Changelog
Dataset updates, releases and significant changes
Updated
Apr 2026
2026
Apr
2026
Release
V1 platform launch — all four catalogs live
Initial public release of FuelMaps. Fuel Hazard (6 metrics), Fuel Load (GEDI L4A biomass), Fuel Type (NBIC ACS Stage 2 BFC) and Forest Metrics (6 layers) all available for 2017–2025 at 20m resolution.
All catalogs2017–202520m
Mar
2026
Update
2025 annual mosaic ingested for all hazard metrics
2025 COGs processed and deployed for all six fuel hazard layers. Note: 2025 GEDI biomass layer pending — estimated Q3 2026.
Fuel Hazard2025
Feb
2026
Update
Forest Metrics: rh98 LiDAR regression model updated
Canopy height model retrained with additional ALS calibration data from NSW and VIC. R² improved from 0.81 to 0.87. Affects 2022–2025 layers.
Forest Metricsrh982022–2025
Jan
2026
Fix
BFC classification: WUI boundary encoding corrected
WUI codes 920–940 were being masked due to an off-by-one error in the custom colormap JSON. Fixed in the titiler endpoint and COG reprocessed.
Fuel TypeBFCWUI zones
2025
Oct
2025
Release
Forest Metrics catalog added — 6 structural layers
New catalog: Woody Vegetation Extent, FPC, FHD, Canopy Height (continuous + classified) and PAI. Delivered at 10m for 2019–2024.
Forest Metrics10m2019–2024
Jul
2025
Update
Switched base embedding model to AlphaEarth V2
All fuel hazard and biomass layers regenerated using AlphaEarth V2 embeddings. Improved spatial coherence in sparsely vegetated areas.
All catalogsAlphaEarth V2
Apr
2025
Note
Temporal gap: 2019 Fuel Type layer uses 2020 mosaic
2019 BFC produced from a 2020 composite due to insufficient cloud-free coverage in northern QLD and WA. Treat 2019 BFC as approximate in these regions.
Fuel Type2019QLD · WA
Jan
2025
Release
Initial internal beta — Fuel Hazard & Biomass only
Beta access granted to Flare Wildfire internal team. Covered 2018–2024 annual series for hazard and biomass.
Internal betaFuel HazardBiomass
Assumptions
Project-level assumptions and caveats
5 entries
Field calibration data is representative of target biomes
Model
All models assume field-collected fuel assessment data is spatially and temporally representative. Significant gaps exist in arid/semi-arid zones and tropical savannas.
Affects: All catalogsPriority: High
GEDI footprint sampling is unbiased at 20m aggregation
Data
GEDI L4A footprints (~25m diameter) are assumed representative when aggregated to 20m. Slope >40° and canopy cover >95% cause under-sampling in tall dense forests.
Affects: Biomass · Forest MetricsPriority: Medium
Annual mosaic captures end-of-summer fuel state
Temporal
Each annual layer targets Feb–Apr imagery to capture post-summer fuel state. May not reflect peak hazard in tropical north (Oct–Nov).
Affects: Fuel HazardPriority: Medium
BFC classification does not capture post-fire succession
Model
The mapping is a static structural snapshot. Rapid post-fire recovery may cause the BFC layer to lag real-world transitions by 1–2 annual cycles.
Affects: Fuel TypePriority: Medium
Projected coverage is mainland Australia only
Scope
Current datasets cover mainland Australia and Tasmania. Islands and external territories are out of scope.
Affects: All catalogsPriority: Low
Known Issues
Active bugs, artifacts and limitations under investigation
4 open
2020 biomass layer has striping artifact in WA wheatbelt
High
Horizontal striping at ~1km intervals visible between lat −31° and −34°, lon 116°–120°. Likely GEDI orbit gap interpolation. Reprocessing planned Q2 2026.
Affects: Biomass 2020Status: In progress
Surface FH slightly overestimated in recently burned areas
Medium
Pixels burned within 6 months prior to mosaic acquisition can push Surface FH predictions 0.5–1 class above ground-truth. Fire history masking planned for V2.
Affects: Surface FHStatus: Open
FHD metric shows edge artifacts at state boundary extents
Medium
1–2 pixel seam artifact at NSW/QLD and SA/WA boundaries. Caused by different Sentinel-2 tile compositing windows per state.
Affects: Forest Metrics · FHDStatus: Open
Hover point query slow (>2s) for BFC fuel type layer
Low
Custom JSON colormap causes TiTiler point query ~2.1s vs ~0.4s for named colormaps. Pre-baked lookup approach being evaluated.
Affects: Fuel Type · hoverStatus: Open
Methodology
Data sources, modelling approach and processing pipeline
Reference doc
01 — Input data sources
Sentinel-2 multispectral imagery
Primary
Annual median composites from cloud-masked Feb–Apr scenes. Surface reflectance (NBAR) using BRDF correction from DEA. 10–20m bands.
Resolution: 10–20mCoverage: 2017–2025
GEDI spaceborne LiDAR (L2A, L4A)
Structural
NASA GEDI L2A height metrics and L4A aboveground biomass density. ~25m footprint from ISS orbit, non-contiguous sampling.
Footprint: ~25mCoverage: 2019–2024
Field fuel assessments (DSE Report 82)
Labels
~18,000 ground-truth fuel hazard ratings from trained assessors using the DSE Overall Fuel Hazard Assessment Guide. VIC, NSW and SA coverage.
Plots: ~18,000Coverage: VIC · NSW · SA
02 — AlphaEarth embedding model
A self-supervised vision transformer (ViT-B/16) pre-trained on 2.4B Sentinel-2 patches across Australia and Southeast Asia. Encodes 256×256px chips into 768-dimensional feature vectors capturing spectral-temporal patterns.

Downstream task heads (gradient boosted trees or shallow MLPs) trained on frozen embeddings per metric. Model: AlphaEarth V2 (Feb 2026).
03 — Output processing & delivery
Cloud-Optimised GeoTIFF (COG)
Overviewed, tiled COGs in EPSG:3857. Tile size: 512×512px. Compression: Deflate. Hosted on Cloudflare R2.
TiTiler dynamic tile service
Serves XYZ map tiles on demand with dynamic rescaling and colormap application. Point query endpoint for per-pixel hover readout.
04 — Validation
20% held-out test split. Spatial cross-validation (50km blocks) to account for autocorrelation.

Key metrics — Fuel Hazard OFH: weighted kappa = 0.74 · Biomass: R² = 0.72, RMSE = 28 t/ha · Canopy Height rh98: R² = 0.87, RMSE = 2.1m