Geoscience Reference
In-Depth Information
The BRS obtained 156 change images from state agencies and calculated the total change
(woody vegetation) for each image. Due to resource limitations, it was decided that only half the
scenes in the study area would be assessed. Scenes containing only a small percentage of landmass
were excluded, leaving 151 for possible assessment. Within each state, scenes were classified as
having “low” or “high” levels of change. A weighting program that took into account the amount
of change in each scene, the individual states' overall contribution to Australia's LC change, and
the methods used for change detection was used to select scenes for sampling. A total of 67 scenes
were selected for reliability assessment (Figure 8.3).
Analysts who had not previously been involved in the project were selected through a compet-
itive tendering process to prepare the independent sample-based estimates of change for comparison
with the results produced by state agencies. The analysts were provided with coregistered TM
images for the 67 scenes, pixel coordinates for the upper left-hand corners for sample units within
scenes, and 1990/1991 LC maps showing the distribution of woody vegetation, but no LC change
information. They verified the image coregistration, and after some preliminary testing they calcu-
lated the normalized vegetation index (NDVI) for each image then subtracted the NDVI images
and displayed the difference image (Jensen, 1996).
The NDVI difference image for each sample unit was examined at a range of threshold values
to determine the location of positive (increases in woody vegetation) and negative (decreases)
differences. These areas were checked in detail against the 1990/1991 and 1995 images and the
upper and lower thresholds for increases and decreases were recorded. A final classification of
change was performed using selected thresholds, and the change areas were checked against the
LC image. This ensured that only areas that were woody in 1990/1991 could be identified as a
“decrease” and those that were not woody in
1990/1991 as an “increase” in woody cover. The
analysis provided an estimate of the number of pixels of increase and decrease for each sample
unit in the image.
For quality assurance (QA) purposes, four to six sample units from half the images being
assessed were randomly sampled. Change for these sample units was assessed as described above,
but by different operators. Differences in interpretation were discussed and evaluated statistically
using a paired
-test. A sample unit failed QA if the average of the differences found by two
operators was not the same. In this case the main operator reexamined all the sample units for the
image. The analysts provided BRS with a spreadsheet for each image containing the sample
locations and the number of pixels of increase and decrease for each sample in the image, plus
notes on any other areas of possible change identified. The BRS implemented an automatic analysis
to evaluate the differences between the state estimate of change and those provided by the consult-
ants. For scenes where the state's and the analyst's estimates of change differed substantially, the
BRS investigated the possible reasons. The approximate spatial distribution of the change in the
state change map was also examined to determine whether a lack of acceptance was due to highly
localized changes difficult to sample effectively using the current method.
The investigation of lack of acceptance included inspecting the state's sources of information
used for initial checking of the change (i.e., aerial photography, other satellite imagery, or field
data) (Kitchin and Barson, 1998). Discrepancies were forwarded to respective states for advice on
likely reasons for such differences. If no reason could be identified (e.g., where severe drought had
led to leaf drop so that spectrally the area appeared to have been cleared, but ground inspection
showed that it had not), the scene was reprocessed by the analysts.
t
8.3 RESULTS
In the first assessment, 60 of 67 scenes met the acceptance criteria described above. The seven
noncompliant scenes were forwarded to the states for comment, and a new set of 500-
500-pixel
sample units was generated for reprocessing these scenes. On reprocessing, five additional scenes
¥
