This dataset follows the TACO specification.

DynamicEarthNet-video: Daily PlanetFusion Image Cubes Compressed as Videos

Description

📦 Dataset

DynamicEarthNet-video is a storage-efficient re-packaging of the original DynamicEarthNet collection. The archive covers seventy-five 1024 × 1024 px regions (≈ 3 m GSD) across the globe, sampled daily from 1 January 2018 to 31 December 2019. Each day is delivered as four-band PlanetFusion surface-reflectance images (B04 Red, B03 Green, B02 Blue, B8A Narrow-NIR). Monthly pixel-wise labels annotate seven land-cover classes: impervious, agriculture, forest, wetlands, bare soil, water and snow/ice.

All original GeoTIFF stacks (≈ 525 GB) are transcoded with xarrayvideo to 12-bit H.265/HEVC, yielding dramatic size savings while preserving scientific fidelity:

Version	Size	PSNR	Ratio
Raw GeoTIFF	525 GB	—	1 ×
DynamicEarthNet-video	8.5 GB	60.1 dB	62 ×
Extra-compressed (optional)	2.1 GB	54 dB	249 ×

Extensive tests show that semantic change-segmentation scores obtained with U-TAE, U-ConvLSTM and 3D-UNet remain statistically unchanged (Δ mIoU ≤ 0.02 pp) when the compressed cubes replace the raw imagery.

The compact video format therefore removes I/O bottlenecks and enables:

• end-to-end training of sequence models directly from disk,
• rapid experimentation on 4-band daily time-series,
• efficient sharing of benchmarks for change detection and forecasting.

🛰️ Sensors

Instrument	Platform	Bands	Native GSD	Role
PlanetFusion	PlanetScope / SkySat fusion	RGB + NIR	3 m	Daily image sequence

👤 Creators

Name	Affiliation
Achraf Toker	Technical University of Munich (TUM)
Lisa Kondmann	TUM
Manuel Weber	TUM
Martin Eisenberger	TUM
Alfonso Camero	TUM
Jing Hu	TUM
André Pregel Höderlein	TUM
Çagatay Şenaras	Planet Labs PBC
Tyler Davis	Planet Labs PBC
Daniel Cremers	TUM
Guido Marchisio	Planet Labs PBC
Xiao Xiang Zhu	German Aerospace Center (DLR) / TUM
Laura Leal-Taixé	TUM

📂 Original dataset

Download (TUM Mediatum): https://mediatum.ub.tum.de/1650201

🌮 Taco dataset

⚡ Reproducible Example

import tacoreader
import xarrayvideo as xav
import xarray as xr
import matplotlib.pyplot as plt

# Load tacos
table = tacoreader.load("tacofoundation:dynamicearthnet-video")

# Read a sample row
idx = 0
row = dataset.read(idx)
row_id = dataset.iloc[idx]["tortilla:id"]

🛰️ Sensor Information

Sensors: planet

🎯 Task

• Semantic change detection and land-cover mapping on daily 4-band sequences.
• Benchmarks include U-TAE, U-ConvLSTM, 3D-UNet (official splits A/B/C) .
• DynamicEarthNet-video can also serve for next-frame prediction and self-supervised representation learning on high-frequency optical data.

📚 References

Publication 01

• DOI: 10.48550/arXiv.2203.12560
• Summary: Toker et al. introduce DynamicEarthNet, a benchmark of 75 daily 4-band PlanetFusion image cubes (3 m, 2018-2019) with monthly 7-class land-cover masks for semantic‐change segmentation. The paper establishes U-TAE, U-ConvLSTM and 3D-UNet baselines and proposes spatially blocked cross-validation to limit autocorrelation. ([arXiv][1])
• BibTeX Citation

@inproceedings{toker2022dynamicearthnet,
  title     = {DynamicEarthNet: Daily Multi-Spectral Satellite Dataset for Semantic Change Segmentation},
  author    = {Toker, Aykut and Kondmann, Leonie and Weber, Markus and Eisenberger, Marvin and Camero, Alejandro and others},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year      = {2022},
  doi       = {10.48550/arXiv.2203.12560}
}

💬 Discussion

Chat with the maintainers: https://huggingface.co/datasets/tacofoundation/DynamicEarthNet-video/discussions

🤝 Data Providers

Name	Role	URL
Planet Labs PBC	Imagery provider	https://www.planet.com

👥 Curators

Name	Organization	URL
Oscar J. Pellicer-Valero	Image Signal Processing (ISP)	Google Scholar
Cesar Aybar	Image Signal Processing (ISP)	Google Scholar
Julio Contreras	Image Signal Processing (ISP)	GitHub