Materials and Methods

The study area was determined by the tidal range area of the Peel-Harvey Estuary and Harvey, Serpentine and Murray rivers (Figure 1). The Peel-Harvey system is 75km south of Perth. It consists of the Peel Inlet and Harvey Estuary which together form a large shallow waterbody of some 133km2. Three rivers, the Serpentine, Murray and Harvey flow into this system. The whole study area lies on the Swan Coastal Plain of Western Australia. Lines of coastal sand dunes consisting of Safety Bay sand and Tamala Limestone (Dept. of Mines, 1977) separate the system from the Indian Ocean. Goemorphology and vegetation of the system have been described elsewhere (for example Semeniuk, 1990).

Data processing was performed on a Sun SparkStation IPX with 24bit RasterOps video display. The workstation has 40Mb of RAM and a total of 5Gb of hard disk storage over 3 separate hard disks. Data storage backup is made to a DAT tape with a storage capacity of 3Gb. A raster image of the entire study area at the supplied image resolution would result in a digital photomosaic of approximately 4Gb, exceeding the system storage capacity. For this reason the images were resampled to 3m resolution and mosaiced in six sub images based on the flying dates and continuity. The entire scene mosaic then amounted to 510Mb. A detailed technical description of the digital mosaic process is contained in McComb et al (1995). This study differs from McComb’s study in that digital scans of air photographs were used previously whilst this study used digital video. The image processing was completed using ERDAS Imagine 8.1.0 software.

DMSV Instrument

The DMSV images in native Targa format were supplied by SpecTerra Systems on DAT tape. This is possibly one of the only suitable storage media for provision of data of this type given the large number of individual files where the total volume exceeds that of a CD-ROM (>600Mb). In a DMSV image file there are four spectral bands, that is each picture element (pixel) has four values associated with the amount of radiance that each CCD camera measures. There are four cameras, each measures a discrete portion (band) of the spectral frequency. The bands and the wavelengths and optical features of the cameras are shown in Table 2. A digital file that arranges pixel values of a image in a grid is known as a raster file. The size of a raster file is a function of the number of rows, columns and bands that comprise the file, for instance a image that has four bands and is 740 pixel wide by 578 pixel high will comprise of 740 x 578 x 4 = 1710880 (1.7Mb) discrete data values. Computer image processing software such as Imagine 8.1.0 will frequently add small amounts of meta-data to the image file for housekeeping purposes (file headers/trailers), marginally increasing file sizes.

Image Processing

The DMSV images were digitally joined together by georeferencing each image with minimal distortion and orientating the images to conform with mapping conventions. McComb et al (1995) used a digital orthophoto to georeference scanned air photos. The initial georeferencing of the video frames for this study was carried out using the 1994 digital photomosaic compiled by McComb et al (1995). Limitations of the Imagine 8.1.0 software restrict mosaicing images to a software array of 10,000 x 10,000 pixels and another apparent limit is on the number of files in a mosaic task. The final mosaic was composed of initially four and ultimately six smaller mosaics. These were logically divided on the basis of flight runs conducted at similar times. The Harvey West and Goegrup images were compiled from additional data not supplied at the original time, indeed their classification processing was started after the initial four images were completed. The runs, timings and dates acquired are shown in Table 3.

All flights were between 7:30am and 10:00am. The timing of the missions was designed to obtain maximum water penetration and to avoid strong seabreezes (waves) around noon. Prior to mosaicing the images, they were histogram matched to enable correct colour balance between the sub components. The final mosaic image was compiled from approximately 450 sub-images. A flow diagram of the entire mosaic process is summarized in Appendix B. A table showing file handling procedures is contained in Appendix C.

As part of preliminary data interpretation the entire mosaic was subject to an unsupervised classification algorithm (ISODATA) to determine the extent of landcover classes, which could be discerned by parametric methods. The ISODATA classification resulted in 20 classes and took 25 hours of processing time. Based also on the outcome of the ISODATA classification and a trial vegetation index of the Harvey image it was thought the use of vegetation indices would not be productive.

Following evaluation of the results of the unsupervised ISODATA classification we used supervised classification techniques (maximum likelihood) to group the vegetation classes. The classification used classes based on vegetation association thought to be sufficiently spectrally distinct to eliminate misclassification, and informative in the determination of vegetation affected by variations in salinity. A table of the vegetation classes used is contained in Table 4. This method will be described in the following section. The classification process is summarised in Appendix D.

Method for Determining Classification Seed Areas

In order to carry out a supervised classification on the mosaic image, it was necessary to firstly define for information content, the vegetation classes present in the image area, and secondly to outline examples of these areas in the mosaic. An initial list of vegetation classes was compiled for the study area from reviewing maps, management plans and journal articles (Dames and Moore, 1987; DCE, 1978; EPA, 1993; Semeniuk and Semeniuk, 1990; Siemon et al, 1993; Woodcock, 1992; Kinnaird et al, 1979). The vegetation classes were morphologically based. A list of 18 classes was compiled, which could be reduced to a minimum of 10. This initial list relied on the accuracy scale and coverage of the sources. These three factors made this list provisional. The vegetation class descriptions were subsequently revised as the classification process proceeded.

The classes were refined and more added once ground truthing data had been included. A field trip was conducted in March 1995 to determine vegetation associations within the study area to use as a basis of seed sites for image interpretation and classification. Ground truthing involved visiting sites and recording (photographing) the broad vegetation characteristics of these sites. These characteristics included major species and the vegetation structure. An example would be Melaleuca and Acacia low closed forest. Appendix E shows some of the field sites visited for ground truthing. This information, and previous personal observations, when included in the compilation of the vegetation classes, resulted in a list of 23 classes. These can be seen in Table 4.

The next step was to create a digital overlay of the mosaic which contained examples of these classes. The initial mosaic was comprised of the original four subimages. The additional subimages were added after the original mosaic classification was completed. Firstly, areas where each class occurred were drawn on mylar sheets over colour air photographs. These were then transferred to an class seed polygon overlay on the digital mosaic. Each polygon drawn belonged to one of the 23 classes and was labeled as such. Multiple examples of each of the vegetation classes were included to aid the classification processes. In total there was 121 class seed polygons distributed throughout the original study area. In addition to the vegetation classes polygons were created for non vegetation classes such as water, farmland, bare ground, roads, and urban areas. These non vegetation classification classes were then used to eliminate these classes from the final vegetation classification result.

The next step was to convert class seed polygons into spectral signatures to be used as seeds for the supervised classification. Two classifications were conducted of the entire original mosaic image using these signatures. The first involved merging the signatures of the same class, resulting in 23 vegetation signatures for the classification. The second involved conducting the classification without merging the signatures of each class.

Both of these initial classifications were deemed to be unsatisfactory. A trial classification of the Harvey image using a maximum likelihood classifier was assessed to determine the accuracy of classification based on the vegetation and land cover classes It was evident that the water spectral component was determinant in class separation so it became necessary to eliminate the water bodies from the image to correctly separate vegetation classes.

The mosaic images were then further subset into two areas of interest, water body, and vegetation bounding the water body. These areas of interest were then subsequently classified again using a maximum likelihood classifier using only the relevant classes. To improve the classification it was found necessary to classify the initial four, then six subimages separately. Using the existing class seed polygons and taking into account the subimage area boundaries, spectral signatures were created for each subimage.

Subsequently, the 121 class seed polygons were split up between the initial four subimages depending on which subimage they were located on. Some deletions and modifications of the class seed polygons was carried out in an attempt to resolve those which crossed boundaries between the four subimages. As before, non vegetation classes such as bare ground, urban, turf, water and farmland signatures were included where they occurred. Goegrup and Harvey West were treated separately and class seed polygons for the vegetation classes for these images were defined after the initial classification processes of the initial four were complete. Harvey West and Goegrup class seed polygons were determined by air photo interpretation.

As a result, differing subimages had differing classes applied to the classification process. These class allocations to subimage can be seen in Table 5.

The class seed polygons used to generate the class spectral signatures were constantly revised to improve discrimination between classes. For instance due to the small ground resolution, Tuart forest class was made up of a number of discrete classes that actually defined portions of individual tree canopy. This was due to the differing reflectance associated with a large canopy and the fine spatial resolution of the sensor. Where Tuart were found in association with other classes such as Marri and Jarrah, particularly where the forest canopy was continuous, it became spectrally inseparable from other classes.