ruzcko/PEAK-Dataset

PEAK: Precipitation over East Asia and Korea

PEAK is a multi-source precipitation dataset and benchmark for East Asia and Korea. It brings together geostationary satellite imagery (GK2A), Korean radar (HSP, HSR), satellite precipitation (IMERG), rain gauges (AWS), lightning and terrain on documented grids with exact UTC timestamps and validity masks. Each product keeps its own physical units and native time step.

PEAK defines three benchmark tasks: precipitation nowcasting, satellite precipitation retrieval and precipitation downscaling (super-resolution).

The code, task definitions and docs are in ruzcko/PEAK-Dataset, which is not public yet.

Release

The original release (2026-10-11) is pinned at revision 6206d9169ae29b888f541837ec71bebdceca42f2. It holds 1,373 monthly inputs (about 584 GB in 53,307 files), compact metadata, statistics, daily browser views and the five catalogue 1.0.0 task lists. The observations at that revision never change; this card and the experimental/ folder may be updated.

Catalogue 1.0.0 is the benchmark used in the PEAK paper. Later catalogues add new task views alongside it and leave it unchanged.

The Hub table preview shows catalogue rows. Dense observations are read with the PEAK package.

Regions and grids

View	Grid	Coverage
KR fine	320 × 320 at 2 km	Radar-covered Korea, 640 × 640 km
KR coarse	64 × 64 at 10 km	Same extent; each coarse cell holds 5 × 5 fine cells
EA	512 × 512 at 10 km	East Asia, 5,120 × 5,120 km
KE	512 × 512 at 2 km	Extended Korea, for station points only

All grids use the GK2A Lambert conformal projection. Grid spacing describes the processed product, not the sensor's native resolution. Radar and gauges cover Korea only.

Products

Product	Views	Quantity	Time step	Coverage
GK2A IR105, IR123, WV063, WV073	KR, EA	Brightness temperature, K	10 min	Jul 2019 – Dec 2025
HSP (KMA radar)	KR fine	Rain rate, mm/h	10 min	Jan 2021 – Dec 2025
HSR (KMA radar)	KR fine	Rain rate, mm/h	10 min	Jul 2019 – Dec 2025
IMERG Final V07B	KR coarse, EA	Precipitation rate, mm/h	30 min	Jul 2019 – Sep 2025
AWS gauges	KR, KE points	Rate, mm/h; accumulation, mm	10/30/60-min windows	Jul 2019 – Dec 2025
Lightning	KR	Events and a 10-min stroke-count grid	Event time / 10 min	Jul 2019 – Jul 2025
DEM	KR (SRTM), EA (Copernicus GLO-90)	Elevation, m	Static	—

IMERG V07 Final ends in September 2025, and six IMERG monthly files for October–December 2025 are absent. Coverage dates are calendar ranges; exact frame availability is in the manifest and metadata bundle.

User-derived products. GSMaP (GSMaP_Gauge, hourly) and MSWEP are not distributed here. PEAK's processing code turns your own copy of the source files into the same grids and format:

peak process precipitation --product GSMaP --region EA-ROI --year-month 202307 \
  --source-root /path/to/gsmap_gauge --data-root ./peak-data

The PEAK paper uses GSMaP, derived this way, in its product comparison and as an optional input. MSWEP is supported in code only and is not part of the paper.

Format and missing data

Gridded products are monthly Zarr v3 stores (one per product, region, channel and month) with named (time, y, x) dimensions, coordinates and masks. Values are stored as scaled integers with units and scale/offset metadata; the PEAK readers return physical values. Stations, lightning events and catalogues are Parquet. Both are standard formats, so xarray, zarr-python and pandas read them directly.

• Frames. A frame is present only if its exact UTC slot was observed. Missing frames are never filled from neighbouring times.
• Pixels. Use the stored validity mask. A zero value or fill value on its own does not mean missing.
• HSP benchmark mask. Benchmarks use frame availability plus the static radar coverage, so dry zeros inside coverage count. The stored per-frame HSP mask excludes dry zeros and is not the evaluation mask.
• HSR cap. 655.3 mm/h is a valid, capped rain rate.
• Time labels. Join products in UTC. IMERG labels the start of its 30-minute interval (09:30 is [09:30, 10:00)). The offset-8 AWS window extends past its label, so take care with it in forecasting.

Benchmark tasks (catalogue 1.0.0)

Five task variants, 502,053 samples in total. Samples may overlap in time within a split. No rainfall threshold is used to select samples.

Task ID	Inputs → target	Train	Validation	Test
hsp-nowcasting-all-v1	12 HSP frames (T−110 … T) → 12 frames (T+10 … T+120)	25,778	3,891	4,290
kr-multimodal-retrieval-v1	GK2A 4 channels + DEM at T → HSP at T	25,661	8,573	8,132
ea-multimodal-retrieval-v1	GK2A 4 channels at T, T+10, T+20 + DEM → IMERG [T, T+30)	50,847	17,056	12,759
imerg-hsp-downscaling-v1	IMERG 10 km → mean of HSP at T, T+10, T+20	52,417	17,245	13,065
hsp-coarsening-downscaling-v1	HSP coarsened 5 × 5 → HSP 2 km	157,517	52,362	52,460

Splits are UTC years: train 2021–2023, validation 2024, test 2025. Every input and target of a sample falls inside its split; 2019–2020 data are outside the benchmark. Sample spacing is 60 min in nowcasting training and 120 min in its validation and test, 60 min for Korea retrieval, 30 min for East Asia retrieval and IMERG-to-HSP, and 10 min for coarsened HSP.

reduced_train marks a fixed 10% training subset: in each training month, floor(0.10 × eligible samples) chosen by SHA-256 of 42:sample_id.

Task notes:

• Satellite precipitation retrieval keeps the same samples and the full-input evaluation mask when you select fewer input channels, so scores stay comparable across input choices.
• The IMERG-to-HSP target is the mean of three HSP rates over the IMERG interval, not a native 30-minute accumulation.
• HSP coarsening needs at least 24 valid fine cells per coarse cell.

Each catalogue row has sample_id, split, reference_time_utc and reduced_train. On the Hub, the catalogue split shows the whole table; the split column holds train, validation or test. The YAML task definitions in PEAK-Dataset specify timing, operators, grids and masks, and a membership hash identifies each catalogue.

Getting started

Install the PEAK package (Python 3.12 for the Zarr and release tools):

python -m pip install 'peak-dataset[zarr,catalogues,release,mcp] @ git+https://github.com/ruzcko/PEAK-Dataset.git'

Point the tools at the pinned release:

export PEAK_RELEASE_MANIFEST=https://huggingface.co/datasets/ruzcko/PEAK-Dataset/resolve/6206d9169ae29b888f541837ec71bebdceca42f2/release.json

# Metadata and static files only (no dense observations)
peak download --release "$PEAK_RELEASE_MANIFEST" --metadata-only \
  --output-root ./peak-release --fetch

# Rebuild the five catalogues from metadata and check them against the release
peak catalogues build --metadata ./peak-release/metadata/inventory \
  --definitions ./peak-release/metadata/definitions --output ./rebuilt-catalogues
peak catalogues verify --catalogues ./rebuilt-catalogues \
  --definitions ./peak-release/metadata/definitions \
  --metadata ./peak-release/metadata/inventory \
  --reference ./peak-release/metadata/catalogues

# Plan a download of selected monthly stores (add --fetch to download)
peak download --release "$PEAK_RELEASE_MANIFEST" --product HSP --region KR-ROI --months 202308

Load one benchmark sample straight from the Hub:

import os
from peak.tasks import load_sample

sample = load_sample(
    "kr-multimodal-retrieval-v1:20230801T0000Z",
    definitions="./peak-release/metadata/definitions",
    release=os.environ["PEAK_RELEASE_MANIFEST"],
    inputs=["kr-ir105"],
)
print(sample.sample_id, sample.inputs.keys(), sample.evaluation_mask.shape)

load_sample returns physical-unit NumPy inputs and targets, source masks, the evaluation mask and provenance. Downloads resume and verify checksums. For a local copy, pass data_root instead of release.

You can write your own task YAML (sources, dates, windows, strides, splits, subsets). Give it a new task ID so results are not confused with the benchmark.

Experimental extensions

The experimental/ folder holds products outside the original release and its catalogues, each with its own README, inventory and checksums.

• experimental/hsp500-0.1.0/: KMA HSP rebuilt from raw files at 500 m over the extended Korea (KE) region, 2021–2025, 60 monthly HDF5 files (64.9 GB).

Sources and terms

The PEAK code is MIT-licensed. The data keep the terms of their providers; license: other means "see this section". Please cite the source products you use.

Product	Source	Terms
GK2A, HSP, HSR, AWS, lightning, radar coverage	Korea Meteorological Administration (KMA)	Redistributed with KMA approval for this project; credit KMA
IMERG	NASA GPM IMERG Final V07B, doi:10.5067/GPM/IMERG/3B-HH/07	NASA GPM data policy
EA DEM	Copernicus DEM GLO-90 (2021), doi:10.5270/ESA-c5d3d65	Copernicus DEM terms
KR DEM	USGS SRTM 1 Arc-Second Global	USGS SRTM, public domain
GSMaP (user-derived)	JAXA GSMaP_Gauge, hourly	JAXA terms

Copernicus notice for the EA DEM:

produced using Copernicus WorldDEM-90 © DLR e.V. 2010-2014 and © Airbus Defence and Space GmbH 2014-2018 provided under COPERNICUS by the European Union and ESA; all rights reserved

Processing in brief: radar and GK2A keep their projected KMA lineage; IMERG and GSMaP are conservatively remapped (at least 95% valid overlap per cell, no extrapolation); AWS rates come from QC'd one-minute increments; lightning keeps event IDs and signed intensity, with only exact duplicates removed; the EA DEM is a surface model averaged onto the EA grid.

Using PEAK well

• IMERG is the target for East Asia retrieval. Matching IMERG is not the same as matching true rainfall; radar targets exist for Korea only.
• Keep masks, units, time windows and splits fixed when comparing models, and fit normalization on training data only.
• Dropping an input at test time is not the same as retraining without it.
• PEAK is a research benchmark, not an operational forecasting or warning system.

Citation

Tobias, R. R. PEAK: Precipitation over East Asia and Korea. Hugging Face dataset, 2026. https://huggingface.co/datasets/ruzcko/PEAK-Dataset

When reporting results, record the data revision, code commit, task ID, catalogue membership hash, the inputs you used and the evaluation mask.

Contact

Rogelio Ruzcko Tobias (ruzcko@dilab.kr). Questions and issue reports are welcome by email or in the Community tab.