The study area occupies the west part of Nitmiluk National Park and is close to the southwest boundary of Kakadu National Park that is inscribed in the World Heritage list for both its cultural and natural values. The major land uses in terms of area are Aboriginal land, pastoralism and nature reserves. The regional climate is characterised seasonally by rainfall with over 90% occurring in the summer wet season (from November to March). The average rainfall is 1 200 mm. Maximum daily temperatures average above 30oC over the year (Russell-Smith et al 1995). The wildfire pattern in the study area is characterised in terms of burning period during the dry season. The rugged Arnhem Land plateau exists in the east part of the study area with an elevation below 400 m. An undulating Cainozoic plain stretches away from the plateau margins and is composed of weathered and coarse-grained sediments (Williams 1991). An open forest/woodland type dominated by Eucalyptus spp. over a typically grassy understorey is the main vegetation type, while other freshwater floodplain communities ranging from sedgelands to open forests dominated by Melaleuca spp. distribute along the creeks in lowlands (Wilson et al. 1990, Russell-Smith 1995).
3. Methodology
In this project, one georeferenced Landsat Thematic Mapper image was acquired in June 2000 to demonstrate an operational method for the accurate mapping of fire scars that can be employed to extract a fire history for the past decade using this data. The accuracy assessment is based on field data and an IKONOS high-resolution data acquired within one week of the Landsat data. A 9-second Digital Elevation Model (DEM) acquired from Australian Survey and Land Information Group was available for the study area.
The interactive approach developed here employs a two-tired approach as can be seen in the flow diagram (Fig 2). On one hand, a conventional pixel-based, spectral classification is carried out to identify potential fire areas. On the other hand, areas that can be recognized as fire scars by the operator are manually delineated to provide the coverage of definite fire areas. The spectral classification is based on an unsupervised ISODATA algorithm of Landsat TM bands 3,4, and 5 generating 100 spectral classes. Considering that vegetation is highly sensitive to fire, the visible red, NIR and MIR bands were used in the image classification process, since this band combination has widely been used as a standard in vegetation studies (Conese and Maselli 1993). Each of the resulting classes is displayed on top of the false colour composite and assessed as being a fire scar or other land cover. This process is supported by available field data and the DEM to indicate potential water bodies and shadow areas. The classes are then aggregated to a binary image containing a value of '1' for all pixels that are potentially affected by fire. The manual delineation of definite fire areas was achieved by digitising the boundaries from the enhanced false colour composite with the help of field knowledge and the DEM. The boundary of fire patches was followed as closely as possible, attempting to eliminate shadows, water bodies, and other topographic features that may spectrally overlap with fire signatures. The digitised map was converted to a binary image with a value of '1' for the identified fire areas. The final fire map is produced by multiplying the binary maps from the classification and manual masking process.
In this project, two types of data were used to carry out accuracy assessment. Ground-truthing data were collected by recording fire locations and its coordinates in each 50 m interval by walking along four transects of 5 km length each set up uniformly in the study area. This fieldwork was carried out at the time of Landsat TM image acquisition. A total of three hundred field observations was used to assess the overall accuracy rate of fire mapping. Additionally, five hundred random points were tested on the high spatial resolution IKONOS data.
Figure 2. Flow chart of image processing steps for fire mapping in Australian savannas
4. Results and Discussion
The unsupervised spectral classification resulted in extensive confusion between fire classes, water bodies, shadows, and other topographic features such as mixed pixels in catchment areas where vegetation signatures are dominated by standing water. A more detailed classification, possibly using more spectral bands could be implemented, but there is doubt as to whether the problem could be solved. The real difficulty in this task is the nature of the fire scars to be mapped. Within the study area, the frequent fire events throughout the dry season lead to a great spectral variation to be tackled in the classification process. The fires in this environment burn the dry grass, but leave the sparse tree canopies unaffected. Thus the spectral signature is a mixture of the canopies and the burnt grasses, which are black at first, but rapidly, become lighter as the ash disperses, and eventually show signs of re-growth. For this reason, a relatively large number of spectral classes were generated to minimise the spectral overlap, but a more sophisticated classification methodology was not followed as its main goal is to become operational for the mapping of long-term fire history. The binary map of potential fire classes is shown in figure 3A.
