale ranging from healthy
to dead (WRG 2005). The Wandoo Recovery Group has
monitored the crown condition at 24 sites between 2006 and
2010 and determined trends in the crown condition over this
time period. Surveys were of individual trees along transects on
average 100m long and 20m wide. Jack Mercer established
three long range transects with several sites along each and
made observations of wandoo crown condition and stand health
in 1991, 2002 and 2008 with the 2002 and 2008 field data being
the most comparable (Mercer 1991, 2003, 2008).
In 2010 (10) and 2011 (8) sites were assessed for crown decline
stage and canopy cover by the authors. These sites were chosen
to be appropriate for use with Landsat imagery and were
selected in areas of varying trends in vegetation cover. One
hectare homogeneous sites were identified and three to five trees
within a site assessed.
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B3, 2012
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia
2.3 Vector Data
The study area has been restricted to the mapped occurrence of
wandoo. The boundaries of wandoo occurrence were
interpreted from airphotos and ground knowledge (figure 1.
Wandoo does occur with other species (jarrah and marri) and
can be mistaken for another species, powderbark wandoo (E.
accedens). The mapped extent does not capture all wandoo
woodlands and does include woodlands that are not wandoo.
However restricting the analysis to this region enables broad
observations of the wandoo woodlands to be made.
3. PROCESSING
3.1 Index images
Input to the trend analysis requires spectral index images that
represent the land cover being investigated and vary linearly
with changes. Index images were created using a combination
of Band 3 and Band 5 of the Landsat imagery shown in equation
(1).
(Band 3 + Band 5) / 2 (1)
This band combination has been found in other studies in the
south west of Western Austalia to be strongly related to changes
in projected foliage cover (Zdunic and Behn 2009, Garkalis and
Behn 2009). The index was calculated for all available images
1990 to 2010. The index for the 2000 image was later removed
as this image has much ephemeral understorey growth not
usually present in dry season imagery. In the winter of 1999
and into the summer months the rainfall was higher than
average. Keeping the 2000 image in the analysis meant that it
dominated the canopy cover in the image sequence and caused
false gain and loss in the trend epoch calculations.
3.2 Trend calculation
The linear trends were calculated for several time periods
(epoch) using orthogonal polynomials (Draper and Smith 1980)
with software provided by CSIRO Mathematics, Informatics and
Statistics. The output trend image contains the slope value
scaled to fit 8-bit integer data range of 0-255. Firstly the trend
for time periods 1990-2002 and 2002-2010 were calculated.
These longer time periods enable examination of long term
trends and are not as affected by individual image date
conditions.
The long term linear trends were classified into five classes;
decline, small decline, stable, small gain and gain cover classes.
This enables areas to be calculated and the classes compared
between the two time periods. Class boundaries were modified
from previous products produced by the Land Monitor Project
(Furby et al 2008).
A series of short epoch trend images were calculated using four
image dates. In all eleven four date trend images were created
by moving through the image sequence one date at a time. This
sequence was produced as WCD can occur over several years
not like other impacts like fire, flood or drought which have a
strong impact and then recovery signal. The decline does not
occur at the same time in all locations, but is diverse in its effect
in time and space. By reducing the trend epoch length and
calculating it for every available date the opportunity for
ident
incre:
4.1
Thes
wanc
year
and
high
cove
areas
2008
% Area of Wandoo Occurrence
> d
> uu Dg 2 y wo 555
Lo
Fig]
are
Wa
The
hap|
bys
are.
OCCI
obs
200
199
still
dec
200
469
has
the
2%
pre
(WI
4.2
Rec
geo
The
ind
app
(20
