When Can We Predict Fire Severity?

A temporal analysis of the 2025 Grampians (Gariwerd) bushfire

Background

Fire severity maps are critical for post-fire ecological management, guiding decisions about erosion control, salvage logging, seed dispersal, and wildlife habitat recovery. However, the standard approach—differenced Normalised Burn Ratio (dNBR) from cloud-free post-fire satellite imagery—typically requires 2–4 weeks of waiting after fire containment. During large, complex fires, managers need severity estimates much earlier.

The 2025 Grampians bushfire provides an ideal case study. Ignited by dry lightning on 17 December 2024, the fire burned approximately 135,000 hectares across the Grampians National Park (Gariwerd) and surrounding farmland before containment in mid-February 2025. A second lightning-ignited wave on 27 January 2025 extended the fire into the southern ranges. The burned area encompassed roughly 80% of the national park, spanning diverse terrain including sandstone ranges, eucalypt forest, heathland, and riparian gullies.

This analysis asks: at which point in time can we predict burn severity, and how much does each data source contribute?

Methods

Study area and response variable

Analysis was restricted to the burned area within the Grampians National Park boundary (WDPA), excluding surrounding farmland where crop phenology confounds the dNBR signal. The response variable is dNBR (differenced Normalised Burn Ratio), computed as pre-fire NBR minus post-fire NBR from Sentinel-2 surface reflectance (bands B8A and B12, 20m resolution). Pre-fire composite: median of cloud-free scenes from 1 Oct – 10 Dec 2024. Post-fire composite: median from 15 Feb – 15 Apr 2025. Cloud masking used the Scene Classification Layer (SCL classes 4–7). Severity was classified using standard USGS thresholds: Low (0.10–0.27), Moderate-Low (0.27–0.44), Moderate-High (0.44–0.66), High (>0.66).

Sampling

10,000 sample points were extracted using stratified random sampling (2,500 per severity class) within the park burnt area at 100m spacing (seed=42). For each point, all predictor values were extracted from the corresponding satellite imagery via Google Earth Engine. Points retain their geographic coordinates for spatial cross-validation.

Predictor variables

Predictors were organised into six cumulative temporal tiers, reflecting the order in which data becomes available during a fire event:

T1 — Landscape (pre-fire): Pre-fire NDVI (Sentinel-2 median, Oct–Dec 2024) as a fuel load proxy. Slope, northness (cosine of aspect; positive = north-facing = drier in the southern hemisphere), elevation, and Topographic Position Index (TPI, elevation relative to 300m neighbourhood mean, distinguishing gullies from ridges) from SRTM 30m DEM. ESA WorldCover v200 vegetation type, one-hot encoded (tree/shrub/grass).

T2 — + Drought (pre-fire, time-varying): Antecedent soil moisture (ERA5-Land volumetric soil water layer 1, mean of 30 days before fire start: 17 Nov – 16 Dec 2024). Pre-fire vapour pressure deficit (VPD) derived from ERA5-Land 2m temperature and dewpoint, Oct–Dec 2024, using the Magnus formula: VPD = e_s(T) − e_a(T_d), where e_s = 0.6108 × exp(17.27T / (T + 237.3)).

T3 — + Fire weather (during fire): Each sample point was assigned a burn date from the earliest VIIRS active fire detection (system:time_start) at that location. ERA5-Land daily maximum temperature, 10m wind speed (from U/V components), and VPD were then extracted for that specific date. ERA5 resolution is ~9km, so points sharing a burn date receive identical weather values—the model learns which day of burning matters, not spatial weather variation.

T4 — + Satellite intensity (during fire): VIIRS (S-NPP + NOAA-20, 375m): maximum brightness temperature (Bright_ti4), maximum fire radiative power (FRP), and detection count across all passes during Dec 2024 – Feb 2025. Himawari-9 (JAXA P-Tree L2 Wildfire product, ~2km, 10-min intervals): cumulative FRP, maximum FRP, and fire duration (hours between first and last detection within 2km of each sample point).

T5 — + Post-fire indices (2–8 weeks): dCIre (differenced red-edge Chlorophyll Index: B7/B5 − 1, pre minus post), sensitive to canopy chlorophyll loss. dVH (differenced Sentinel-1 SAR VH backscatter, IW mode, pre minus post), detecting structural canopy damage independent of cloud cover.

T6 — + Recovery (months): NDVI recovery rate, computed as the linear slope of 14 monthly NDVI composites from Feb 2025 to Mar 2026.

Alpha Earth foundation model comparison

To test whether learned representations outperform hand-engineered pre-fire features, two additional standalone models were trained using 64-dimensional embeddings from Google DeepMind’s AlphaEarth Foundations (GOOGLE/SATELLITE_EMBEDDING/V1/ANNUAL in GEE). AlphaEarth is a foundation model trained on multi-sensor time series (Sentinel-2, Sentinel-1, Landsat) that encodes vegetation characteristics, phenology, surface moisture, and topographic context into dense unit-length vectors at 10m resolution (AlphaEarth Foundations, 2025).

AE_2023: 64 embedding dimensions from the 2023 annual composite — guaranteed pre-fire, representing landscape conditions one year prior to the fire.

AE_2024: 64 embedding dimensions from the 2024 annual composite — the most recent pre-fire snapshot, but potentially contaminated by ~2 weeks of fire signal (fire started 17 Dec 2024). Comparing AE_2024 vs AE_2023 provides a leakage check: a large difference would indicate the 2024 composite has encoded fire damage rather than pre-fire landscape.

These models were trained head-to-head against T1 using the same 10,000 samples, spatial block CV, and max_features tuning. Unlike the cumulative T1–T6 models, AE models use only embedding features — no hand-picked variables.

Bivariate analysis

Pearson correlations between each predictor and dNBR were computed from 5,000 stratified sample points (1,250 per severity class) to assess individual predictor strength.

Random Forest regression

Six cumulative Random Forest models (T1–T6) were trained using scikit-learn (RandomForestRegressor, 500 trees, oob_score=True, random_state=42). Each model adds a tier of predictors to the previous, quantifying the incremental variance explained at each temporal stage.

Cross-validation

Three R² estimates are reported for each model:

• Spatial block CV (primary): sample points were assigned to ~2km grid blocks based on their coordinates. GroupKFold (5 folds) holds out entire blocks, preventing the model from memorising spatial patterns. This is the most conservative and honest estimate.
• Random CV: standard 5-fold CV with shuffling. Expected to be higher than spatial CV due to spatial autocorrelation leakage.
• Out-of-bag (OOB): the inherent RF bootstrap estimate.

Model diagnostics

Permutation importance (10 repeats) from the full T6 model identifies which predictors drive predictions. Partial dependence plots (50-point grid) for the top 6 predictors reveal non-linear relationships and thresholds. A 2D interaction PDP for the top two predictors from different time groups shows whether predictor effects are conditional on each other. Predicted vs actual scatter uses spatial-CV held-out predictions to visualise model calibration.

FRP validation

To test whether satellite-measured fire intensity corresponds to spectral severity, Pearson and Spearman correlations were computed between dNBR and each FRP metric (VIIRS FRP, VIIRS brightness temperature, Himawari cumulative and maximum FRP). To address the resolution mismatch between VIIRS (375m) and dNBR (20m), sample points were aggregated to ~375m grid cells before computing correlations, with mean dNBR compared against the (constant) FRP value within each VIIRS pixel.

Signal validation

A 4-zone comparison (park vs farmland × burnt vs unburnt) validates that dNBR measures fire effects rather than seasonal phenology. Unburned park areas should show near-zero dNBR; farmland dNBR is expected to be confounded by crop senescence.

Results

1. Signal validation: dNBR works in native vegetation but not farmland

A 4-zone comparison (park vs farmland × burnt vs unburnt) revealed a critical confound in raw dNBR. Within the national park, the signal is unambiguous:

The unburned native vegetation shows effectively zero dNBR change, confirming that the index is measuring fire effects and not seasonal phenology. However, farmland tells a different story:

Unburned farmland shows a median dNBR of +0.367—comparable to genuine low-severity fire—driven entirely by crop senescence between the spring pre-fire and late-summer post-fire composites. The high standard deviation (0.287) reflects the patchwork of irrigated, dryland, and fallow paddocks. This confirms that dNBR-based severity classification is only meaningful within native vegetation. All subsequent analysis uses the Grampians National Park boundary (from the World Database on Protected Areas) to exclude agricultural confounders.

To test whether SAR-based severity avoids this confound, the same 4-zone comparison was applied to Sentinel-1 dVH:

Within the park, dVH discriminates cleanly: burned vegetation shows a median backscatter loss of +2.2 dB versus near-zero in unburned areas. However, farmland also shows substantial dVH change even without fire (median +3.4 dB), likely from crop harvest and ploughing altering surface roughness between the spring pre-fire and late-summer post-fire SAR composites. While dVH still separates burned from unburned farmland (+3.3 vs +3.4), the high non-fire baseline means neither optical (dNBR) nor SAR (dVH) indices reliably isolate fire effects in agricultural landscapes. This reinforces the decision to restrict analysis to the national park.

2. Severity distribution: a predominantly high-severity fire

Within the burned portion of the national park, severity was classified using standard USGS dNBR thresholds:

The fire was predominantly moderate-to-high severity, with moderate-high (38,045 ha) and high severity (29,952 ha) together accounting for 63% of the classified burned area. This is consistent with the fire conditions: multiple ignitions during extreme heat, strong winds, and prolonged drought.

Note: these thresholds derive from US FIREMON standards. He et al. (2024) showed that vegetation-specific thresholds improve severity classification accuracy in SE Australian forests. The area estimates above should be interpreted as approximate; the Random Forest regression below uses continuous dNBR values and is not affected by threshold choice.

3. Temporal prediction: when does each data source contribute?

The central research question—at which point in time can we predict severity—is addressed by examining correlations between dNBR and predictors available at each time step:

During-fire data adds meaningful signal. VIIRS detection count (r = -0.370) is a stronger individual predictor than any pre-fire variable. Pixels detected as active fire on more satellite passes tended to burn less severely, likely because repeated detections indicate slower-moving, lower-intensity flanking fire rather than a single high-intensity passage. However, detection count is also confounded by cloud and smoke obscuration, orbit timing, and scan angle (Bradstock et al. 2010), so this interpretation should be treated with caution. Maximum brightness temperature was less informative (r = -0.125), possibly due to the coarse 375m VIIRS resolution mixing high and low intensity pixels.

Post-fire indices carry the strongest signal. The red-edge chlorophyll index dCIre (r = +0.678) and SAR structural change dVH (r = +0.513) are both strongly correlated with dNBR. These are available within 2–4 weeks of fire containment, confirming that early post-fire imagery—even a single cloud-free Sentinel-2 pass—is far more informative than any pre-fire or during-fire data alone. Note that dCIre and dNBR share the same Sentinel-2 imagery (Fernandez-Manso et al. 2016), so their strong correlation partly reflects spectral redundancy between red-edge and SWIR bands rather than independent confirmation. In contrast, dVH from Sentinel-1 SAR provides the only truly sensor-independent severity measure, detecting structural canopy damage through an entirely different sensing modality (Hosseini & Lim 2023).

Recovery trajectory is the strongest single predictor (r = +0.710), but requires months of post-fire monitoring. This is consistent with the ecological interpretation that severity is ultimately defined by the biological response, not the spectral snapshot.

4. NDVI recovery: slow but progressing

Monthly NDVI monitoring shows the burned area began at ~0.30 immediately post-fire (Feb 2025) and recovered to ~0.57 by March 2026—a 14-month gain of approximately 0.27 NDVI units. However, this remains well below the unburned reference of ~0.65, indicating that full canopy recovery has not occurred after one year. The recovery trajectory shows a seasonal dip during winter (Jun–Aug 2025) before resuming upward, consistent with the known phenology of eucalypt resprouting in this region. However, resprouter-dominated communities (common in the Grampians) recover faster spectrally than obligate seeders, and stratifying recovery by vegetation type would strengthen this analysis (Gibson & Hislop 2022; Caccamo et al. 2015).

Discussion

Bivariate correlations

The initial hypothesis—that pre-fire fuel load and topography might explain a large proportion of severity variance—was not supported by bivariate analysis. Individual pre-fire predictors explained less than 7% of variance each. The progression of bivariate predictive power follows a clear temporal gradient:

Multivariate Random Forest analysis

To move beyond bivariate correlations, a Random Forest regression was trained on 10,000 stratified sample points with cumulative temporal predictor sets (T1–T6). All R² values below are from spatial block cross-validation (~2km blocks, GroupKFold), which guards against inflated accuracy from spatial autocorrelation.

5. Alpha Earth embeddings vs mechanistic features

The Alpha Earth comparison is informative but does not support replacing mechanistic features with learned embeddings. While AE_2023 outperforms T1’s 8 landscape features alone, it only matches what 10 interpretable variables (T2) already achieve — using 64 black-box dimensions to do so. The embeddings add no new predictive signal beyond what hand-engineered features plus antecedent drought already capture, and they sacrifice interpretability entirely. The 64 dimensions have no physical meaning, making them unsuitable for understanding why certain areas burn more severely.

Critically, the embeddings cannot incorporate fire-day weather or satellite-detected fire intensity (T3, T4), which add a further +0.13 R² beyond T2. For pre-fire severity prediction, domain-specific features with temporal specificity outperform general-purpose landscape representations.

The key insight is that ~46% of severity variance is predictable before or during the fire (T4 spatial CV R² = 0.46), a meaningful improvement over the bivariate result. The remaining variance is only recoverable with post-fire satellite imagery. For operational forecasting, this means landscape + drought + weather + VIIRS data could produce a useful preliminary severity map within hours of fire passage, well before cloud-free optical imagery becomes available.

Foundation model embeddings as landscape descriptors

The Alpha Earth comparison tested whether a general-purpose foundation model could outperform domain-specific feature engineering for pre-fire severity prediction. The answer is no: AE_2023 (R² = 0.33) matches but does not exceed the combined explanatory power of landscape features plus antecedent drought (T2, R² = 0.33). The 64-dimensional embeddings likely encode a mixture of the same information captured by NDVI, terrain, and vegetation type, plus some implicit drought signal from seasonal spectral patterns — but this does not translate into additional predictive power.

The result underscores the value of mechanistically-reasoned features for fire severity analysis. Hand-picked variables like soil moisture and VPD provide equivalent predictive power to a foundation model trained on petabytes of global satellite data, while remaining interpretable and directly informative for land management. The temporal specificity of features like burn-date weather and satellite-detected fire intensity (T3, T4) — which annual embeddings cannot capture — adds a further +0.13 R² that no landscape embedding can provide.

Alpha Earth embeddings may still have value for rapid transferability to new fires where the full GEE feature extraction pipeline has not been set up, since the embeddings are pre-computed globally. But for a single well-studied fire, purpose-built features are preferable.

Signal validation

The 4-zone validation (park/farmland × burnt/unburnt) highlights an important methodological caution: raw dNBR is unreliable in agricultural landscapes. Seasonal crop phenology produces dNBR signals comparable to low-severity fire. The Grampians National Park boundary from WDPA provided a clean analytical mask, and the near-zero dNBR in unburned park areas confirms the index is valid within this domain. Note that the analysis is restricted to burned area within the national park, which represents a subset of the total ~135,000 ha fire extent; severity patterns in surrounding farmland and state forest are excluded.

Spatial block cross-validation R² is consistently lower than random cross-validation (e.g., T6 spatial 0.81 vs random 0.84), confirming that spatial autocorrelation inflates naive accuracy estimates. The spatial CV results provide an honest assessment of how the model would generalise to unseen areas of the fire.

Limitations

The weak FRP–severity correlations represent a confirmed null result rather than a data gap. Resolution mismatch is the primary driver: dNBR is measured at 20m, VIIRS FRP at 375m, and Himawari FRP at ~2km, meaning each FRP pixel integrates over vastly heterogeneous severity outcomes. This finding is consistent with Nguyen et al. (2024), who found that FRP–severity relationships weaken substantially at coarser resolutions, and Heward et al. (2013), who reported similarly weak correlations in Australian eucalypt forests. ERA5 weather is at ~9km resolution, meaning all points on a given burn date share the same weather values; finer-resolution weather (e.g., ACCESS or downscaled reanalysis) could improve T3. Sentinel-1 SAR data (dVH) may underperform in areas of gentle topography where geometric distortion is minimal. The recovery trajectory is based on 14 months of data; longer monitoring would improve severity confirmation.

Next steps

Integration of Himawari geostationary FRP (10-minute temporal resolution, ~2km spatial) could fill the gap between VIIRS passes and improve during-fire severity prediction. Higher-resolution fire weather data from ACCESS or BoM station observations could improve T3. Comparison with the official DELWP fire severity mapping (when released) would provide independent ground validation. Extending the approach to other recent fires would test generalisability beyond the Grampians.

Data Sources

Analysis by James Maino. Built with Google Earth Engine, Leaflet, and Chart.js.

Grampians 2025 Fire Severity