After some time the data were resupplied, however was found to be in the same condition. Staff of SpecTerra Systems revealed that a uniform correction was applied to all frames in each run. This had the effect of correcting bad frames and turning good frames bad. It should be self evident that aircraft roll will be different for each frame and that corrections should be applied to individual frames and not entire flight runs. A representative of SpecTerra Systems was invited to view the images on the processing workstation at Murdoch University and the problem again detailed with the advantage of visual representation of the perceived problem. The data were subsequently returned for further correction.
The data were supplied again with most images corrected however some frames still had offsets of 3-4 pixels or some 6m on the ground. Some portions of the data were not supplied and one frame was corrupted. Deadlines for mosaic processing were running out and the work proceeded without further requests for correction of the data.
The mosaic image was used to compile two posters for display purposes. The first poster shows the mosaic file as a subset scene of the vegetation showing the classification classes and the second as a false colour infra-red picture. The final mosaic is shown in Figure 4 and comprises approximately 400 sub images. The ground resolution of the final mosaic is 3m and the image file size is 520Mbytes. The final vegetation classification image is shown in Figure 5. Due to the high level of detail shown in both these images, full size reproductions are attached (Attachment 1 and Attachment 2).
It was evident from the mosaic that the flight of the aircraft had not been in accordance with specifications for overlap and coverage. In particular, the western edge of the Harvey Estuary was initially not supplied and the eastern flight path of the Harvey Estuary veered away from the overlapping flight path leaving no overlap in the data. Portions of the Peel Estuary suffered from lack of overlap and the portions of the flight paths in the north western corner of the Peel Estuary were either terminated too early to provide sufficient coverage or not supplied. Portions of the Creery Wetlands were missed due to lack of overlap. The Serpentine River above the mid portion of Lake Goegrup to Amarillo was initially not supplied. Small portions of data were also missing, due to lack of overlap, in the area of the Pinjarra Golf Club, Roberts Bay and Austin Bay (Figure 4).
Because the data were acquired over various days, portions of the study area were treated separately in the classification processing. Areas of interest were centered on those areas collocated with transects designed to be revisited for detection of vegetation change over time. The interest areas are located in the south-eastern edge of the Harvey Estuary, the Austin Bay areas, The Roberts Bay area, Creery Island, and Goegrup Lakes area. A map showing transect localities is shown in Figure 1. Each area of interest will be dealt with individually in this section.
Proposed future monitoring areas were based on the same locations chosen for
a previous study on samphire saltmarshes of the Peel-Harvey (Glasson
et al, 1995). Each area was extracted from the classified image
and the number of pixels of each class in the scene determined. The area
of each class was then determined based on a pixel
of 9 square meters. Table 6 shows the UTM
boundary coordinates for the proposed monitoring areas. Table
7 gives the area of classes by sub image. Table 8
gives the area of class by proposed monitoring area. The subset images
and the resampled classifications of the proposed monitoring areas are shown
in Figure 6, Figure 7, Figure
8, Figure 9 and Figure
10.
The classification accuracy is based on visual allocation of pixels to their classified class as against their machine classified class. The pixels were selected on a random stratified basis, that is, the number of pixels from each class selected was based on the overall number of pixels allocated to that class in the image. Where classes were made up of subclasses the selection was limited to a single subclass having the majority of pixels. This was done to avoid the interpreter having to decide whether a pixel belonged to an individual class where there was visually no definitive clue, for example Farmland1 or Farmland2. Fourteen vegetation classes were selected as representative of the classified image for accuracy assessment. Initially 250 pixels were to be selected which were confined to a near neighbor of 9 majority pixels. This means that the pixel selected had to be surrounded by pixels of the same class. This was to avoid the interpreter having to classify visually edge or mixed pixels. The machine selection of the pixels was exceedingly slow and after 5 hours of processing eventually 110 pixels were selected. These pixels were used for the final classification accuracy assessment. An overall classification accuracy of 76% was achieved.