Wildfire Risk in Australia

Australia contains two substantially different broad fire environments. Across northern Australia, savanna and spinifex burn extensively during the dry season in most years. These large fires are a recurring feature of the northern fire regime. In southeastern Australia, eucalypt forests burn less frequently but can produce severe fires when high temperatures, strong winds, and dry conditions occur together. Both environments feature fuel-driven spread, pronounced dry seasons, and fire regimes that Aboriginal and Torres Strait Islander peoples have managed with fire for millennia. Their primary differences are fire frequency, burned area, and the level of human and property exposure.

The 2019 to 2020 season illustrates the severity of the southeastern fire regime, and Table 1 summarizes its extent. Approximately 5.5 million hectares burned in New South Wales, where 26 lives were lost. An additional 1.5 million hectares burned in Victoria during the same season. A national model must therefore represent both broad fire environments.

SWIM uses fuel behavior models from the Australian Fire Danger Rating System rather than translating Australian vegetation into an American fuel classification. The spread equations, fuel families, and moisture relationships are those used by Australian fire agencies. Table 2 compares the burned area and modeled losses associated with the two environments.

Table 1. Reported burned area for selected Australian fire seasons, as published by the New South Wales Bushfire Inquiry, the Inspector-General for Emergency Management Victoria, and the Victorian Department of Energy, Environment and Climate Action. New South Wales is 3.6 times the area of Victoria. As a share of each state, however, the 2019 to 2020 season burned a similar area in both jurisdictions.

SWIM estimates a range of possible final fire sizes rather than a single deterministic outcome. The range is expressed as percentiles. The fire-size model is built from 157,933 mapped Australian fire extents and evaluated on geographically separate areas that were excluded from training. A fire is therefore not evaluated against nearby fires from the same landscape.

These 157,933 records form the primary dataset used to fit and evaluate the fire-size model. Each record is a mapped fire perimeter published by an Australian state agency through Geoscience Australia's Bushfire Historical Extents, with a known ignition point and final area. They are mapped fire perimeters rather than individual satellite detections, incident reports, or modeled events. The archive contains 324,741 perimeters. The analysis excludes 155,311 records classified as deliberate burning and another 9,049 records below the model's minimum representable area of 0.09 hectares. The retained dataset contains 157,933 records from 1898 through 2024. Of these, 73,444 are classified by the source as wildfire and 93,538 have no recorded type and are retained.

Fire duration is estimated using a separate Australian model. The rate at which burned area grows through time follows a curve reconstructed from satellite observations of mapped Australian fire perimeters. Historical Australian fires had burned approximately 2.9 percent of their final area after the first 10 percent of their duration, close to the corresponding United States value and different from the Chilean value.

Model results are reported in hectares and Australian dollars, and the underlying quantities are derived from Australian records. Table 3 identifies each component and its primary data source.

Table 3. Primary Australian data sources used for each model component. The fatality rate is calibrated on published civilian death tolls from Australian fire seasons.

Each request returns a range of outcomes rather than a single number. The 90th-percentile estimate represents a final fire size that approximately 90 percent of fires are expected to remain below and 10 percent may exceed.

A well-calibrated percentile should contain approximately the proportion of observed outcomes that it represents. Across 157,933 mapped Australian fires, 90.0 percent of observed final fire sizes fell at or below the model's 90th-percentile estimate. Coverage remained within 1.3 percentage points of the target across the model's prediction range, from fires of 20 hectares to fires of 6,000 hectares.

The model was evaluated using five-fold geographic cross-validation. Australia was divided into 592 blocks measuring 100 km by 100 km and assigned to five groups. For each evaluation round, one group was excluded from model training and used only for testing. One-fifth of Australia was therefore excluded during each round, and every historical fire was evaluated once using a model that had not been trained on fires from the same geographic block. Geographic blocks were used instead of randomly splitting individual records because 102,487 records are in Western Australia and the archive contains repeated burns in the same areas over a century. A random split could place fires from nearly identical landscapes in both the training and test data.

Table 4 reports performance of the 90th-percentile fire-size estimate across the complete dataset. Each of the eight fires in Figure 1, listed in Table 5, finished within 15 percent of the estimate produced from location, season, and landscape alone.

Table 4. Performance of the 90th-percentile fire-size estimate on geographically held-out data. Overall, 90.0 percent of observed fires finished at or below the modeled 90th percentile. Each test excluded complete 100 km geographic blocks from training.

The fire-size model estimates how large a fire could become. This section evaluates whether SWIM also reproduces the geographic area affected by selected historical fires. Each case in Table 7 uses observed weather from the day of the fire and begins at the earliest satellite detection. Its fire-size estimate was generated by a model that was not trained on fires from the same geographic test block.

Beginning each reconstruction at the earliest satellite detection prevents the model from using information about the fire's eventual extent when generating the footprint.

The primary metric is the percentage of the observed burn area contained within the modeled footprint. For planning applications, missed portions of the eventual burn fall outside the modeled planning envelope. Each fire is first evaluated at the 90th percentile of its own fire-size distribution and at the 95th percentile when the fire exceeded the 90th-percentile estimate. These percentiles are reference points within a continuous modeled distribution, not maximum possible fire sizes. The distribution can be evaluated through the 99.9th percentile. Table 6 lists the inputs used for each reconstruction.

Table 6. Inputs used for the historical fire reconstructions. Each simulation begins at the earliest satellite detection, represented by the center of the overpass rather than a single pixel, and uses observed weather from the day of the fire. The footprint is generated using the same rule applied to a live request.