isp-uv-es/earthdial-uhi-taco

EarthDial-Instruct: urban heat island (Landsat-8)

This is a repackaging, not a new dataset. It is EarthDial-Instruct: urban heat island (Landsat-8) by EarthDial authors (MBZUAI / IBM), USGS / NASA (Landsat-8), 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: Apache-2.0, public-domain

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{soni2025earthdial,
  title     = {{EarthDial}: Turning Multi-sensory Earth Observations to Interactive Dialogues},
  author    = {Soni, Sagar and Dudhane, Akshay and Debary, Hiyam and Fiaz, Mustansar and Munir, Muhammad Akhtar and Danish, Muhammad Sohail and Fraccaro, Paolo and Watson, Campbell D. and Klein, Levente J. and Khan, Fahad Shahbaz and Khan, Salman},
  booktitle = {2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  pages     = {14303--14313},
  year      = {2025},
  doi       = {10.1109/CVPR52734.2025.01334}
}

About the data

648 urban-heat-island dialogues over 8-band Landsat-8 tiles at 30 m, 224x224 px, one per (tile, month).

648 samples · splits: test 216 · train 336 · validation 96 · tasks: visual-question-answering, scene-classification

Packaged as TACO v3.

Full description

Records. Each is a three-exchange conversation: whether the ground under the tile runs cooler, mildly hot or extremely hot; what land use drives it; what would mitigate it. The three answers are also lifted into temperature_trend, landuse and mitigation, the schema the release's test half uses.

Bands. The order is B1-B7 then B10, settled from the pixels: the release states only '8 bands'. Band 7's normalised spatial gradient is 0.0199 against 0.080-0.103 for bands 1-6, five times smoother than any reflective band and what a 100 m TIRS band resampled to 30 m looks like.

Radiometry. Values are unscaled digital numbers, with no published scale or offset and no stated processing level, so they are declared digital_number and unknown rather than as reflectance. 55537 is every band's exact maximum, a clip of bright ground rather than an absence of data, so it is left in the pixels and sat_frac records how much of each tile sits at it. No CRS is written: the tile ids are grid indices and the release publishes no geotransform.

Splits. The authors' split is temporal, not spatial: 432 records in their training half, covering 2015/2018, and 216 in their held-out half, covering 2022, over the same 18 of 18 tiles. A number quoted off this test split measures generalisation across years, not across ground; the split is kept verbatim as the authors' protocol. val is carved out of train by tile (96 records over 4 tiles, leaving 336), so train and val are ground-disjoint. With a test half defined over every tile, no carve separates val from test as well.

Getting started

git clone --recursive https://github.com/OscarPellicer/taco
pip install -e "taco/python[ml]"     # builds the reader: C++23, CMake, Ninja, pkg-config, libcurl >= 7.83, OpenSSL >= 3

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

Metadata without decoding anything:

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

Samples

What a sample contains

role	slot	holds	modality	detail
input	image	raster	multispectral	8 band(s), digital_number
input	questions	text_sequence	text
target	answers	text_sequence	text

Licence

Apache-2.0, public-domain

Terms of use:

• Dialogues and annotations: Apache-2.0; Landsat-8 imagery: U.S. public domain.
• A copy of the Apache-2.0 licence must accompany redistributed copies of the annotations.

Required credits:

• EarthDial (Soni et al.), MBZUAI and IBM Research
• Landsat-8 imagery courtesy of the U.S. Geological Survey

Providers: EarthDial authors (MBZUAI / IBM), USGS / NASA (Landsat-8)

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.