fire-fusion-cascades-500m

FireFusion Cascades 500m

Daily spatio-temporal datacube for wildfire ignition and cause prediction over the Eastern Cascades of Washington State, a 272 km square running from the Cascade crest through the Okanogan Highlands and the most fire-active terrain in the state. Ten geospatial products spanning terrain, fuels, weather, human activity, lightning, and fire history are aggregated onto a single daily 500m $\times$ 500m grid spanning 2003-2020 continuously.

• Full daily coverage from 2003-2020, supervised over the May-October fire season (months that carry active fire activity).
• 38 input channels (25 from source processors, 13 derived), normalized, resampled and interpolated into trainable feature layers.
• Custom derived features: Per-cause ignition KDE's, 3x3 cell 7d rolling fire occurrence, NDVI anomalies, 2 and 5-day cumulative precipitation, 100 and 1000-hr dead fuel moisture, decayed lightning load, Fosberg FWI.
• Circular quantities (N/S and E/W aspect and wind-direction components, day-of-year) decomposed into orthogonal components, so no channel carries 0/360 discontinuities.
• Labels for fire ignition time, cause, KDE by cause, 3x3 rolling fire occurrence, and months since last burn.
• Mask layers for water and active fire, for ignition prediction.

---

Built by Tanner O'Rourke as part of FireFusion, independent research conducted on multi-source wildfire ignition modeling. The full pipeline, from raw-source extraction to model training, is at FireFusionNet. If this work is useful to you, a star on the repository helps it reach more people. Feel free to reach out!

@misc{orourke2026firefusion,
  author = {O'Rourke, Tanner},
  title  = {FireFusion: multi-source datacubes and a spatiotemporal ConvFormer for wildfire ignition prediction},
  year   = {2026},
  url    = {https://github.com/tannerorourke/FireFusionNet}
}

---

Files

dataset.zarr

The primary artifact, cut prior to compilation; Every transform with a data-estimated parameter (any z-score, min-max, or scale) is deferred out of it and into the split stage, so it carries no knowledge of where the split boundaries fall. The time axis carries the same off-season gaps as the splits, so the contiguous-block check shown for those applies here unchanged.

Differences from the splits store:

• No statistical normalization. Only the deterministic steps (clip, log1p, to_sin, per_area) are applied, so values sit in native units after those steps. Every z-score and min-max is fit at compile time against whichever years you designate as training, which is what keeps this store free of split-boundary leakage.
• Four cause classes, not three. ign_next_cause runs 0-3 with DEBRIS held separate from INDUSTRIAL, and dataset_manifest.json records n_cause_classes: 4. See Labels.
• 46 channel variables versus the splits' 38. The extra eight are raw layers the derived channels are built from, kept for provenance and dropped by compile: lf_aspect and wind_dir before their component decomposition, modis_ndvi before the anomaly, plus modis_water_mask, rh_max, usfs_burn_occ, usfs_burn_cause, and usfs_perimeter.
• Features are held as named variables rather than a stacked tensor, so a channel loads independent of the other 45.

Schema

Coordinates are time (datetime64[ns], 3312 days), y and x (float64, UTM metres), and burn_cause (the four cause names). Every array is chunked (16, 544, 544), 16 days at full spatial extent. Store size 72 GB.

Loading

ds = xr.open_zarr("dataset.zarr")             # 3312 days, no split boundary

vpd = ds["vpd_max"]                           # (3312, 544, 544) float32
window = ds[["temp_max", "wind_mph", "lightning_load"]].isel(time=slice(0, 10))
y, cause = ds["ign_next"], ds["ign_next_cause"]

# stack an arbitrary channel set into a model-shaped tensor
chans = ["temp_max", "vpd_max", "lf_elevation"]
X = ds[chans].to_array("channel").transpose("time", "channel", "y", "x")

To cut your own splits, set train_yrs / eval_yrs / test_yrs in fire_fusion/config/dataset_config.py and run compile against this store. Every statistic refits on the years you name:

python -m fire_fusion.dataset.build --dataset cascades500 --stage compile

train.zarr / eval.zarr / test.zarr

Train-ready splits. Each is a zarr store holding X with shape (time, channel, y, x) in float32, stacked tensor along with labels and masks. The task is forecasting, these suggested splits are chronological and balanced positive label counts with training data size. Note that any random sampler will leak future weather in training.

The splits are cut for comparability and not the only choice. Cut your own with the compile stage of the FireFusionNet repo, which refits every statistic on whatever you designate as training data.

Schema

channel is a labelled coordinate carrying the names in manifest.json, so ds["X"].sel(channel="temp_max") resolves without an index lookup. time is datetime64[ns], and y and x are float64 UTM metres. X is chunked (16, 38, 182, 182), labels and masks (64, 544, 544).

Loading

import numpy as np
import xarray as xr

ds = xr.open_zarr("train.zarr")              # local path, or an fsspec URL
x = ds["X"].isel(time=slice(0, 10))          # (10, 38, 544, 544)
y = ds["ign_next"].isel(time=9)

# the time axis skips the off-season, so build windows within a contiguous block
days = np.asarray(ds.indexes["time"], dtype="datetime64[D]")
block = np.concatenate([[0], np.cumsum(np.diff(days).astype(int) != 1)])
# a length-W window at t is valid iff block[t] == block[t + W - 1]

Note on Streaming

If training or compiling the data, download first. Streaming is only recommended for inspection; split X is chunked (16, 38, ~182, ~182), the grid split 3 x 3 spatially, about 80 MB per chunk decompressed and ~40 MB on the wire. A full-grid day touches 9 chunks and a 10-day window 9 to 18, so an epoch of full-grid windows moves an order of magnitude more data than downloading the train split once.

Grid Details

The bounds are sized so the 544-cell extent divides evenly under strided encoders and needs no alignment crop. The north edge clamps to the 49th parallel, where the US sources stop.

Time axis

Only in-season days ship, so the time axis is not contiguous: within a year it runs May 1 to Oct 31 day-by-day, then jumps to the next May. The manifests record time.season_months = [5, 10] and time.contiguous = false.

Halo days

Each year is built from a single block running March 22 to November 10, deliberately wider than the window. This halo -a 40-day lead and a 10-day trail - lets temporal derivations enter the supervised window with real history instead of restarting at zero. Every backward-looking channel (decayed lightning load, 2 and 5-day cumulative precipitation, the per-cause ignition KDEs) is computed on the wider index, and the halo is dropped before any normalization statistic or class balance is taken, so those describe exactly the days that ship. The 40-day lead is sized by the longest backward operator in the pipeline, the lightning-load IIR, which decays below 0.1% there.

Halo days are never supervised and never scored. They survive only in the build-time staging cube (cube.zarr), which is not distributed, and are absent from every artifact listed above. A sliding-window loader must still avoid building any window that straddles the year-to-year gap. See Loading.

State that legitimately spans years decays by elapsed time, not by index position. The per-cause ignition KDEs apply a 365-day half-life to the true day count between consecutive entries, so a multi-year prior crosses the roughly 4.5-month off-season gap correctly attenuated instead of stepping across it as a single day.

234 days are extracted per year against the 184 that are supervised. Over 2003-2020 that is 4212 days extracted and 3312 published.

Labels

Note: dataset.zarr carries a fourth class DEBRIS. The splits compile DEBRIS into INDUSTRIAL together, since both are a few hundred cases, versus lightning and human causes which are the overwhelming majority of cases. Regroup them by re-running compile against the published cube; n_cause_classes in each manifest is authoritative for the store beside it.

Masks

Ignition is heavily imbalanced: ign_pos_weight = 3754.89 on the train split. Prevalence per cell falls as cells shrink, but the fire-dense extent pulls the other way: the 500 m base rate (~2.7e-4) lands just below the 2 km statewide tier's ~3.7e-4. Restricting to the fire season removes winter cell-days that are near-uniformly negative, so this is a fire-season base rate and not an annual one. Losses should apply land_mask and no_act_fire_mask; the cause head should additionally apply valid_cause_mask.

Channels

Deterministic steps (clip, log1p, to_sin, per_area) are already applied in dataset.zarr and recorded in dataset_manifest.json. Statistical steps (z_score, minmax) are fit on the train years only and applied in the split stores, recorded in manifest.json. Both are listed together below in the order they compose.

In-depth extraction and feature details can be found in the repo's SOURCING.md.

The four kde_* channels are fire-history kernel density estimators in events per km squared, with a 20 km smoothing radius and a 365-day decay half-life. The per_area step is what puts them in those units; the raw accumulator is mass per cell and would rescale with cell size.

ndvi_anomaly is NDVI minus its day-of-year climatology, averaged over the train years only so no held-out day contributes to the mean it is measured against.

Sources

dataset_manifest.json and manifest.json are the authoritative record of channel order, normalization steps, grid geometry, and class balance. Read them rather than hardcoding.

• PRISM: temperature, dewpoint, vapour-pressure deficit, precipitation
• NOAA AORC: humidity, wind
• MODIS MOD13Q1 / MCD15A2H / MCD64A1: NDVI, leaf area index, burn history, water mask
• LANDFIRE: elevation, slope, aspect
• NLCD: canopy cover, impervious surface
• USFS FOD and fire perimeters: ignition and cause labels
• NOAA NCEI SWDI: lightning
• GPW: population
• USDA WUI: wildland-urban interface
• US Census TIGER: roads vectors

Fire Fusion collection

All FireFusion datasets carry identical channels, labels, masks, and split years. Three are statewide, cascades500 (this one) covers only the Eastern Cascades sub region.

Note: Class balance is not identical; Ignition prevalence per cell rises with cell size. (e.g., ign_pos_weight is 944.85 at 4 km vs. 2683.80 at 2 km). See manifest for exact values.