Catalogue 1.1 · 8 downstream tasks
PEAK
A Multimodal Dataset for Precipitation Deep Learning Applications over East Asia and Korea
What PEAK is
PEAK brings GK2A satellite imagery, the IMERG precipitation product and terrain over East Asia together with Korean radar rainfall, rain gauges and lightning over a nested Korean domain. Each source keeps its own observation times, validity masks and processing record, and versioned task indexes turn the same observations into reproducible samples for precipitation nowcasting, satellite precipitation retrieval and precipitation downscaling.
Dataset at a glance
| Source | Domain | Quantity | Native time step | Coverage |
|---|---|---|---|---|
| GK2A IR105 | EAKR | Brightness temperature, K (10.5 µm, clean infrared window) | 10 min | Jul 2019–Dec 2025 |
| GK2A IR123 | EAKR | Brightness temperature, K (12.3 µm, infrared window with some water vapour) | 10 min | Jul 2019–Dec 2025 |
| GK2A WV063 | EAKR | Brightness temperature, K (6.3 µm, upper-level water vapour) | 10 min | Jul 2019–Dec 2025 |
| GK2A WV073 | EAKR | Brightness temperature, K (7.3 µm, mid-level water vapour) | 10 min | Jul 2019–Dec 2025 |
| IMERG V07B | EAKR | Precipitation rate, mm/h | 30 min | Jul 2019–Sep 2025 |
| HSP radar | KR | Radar rain rate, mm/h | 10 min | Jan 2021–Dec 2025 |
| HSR radar | KR | Rain rate from reflectivity, mm/h | 10 min | Jul 2019–Dec 2025 |
| AWS gauges | KR | Rate, mm/h; accumulation, mm | 10 min | Jul 2019–Dec 2025 |
| Lightning | KR | Stroke events and count grid | events | Jul 2019–Dec 2025 |
| Terrain (DEM) | EAKR | Elevation, m | static | – |
East Asian domain (EA-ROI): 512 × 512 cells at 10 km. Korean domain (KR-ROI): 320 × 320 cells at 2 km, and 64 × 64 at 10 km.
Every source at one time
Frames present, month by month
over 99.5% 97–99.5% 90–97% under 90% no data
—
Downstream tasks

2025-07-17 06:00 UTC · T+10 min

2025-07-16 20:00 UTC

2025-07-17 08:30–09:00 UTC
Samples in train / validation / test: 2021–2023 / 2024 / 2025 (UTC). Rainy: at least 10% of target pixels at 0.1 mm/h or more, or a 99th percentile above 10 mm/h. Heavy: at least 0.1% at 20 mm/h or more.
Quickstart
Python 3.12. Samples stream from the pinned release on Hugging Face. Install works once the code is public
python -m pip install "peak-dataset[release] @ git+https://github.com/ruzcko/PEAK-Dataset.git"
hf download ruzcko/PEAK-Dataset --repo-type dataset \
--revision 377f2e8646e766d1fc5bd6d8992af6a6ce8b1aad \
--include "metadata/catalogue-1.1/*" --local-dir ./peak
from peak.tasks import load_sample, samples
from peak.zarr import read_sequence
hf = "https://huggingface.co/datasets/ruzcko/PEAK-Dataset/resolve/377f2e8646e766d1fc5bd6d8992af6a6ce8b1aad"
defs = "./peak/metadata/catalogue-1.1"
sample_id = samples("hsp-nowcasting-v1", definitions=defs, split="train")[0]
s = load_sample(sample_id, definitions=defs, release=f"{hf}/release-1.1.json")
print(s.inputs["kr-hsp"].shape, s.targets["kr-hsp"].shape) # (6, 320, 320) (12, 320, 320)
# One radar frame by UTC time
frame = read_sequence(f"{hf}/data/KR-ROI/HSP/2025/HSP_KR_202507.zarr", ["202507170900"])
print(frame.rain_rate.shape) # (1, 320, 320)
Cite PEAK
BibTeX coming soon