isp-uv-es/ebd-taco

EBD

This is a repackaging, not a new dataset. It is EBD by Zeyu Wang, Feng Zhang, Chuyi Wu (Zhejiang University), Maxar Open Data Program, 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 · licence: CC-BY-NC-4.0, 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:

@misc{wang2024ebd,
  author    = {Wang, Zeyu and Zhang, Feng and Wu, Chuyi},
  title     = {An Extensible Building Damage (EBD) dataset constructed from disaster-related bi-temporal remote sensing images},
  publisher = {figshare},
  year      = {2024},
  doi       = {10.6084/m9.figshare.25285009.v2}
}

About the data

18215 bi-temporal 512x512 px pre/post-disaster pairs over 12 events, each with two rasters of truth: a binary building-localisation mask on the pre-event frame and a five-level damage mask on the post-event one. EBD is "Extensible" Building Damage, not "Earthquake": one of the 12 events is an earthquake.

18,215 samples · splits: test 3,856 · train 12,545 · validation 1,814 · tasks: semantic-segmentation

Packaged as TACO v3.

Full description

Labels. Machine-produced by a model pre-trained on xBD and fine-tuned per event under manual supervision, a pipeline meant to be re-run on the next disaster.

• Both legends are measured over all 36,430 masks: every pre mask is a subset of {0, 255} and every post mask a subset of {0, 1, 2, 3, 4}, with no file mixing the ranges.
• The localisation mask is remapped 255->1 and declared ml:binary. The damage mask keeps the release's own codes, where 0 is background rather than a grade; no ignore_index is set, the release publishing no unclassified level.
• 5,955 pairs are empty on both mask sides, and they are the same 5,955: a tile with no buildings has no damage either, which shifts any mean IoU over the whole set.

Encoding. Imagery is Maxar Open Data 8-bit natural colour, copied byte-for-byte; a lossless COG measures 1.42x the source PNG on 512 px tiles with no georeferencing to earn it back. Masks are re-encoded as COG at 0.96x. No CRS, geotransform or GSD is published, so none is written, and radiometry is requantised with no approx_scale.

Splits. The release publishes none. split is a constructed partition that keeps each event whole (test: HURRICANE-IDA, HURRICANE-LAURA, TEXAS-TORNADOES; validation: HURRICANE-IRMA, MOUNT-SEMERU-ERUPTION, TONGA-VOLCANO; the other six are train), and event is the column to group on. The events differ in size by more than an order of magnitude. Pairs per event: HURRICANE-IAN 5641, PAKISTAN-FLOODING 3540, HURRICANE-IDA 2638, HURRICANE-DELTA 1364, EARTHQUAKE-TURKEY 944, HURRICANE-IRMA 788, HURRICANE-LAURA 764, MOUNT-SEMERU-ERUPTION 592, HURRICANE-DORIAN 540, STVINCENT-VOLCANO 516, TEXAS-TORNADOES 454, TONGA-VOLCANO 434.

Overlap. RSCC (BiliSakura/RSCC) carries a change caption for one 512x512 crop-pair of every pair here. Filtered to image_pool == 'EBD', it is the caption layer over exactly these samples, and the two are not independent data.

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:

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

path = hf_hub_download("isp-uv-es/ebd-taco", "ebd.zip", repo_type="dataset")
ds = Dataset(path)
plot_sample(ds[0])

or from a local copy:

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

Metadata without decoding anything:

import taco
taco.read("ebd.zip")              # one Arrow table, levels joined

Samples

What a sample contains

role	slot	holds	modality	detail
input	pre	raster	optical	3 band(s), requantised
input	post	raster	optical	3 band(s), requantised
target	buildings	mask	label_raster	2 classes
target	damage	mask	label_raster	5 classes

Licence

CC-BY-NC-4.0, CC-BY-4.0

Terms of use:

• Non-commercial use only (CC-BY-NC-4.0).
• Credit the original authors and the data sources listed below.
• Imagery (Maxar Open Data Program): CC-BY-NC-4.0; annotations: CC-BY-4.0.

Required credits:

• EBD (Wang, Zhang, Wu), Zhejiang University, doi:10.6084/m9.figshare.25285009
• Imagery: Maxar (now Vantor) Open Data Program, licensed under CC BY-NC 4.0

Providers: Zeyu Wang, Feng Zhang, Chuyi Wu (Zhejiang University), Maxar Open Data Program

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.