Datasets:
MPEFire
MPEFire holds daily inputs for wildfire forecasting in Europe, with the channels, pixel size and 8-bit codes of MPCFire, a datacube for Canada. It covers 22 fire seasons, April 1 to November 4 of 2005-2026 (2026 ends on October 31), on 1025 × 835 pixels of 0.0449° (5.0 km north-south, 1.6-4.1 km east-west) between 11°W and 35°E and between 34°N and 71.5°N. Each day holds 42 channels: MODIS active fire, ERA5-Land weather, the six indices of the Fire Weather Index System, vapor pressure deficit and MODIS land-surface observations. Each season also holds 13 static channels, a land mask and a region mask.
Left: fire pixel-days of all 22 seasons in each cell of 5 × 5 pixels (0.22°) of the forecast area; dark gray marks forecast area without any, light gray other land. Right: fire pixel-days in the forecast area per season, with the means of 2005-2009, 2010-2019 and 2020-2025 (dashed lines); the 2026 season ends on October 31.
| Region | land of the countries that the US Department of State's boundary data (LSIB, 2017) assign to Europe, west of 34.7°E: 403,209 pixels, the forecast area |
| Grid | EPSG:4326, pixels of 0.0449158° (5.0 km north-south, 1.6-4.1 km east-west), 1025 × 835, 11.0°W-35.0°E, 34.0°N-71.5°N |
| Seasons | April 1 to November 4 of 2005-2026, 2026 to October 31: 22 seasons, 4,792 days |
| Daily channels | 42: active fire, 18 weather, 6 fire weather, vapor pressure deficit, NDVI, EVI, LAI, FPAR, 4 surface reflectance, 4 land-surface temperature and emissivity, 4 brightness temperature |
| Static channels | 13: distances to built-up land (3) and to water, population, 7 terrain, land cover |
| Values | uint8 codes on MPCFire's scale for each channel (ranges in channels.csv); 255 marks missing data (with two exceptions, see Encoding) and values at or above the range |
| Fire | 100,044 fire pixel-days in the forecast area, 5.2 × 10⁻⁵ of its pixel-days |
| Size | 22 archives of 3.7 GB, 82 GB in all; 7.8 GB per season unpacked, 172 GB in all |
| License | CC BY 4.0; the static channels ODbL 1.0 (LICENSE.md) |
| Version | 1.0, 2026-11-13 |
Files
| File | Content |
|---|---|
seasons/exp-eu-s<year>.tar.zst |
one season, 2005-2026 |
channels.csv |
the 42 daily and 13 static channels: names, sources, units, encoding ranges, processing |
grid.json |
the grid: CRS, affine transform, bounds, pixel centers |
CONTENTS.sha256 |
SHA-256 of every file inside the archives |
SHA256SUMS |
SHA-256 of the archives, CONTENTS.sha256, channels.csv and grid.json |
LICENSE.md |
licenses and the notices the sources require |
figures/fire_pixel_days.png |
the figure above |
A season unpacks to one folder:
exp-eu-s2025/
├── data_manifests/
│ ├── d4_cube.json days, shape, file names (relative to the folder) and channel names
│ └── d0_masks.json grid size
└── derived_data/
├── cube/
│ ├── dyn_2025.mmap daily channels: raw uint8 array (T, 835, 1025, 42), C order, no header
│ └── stat_2025.npy static channels: uint8 (835, 1025, 13), the same in every season
├── masks/
│ ├── land.npy bool (835, 1025): land, MODIS land cover other than water
│ └── canada.npy bool (835, 1025): the region (LSIB Europe west of 34.7°E); the name is kept from the Canadian version of this layout
└── events/
└── fire_index.npz dates: the season's days
The forecast area is land.npy & canada.npy; every count about Europe on this page refers to it, except the MOD14A1
coverage under Known issues. T is 218 days (214 in 2026), and row 0 of every array is the northern edge.
Download and unpack
pip install -U huggingface_hub
# one season and the metadata; leave out the file names to get all 22 seasons (82 GB)
hf download anfljsnfiawnfd/MPEFire --repo-type dataset --local-dir MPEFire \
seasons/exp-eu-s2025.tar.zst channels.csv grid.json CONTENTS.sha256 SHA256SUMS
cd MPEFire
sha256sum -c --ignore-missing SHA256SUMS
tar --zstd -xf seasons/exp-eu-s2025.tar.zst # GNU tar 1.31 or later; else: zstd -dc <archive> | tar -xf -
sha256sum -c --ignore-missing CONTENTS.sha256
Read a season
From the MPEFire folder, with NumPy only:
import csv, json
from pathlib import Path
import numpy as np
root = Path("exp-eu-s2025")
manifest = json.loads((root / "data_manifests/d4_cube.json").read_text())
season = manifest["per_year"]["2025"]
days = season["days"] # "2025-04-01" .. "2025-11-04"
dyn = np.memmap(root / season["dyn_file"], dtype=np.uint8, mode="r",
shape=tuple(season["dyn_shape"])) # (218, 835, 1025, 42)
stat = np.load(root / season["stat_file"], mmap_mode="r") # (835, 1025, 13)
area = np.load(root / "derived_data/masks/land.npy") & np.load(root / "derived_data/masks/canada.npy")
t = days.index("2025-08-01")
fire = dyn[t, :, :, 0] # 1 where MODIS saw a fire
print(int(fire[area].sum()), "fire pixels in the forecast area on", days[t])
# codes to values for a linear channel, here the daily mean 2 m temperature in K
channels = {r["name"]: r for r in csv.DictReader(open("channels.csv")) if r["array"] == "dynamic"}
ch = channels["temperature_2m"]
code = dyn[t, :, :, int(ch["index"])].astype(np.float32)
lo, hi = float(ch["vmin"]), float(ch["vmax"])
t2m = np.where(code == 255, np.nan, lo + (code + 0.5) * (hi - lo) / 255) # bin centers; 255: missing or >= vmax
# coordinates of the pixel centers
grid = json.loads(Path("grid.json").read_text())
lon = grid["x0"] + (np.arange(grid["width"]) + 0.5) * grid["res"]
lat = grid["y0"] - (np.arange(grid["height"]) + 0.5) * grid["res"]
The manifest's file names are relative to the season folder. A loader that opens dyn_file and stat_file as
given, from another working directory, needs them absolute; rewrite them once (afterwards d4_cube.json no longer
matches its line in CONTENTS.sha256):
for s in manifest["per_year"].values():
for k in ("dyn_file", "stat_file"):
s[k] = str((root / s[k]).resolve())
(root / "data_manifests/d4_cube.json").write_text(json.dumps(manifest, indent=1))
Channels
| Daily channels | Index | Source | Processing |
|---|---|---|---|
| Active fire | 0 | MOD14A1 (Terra), 1 km | 1 if any of the pixel's 5 × 5 one-kilometer pixels has fire-mask class 9 (high confidence), else 0 |
| Weather | 1-18 | ERA5-Land, hourly, 0.1° | 11 daily means, 2 daily maxima (2 m and skin temperature), 5 daily sums, over the 24 hourly steps 00-23 UTC; bilinear |
| Fire weather | 19-24 | ERA5-Land | FFMC, DMC, DC, ISI, BUI, FWI, computed every day from January 1 of the previous year (2023 for 2024-2026) |
| Vapor pressure deficit | 25 | ERA5-Land | Tetens' formula from the daily mean temperature and dew point |
| NDVI, EVI | 26-27 | MOD09GA, MYD09GA | from the codes of channels 30-32, not from reflectance |
| LAI, FPAR | 28-29 | MCD15A3H, 4-day, 500 m | quality-filtered mean per pixel of the newest composite; earlier composites fill gaps |
| Surface reflectance | 30-33 | MOD09GA, MYD09GA, 500 m | bands 1, 2, 3, 7; clouds removed; sensor chosen by quality flag; gap fill; 10 × 10 mean |
| Land-surface temperature | 34-37 | MOD11A1, MYD11A1, 1 km | day and night temperature, emissivity in bands 31 and 32; sensor chosen by quality flag; gap fill; 10 × 10 mean |
| Brightness temperature | 38-41 | MOD09CMG, MYD09CMG, 0.05° | bands 20, 21, 31, 32; Terra, else Aqua; gap fill; bilinear |
| Static channels | Index | Source | Processing |
|---|---|---|---|
| Distances to built-up land | 0-2 | GHSL built-up surface, 2020 | log(1 + distance in pixels) to pixels whose built-up share exceeds 0.5%, 2% and 1%; stand-ins for MPCFire's road, power-line and building densities |
| Distance to water | 3 | MODIS land cover, 2024 | log(1 + distance in pixels) to water pixels |
| Population | 4 | WorldPop, 2020 | log(1 + the pixel's mean population per 100 m cell) |
| Terrain | 5-11 | MERIT DEM | elevation, slope, aspect (sine, cosine), hillshade, topographic position (11 × 11 pixels), a wetness index from slope |
| Land cover | 12 | MODIS MCD12Q1, 2024 | IGBP class, the most common over the pixel |
Gap fill: each pixel of the source grid keeps its most recent valid value, with no age limit; the fill starts 60
days before the season (30 days for MOD09CMG and MYD09CMG; for channels 30-37 from June 26 on, see Known issues).
10 × 10 mean: mean of the valid one-kilometer values in the window of the pixel's 5 × 5 block and its neighbors to
the north, west and north-west, if at least 50 are valid. channels.csv gives each channel's name and modality in the
manifests, its source, unit, encoding range and processing.
Encoding
Every daily channel except fire holds
code = floor(255 · clip((x - vmin) / (vmax - vmin), 0, 1)), computed in float32,
with the channel's range [vmin, vmax] from channels.csv. Values at or above vmax get 255, values below vmin get 0,
and a code k below 255 covers x from vmin + k·Δ to vmin + (k + 1)·Δ, with Δ = (vmax - vmin) / 255. Missing values
get 255, with two exceptions. ERA5-Land has no values over the sea, and those cells are set to 0 before the bilinear
interpolation, so the weather channels 1-18 hold code 0 where only sea cells are near and values pulled toward 0 near
coasts (Known issues). In LAI and FPAR (28-29), a pixel without any valid composite holds 0. The MODIS channels 28-37
encode the products' stored integers (channels.csv gives the scale factors), the brightness temperatures 38-41 are
in kelvin, and NDVI and EVI come from the reflectance codes divided by 255. The fire channel holds 0 or 1.
Static channels 0-11 hold codes 0-254 on MPCFire's Canadian scale: a monotone map matches the quantiles of each stand-in over Canadian land to those of MPCFire's codes, so the codes rank pixels but do not convert back to physical values. Static channel 12 holds the land-cover class.
These ranges and conventions reproduce MPCFire's codes on Canada (Agreement with MPCFire on Canada), so that a code means the same in both datacubes.
Grid
grid.json defines the grid: EPSG:4326, pixel size 0.04491576420597608° (five pixels of Earth Engine's 1 km grid in
EPSG:4326, the pixel size of MPCFire's rasters), outer upper-left corner at 11.0044°W, 71.5059°N. The center of
pixel (row, col) lies at lon = x0 + (col + 0.5)·res, lat = y0 - (row + 0.5)·res.
How we built it
We computed every daily channel through Google Earth Engine from MODIS and ERA5-Land, the sources of MPCFire's daily channels. Where a processing choice was open, we took the setting whose Canadian rasters came closest to MPCFire's released ones (Agreement with MPCFire on Canada).
- Fire: MOD14A1 from Terra only, which matches MPCFire's fire channel. The fire channel of 17 days comes from NASA's archive (below).
- Weather: ERA5-Land hourly, aggregated to days over 00-23 UTC, then interpolated bilinearly from 0.1°.
- Fire weather: the six indices of the Canadian Fire Weather Index System on the ERA5-Land grid, every day of the year. Each season's recursion starts from the standard values (FFMC 85, DMC 6, DC 15) on January 1 of the previous year, at least 15 months before the season (January 1, 2023 for 2024-2026).
- MODIS land surface: Terra and Aqua combined, by quality flag for MOD11A1/MYD11A1 and MOD09GA/MYD09GA and with Terra first for MOD09CMG/MYD09CMG.
- Static channels: MPCFire's static channels come from OpenStreetMap, WorldPop, ASTER GDEM and MODIS land cover, as its dataset card lists them, and could not be reproduced exactly. We built open stand-ins with one recipe for Canada and Europe and placed them on MPCFire's code scale with a quantile map fitted on Canada.
Known issues
- Code 255 also means "at or above the range". The ranges come from MPCFire's Canadian scale: the daily mean 2 m temperature, for instance, reaches code 255 at 24.6 °C. Over the forecast area, 4.2% of the pixel-days of daily mean 2 m temperature hold code 255, 11% in July and August and 34% in July and August south of 45°N. For July and August south of 45°N, the share is 30% for the daily maximum 2 m temperature and VPD, 58% for the maximum skin temperature, 34-35% for DMC, BUI and FWI, 21% for ISI and MODIS band 7, and 15% for DC; over the whole area and season it is 2-8% for these channels. There, a model cannot tell the hottest and driest conditions from missing data. Version 1.0 keeps MPCFire's scale, so that a code means the same in both datacubes.
- Weather near coasts is pulled toward 0. ERA5-Land's sea cells are set to 0 before the bilinear interpolation, so pixels near the coast mix zeros into channels 1-18. On every day we checked (seven days of 2005, 2010, 2018 and 2025), 0.34% of the forecast area holds code 0 in daily mean 2 m temperature and 2.9-3.0% holds a code below 200 (below -39.6 °C), 99% of it within three pixels of the sea or inland water. Our ERA5-Land codes equal MPCFire's on Canadian land, so MPCFire has the same property.
- Negative reflectances enter the window mean. A negative MOD09GA/MYD09GA reflectance is stored as 65000 (a wrap into uint16) and counts as a valid value in the 10 × 10 mean of channels 30-33; a single one raises that mean by about 650, while band 7's range ends at 1,937.
- October 11-23, 2022 has (almost) no fire observations. Terra's MODIS delivered no data from October 11 to 22, 2022, during maneuvers that lowered Terra out of its satellite constellation, and covered 14% of the grid's box on October 23. The fire channel is zero on these 13 days without meaning that no fire burned; leave them out of the labels. Terra's other products are missing as well, so on these days the MODIS land-surface channels rest on Aqua and on values carried forward from before October 11.
- 17 fire days come from NASA's archive. On 16 days (May 11, August 29 to September 4, 2024, and September 14 to 21, 2026), Earth Engine's copy of MOD14A1 is empty or nearly so (0.8% coverage on May 11), while NASA's LP DAAC holds the observations. For these days, and for June 3, 2025 (51% coverage in Earth Engine), the fire channel comes from LP DAAC's granules, sampled onto the 1 km grid as Earth Engine samples them; on nine days held by both, the fire pixels of the two agree everywhere. Neither copy holds a fire pixel in the forecast area on June 3, 2025. For the granule of September 22-29, 2021, the two copies differ in 4 of 36 fire pixels on September 23, so September 22, 2021 (70% coverage) stays as Earth Engine has it. All other days without a fire pixel have at least 99.9% coverage (coverage: the share of the 1 km pixels in the grid's box with a valid MOD14A1 value of any class, cloud included).
- Fewer fire detections in later seasons. Fire pixel-days per season average 6,613 in 2005-2009, 4,514 in 2010-2019 and 3,041 in 2020-2025 (figure). Since its last inclination maneuver on February 27, 2020, Terra's equator crossing time drifts earlier, which could lower its fire detections; the decline starts before 2020, and we have not compared it with Aqua (MYD14A1). It matters when training on early seasons and testing on late ones.
- Approximate channels. The MODIS surface channels approximate MPCFire's (next section), and the static channels are stand-ins; the distance to water and population carry little of the information of MPCFire's channels.
- Gap fill from June 26. We first built every season from June 26 and later extended it to April 1. From June 26 on, channels 30-37, and NDVI and EVI (26-27) computed from them, keep the codes of the first build, whose gap fill started on April 27; they differ from one run started 60 days before April 1 only where a pixel had no valid observation from April 27 to June 25.
- Season end. For event times and censoring, a season ends on November 4 (October 31 in 2026).
Agreement with MPCFire on Canada
We ran the same pipeline on Canada and compared it with MPCFire's released rasters of 2025: the share of Canadian land pixels with the identical code, or, for the static stand-ins, Spearman's correlation over Canadian land.
| Channels | Agreement with MPCFire |
|---|---|
| Fire (1) | identical on 6 of 7 days; on the seventh, ours holds four extra fire pixels |
| ERA5-Land (18) | 16 identical, 2 at least 99.9%; 6 days |
| Vapor pressure deficit (1) | at least 99.9%; 6 days |
| Fire Weather Index System (6) | 50-98% identical, 85-99% within one code; recursion started from MPCFire's codes of September 2, eight days to October 26 |
| MOD11A1/MYD11A1 (4) | 67-89% identical; two windows of 120 × 120 pixels, July 6-15 |
| MOD09GA/MYD09GA (4) | 4-14% identical; one window of 120 × 120 pixels, July 6-15 |
| NDVI, EVI (2) | 5%, 8% identical; same window |
| MOD09CMG/MYD09CMG (4), MCD15A3H (2) | not measured |
| Terrain (7) | Spearman 0.63-0.88 |
| Distances to built-up land (3) | Spearman 0.80-0.84 |
| Distance to water (1) | Spearman 0.03 |
| Population (1) | Spearman 0.14 |
| Land cover (1) | 88.6% identical |
The fire-weather comparison started its recursion from MPCFire's codes; the cold start that MPEFire uses (How we built it) was not compared with MPCFire.
License
CC BY 4.0, except stat_<year>.npy, which is under ODbL 1.0 because its terrain channels derive from MERIT DEM.
LICENSE.md lists the sources and the notices they require.
Citation
The paper that introduces this datacube is under review; we will add its reference here. Please also cite the sources in LICENSE.md.
Changes
- 1.0 (2026-11-13): first release, 22 seasons.
- Downloads last month
- 56