isp-uv-es/digital-typhoon-taco

Digital Typhoon

This is a repackaging, not a new dataset. It is Digital Typhoon by National Institute of Informatics (Kitamoto et al.), JMA / Himawari and GMS, 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{kitamoto2023digital,
  title     = {Digital Typhoon: Long-term Satellite Image Dataset for the Spatio-Temporal Modeling of Tropical Cyclones},
  author    = {Kitamoto, Asanobu and Hwang, Jared and Vuillod, Bastien and Gautier, Lucas and Tian, Yingtao and Clanuwat, Tarin},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS) Datasets and Benchmarks},
  year      = {2023}
}

About the data

192956 geostationary infrared images of 512x512 covering 1116 western North Pacific tropical cyclones from 1978 to 2023, each joined to the JMA best-track record for its hour: centre latitude and longitude, central pressure in hPa, maximum sustained wind in knots and storm grade. 45 years of the same observation over the same basin.

192,956 samples · splits: test 44,023 · train 109,919 · validation 39,014 · tasks: regression, classification

Packaged as TACO v3.

Full description

Radiometry. Pixels are calibrated brightness temperatures in kelvin, declared physical: no scale, no offset, roughly 130 to 313 K. float64 is narrowed to float32, which is lossless at this instrument's precision.

Source. NII's V2 WP.zip, read in place. torchgeo's WP.tar.gz is not used, as its archive is incomplete.

Sample. One image, not one storm: the best-track label is per image and the sequences run from a few frames to over 500. sequence_id and sequence_index make any temporal window a group-by. The image/best-track join is on the release's file_1 column and is exact -- every row with a filename resolves, and no image is left unmatched.

Splits. Digital Typhoon publishes no split. The one here is chronological by season (109919 train, 39014 val, 44023 test images) and keeps whole cyclones together; a random cut over hourly frames would place nearly identical consecutive images of one storm on both sides.

Classes. Grade indices are the JMA codes themselves, not renumbered.

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/digital-typhoon-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("digital-typhoon")
sample = ds[0]                      # {slot name: SlotValue}, arrays decoded
sample["ir"].array.shape

Metadata without decoding anything:

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

Samples

What a sample contains

role	slot	holds	modality	detail
input	ir	raster	thermal	1 band(s), unit K, physical
target	pressure	scalar		unit hPa
target	wind	scalar		unit kn
target	grade	class_index		8 classes

Licence

CC-BY-4.0

Providers: National Institute of Informatics (Kitamoto et al.), JMA / Himawari and GMS

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.