isp-uv-es/wildfirespreadts-taco

WildfireSpreadTS

This is a repackaging, not a new dataset. It is WildfireSpreadTS by KTH Royal Institute of Technology (Gerard, Zhao, Sullivan), NASA/NOAA VIIRS, USGS SRTM, NASA MODIS, gridMET, NOAA GFS, converted to TACO. Pixel values and labels are kept as released except where the description below says otherwise. All credit belongs to the original authors: if you use it, please cite them and follow their licence.

original dataset · paper · licence: CC-BY-4.0

Repackaged into TACO by the Image and Signal Processing Group (ISP), Universitat de València, within the ELLIOT project.

Citation

Please cite the original work:

@inproceedings{gerard2023wildfirespreadts,
  title     = {WildfireSpreadTS: A dataset of multi-modal time series for wildfire spread prediction},
  author    = {Gerard, Sebastian and Zhao, Yu and Sullivan, Josephine},
  booktitle = {Advances in Neural Information Processing Systems 36 (Datasets and Benchmarks)},
  year      = {2023}
}

About the data

Daily multi-modal time series over 607 wildfire events in the western United States (2018, 2019, 2020, 2021), at 375 m nominal resolution: 9 to 94 co-registered days per event.

607 samples · splits: test 201 · train 230 · validation 176 · tasks: semantic-segmentation

Packaged as TACO v3.

Full description

Channels. Per day: VIIRS reflectance (M11/I2/I1), the VNP13A1 vegetation indices NDVI and EVI2, gridMET surface meteorology (precipitation, wind speed and bearing, minimum and maximum temperature, energy release component, specific humidity, Palmer drought severity), a GFS forecast of the same quantities, MODIS IGBP land cover, and the VIIRS active-fire detections. Slope, aspect and elevation are stored once per event rather than once per day, being bit-identical across every day of every event; land cover stays per day because it varies in one event of the 607.

Active fire. The raster stores the UTC granule time of each detection as an HHMM code, with 0 for no detection, rather than a binary mask. Published baselines binarise it with fire > 0, which is an exact function of what is stored; the reverse is not.

Splits. The release publishes no train/val/test split: the authors prescribe leave-one-year-out cross-validation. split carries one of their folds, whole years at a time (2018 val, 2019 train, 2020 test, 2021 train), and year is carried so every other fold stays reconstructible. The fold is lopsided (230 train, 176 val, 201 test): four years over three parts is 2/1/1. It is not spatially disjoint either. An event raster is 111-128 km across and 4,129 of the 183,921 event pairs overlap on the ground; that overlap graph has six components with 595 of the 607 events in one, so no spatially disjoint partition exists.

Caveats.

• The GFS forecast precipitation channel is a daily sum in 5,131 of the 13,607 rasters and the last hourly value in the other 8,476, carried per frame as forecast_precipitation_agg.
• The GFS forecast wind direction spans only -90..+90 degrees, the signature of an atan rather than an atan2, so its quadrant is unrecoverable and it is declared uncalibrated.
• Every raster is tagged UTM zone 10N regardless of where its fire is, so the resolution is measured geodesically per event (361.3-374.1 m) rather than read off the transform.

Getting started

git clone https://github.com/OscarPellicer/taco
pip install -e "taco/python[ml]"     # builds the native reader: needs CMake and Ninja

Read it straight from the Hub:

from huggingface_hub import snapshot_download
from taco.ml import Dataset, plot_sample

path = snapshot_download("isp-uv-es/wildfirespreadts-taco", repo_type="dataset")
ds = Dataset(path)                  # the repository is the container: parts + .tacocat
plot_sample(ds[0])

or from a local copy:

ds = Dataset("wildfirespreadts")
sample = ds[0]                      # {slot name: SlotValue}, arrays decoded
sample["viirs"].array.shape

Metadata without decoding anything:

import taco
taco.read("wildfirespreadts/.tacocat")              # one Arrow table, levels joined

Samples

What a sample contains

role	slot	holds	modality	detail
input	viirs	raster_series	optical	3 band(s), unit 1, scaled
input	vegetation	raster_series	biophysical	2 band(s), unit 1, scaled
input	precipitation	raster_series	atmospheric	2 band(s), unit m, scaled
input	wind_speed	raster_series	atmospheric	2 band(s), unit m s-1, physical
input	wind_direction	raster_series	atmospheric	1 band(s), unit degree, physical
input	forecast_wind_direction	raster_series	atmospheric	1 band(s), digital_number
input	temperature	raster_series	atmospheric	2 band(s), unit K, physical
input	forecast_temperature	raster_series	atmospheric	1 band(s), unit K, scaled
input	humidity	raster_series	atmospheric	2 band(s), unit 1, physical
input	energy_release	raster_series	biophysical	1 band(s), unit 1, physical
input	drought	raster_series	biophysical	1 band(s), unit 1, physical
input	slope	raster	elevation	1 band(s), unit degree, physical
input	aspect	raster	elevation	1 band(s), unit degree, physical
input	elevation	raster	elevation	1 band(s), unit m, physical
input	landcover	raster_series	label_raster	1 band(s), 18 classes, digital_number
target	fire	raster_series	thermal	1 band(s), digital_number

Licence

CC-BY-4.0

Providers: KTH Royal Institute of Technology (Gerard, Zhao, Sullivan), NASA/NOAA VIIRS, USGS SRTM, NASA MODIS, gridMET, NOAA GFS

Acknowledgements

TACO was designed by César Aybar and is specified at https://asterisk.coop/taco/spec/.

Built by Oscar Pellicer within the ELLIOT project at the Image and Signal Processing Group (ISP), Universitat de València.