ruzcko/PEAK-Dataset

PEAK

Public release — 2026-10-11. This repository hosts 1,373 cleared inputs, five frozen catalogue 1.0.0 variants, compact metadata, statistics and daily browser views. The project name and expansion remain provisional.

Use the immutable release manifest at 6206d9169ae29b888f541837ec71bebdceca42f2. Its inventory contains 583,628,811,225 bytes and 53,303 objects; including the manifest, card, source decisions and Hub bookkeeping, that revision contains 583,647,787,208 bytes across 53,307 files. All inventory object sizes and identities were verified. The card may receive documentation updates; this observation pin remains immutable.

Published products are GK2A's four IR/WV channels, HSP, HSR, IMERG, AWS, lightning, both DEMs and the radar mask. GSMaP's 156 inputs and derived statistics are excluded while source records remain unresolved. Six known missing IMERG paths remain explicitly absent. All five catalogue variants have published observation dependencies. Cloudflare deployment and public GitHub code access remain separate steps; the two code repositories are still private.

Dataset summary

PEAK organizes precipitation-related observations and derived products for East Asia and Korea. It supports precipitation nowcasting, retrieval and downscaling through documented grids, exact UTC timestamps and validity masks. Products retain their own physical quantities and native temporal resolution.

The release is designed for three uses: inspect observations and statistics in the project explorer, download selected monthly stores, and reproduce or customize task catalogues. Reproducing a catalogue and loading its samples are implemented; reproducing model training and reported model scores is later work.

The project page also works before observations are connected: it presents documented internal coverage, the five frozen benchmark definitions/counts, task exclusions, access examples and source-documentation status. Pixel statistics and frames require a completed release manifest. See project-page setup.

The Hub table preview shows catalogue rows. Dense observations are accessed through the PEAK package and explorer. The five frozen catalogue variants have published dependencies in this revision. Custom catalogue dependencies must be checked against the manifest.

• Python package and access documentation
• Dataset documentation, definitions and explorer
• Cloudflare project-page deployment is pending. The manifest above records the release identity, actual coverage and source publication decisions.

Domains and observation coverage

View	PEAK grid	Spatial support
KR-ROI fine	320 × 320 at 2 km	Radar-supported Korean core, 640 × 640 km
KR-ROI coarse	64 × 64 at 10 km	Same extent; each coarse cell contains exactly 5 × 5 fine cells
EA-ROI	512 × 512 at 10 km	Satellite and gridded-product view, 5,120 × 5,120 km
KE-ROI	512 × 512 coordinate grid at 2 km	Extended station-point view only, 1,024 × 1,024 km

Coordinates use the GK2A Lambert conformal projection. PEAK grid spacing describes the processed representation, not necessarily native sensor resolution. KMA radar and AWS do not provide EA-wide supervision. Station row/column mappings locate points; they are not gridded rainfall targets.

Product	View	Quantity / units	Native cadence or label	Calendar coverage (GSMaP is unpublished)
GK2A IR105, IR123, WV063, WV073	KR, EA	Brightness temperature, K	10 min	Jul 2019–Dec 2025
HSP	KR fine	Radar rainfall rate, mm/h	10 min	Jan 2021–Dec 2025
HSR	KR fine	Separate radar rainfall rate, mm/h	10 min	Jul 2019–Dec 2025
IMERG	KR coarse, EA	Precipitation rate, mm/h	30 min	Jul 2019–Sep 2025
GSMaP	KR coarse, EA	Precipitation rate, mm/h	60 min	Internal: Jul 2019–Dec 2025; unpublished
AWS	KR, KE points	Rate, mm/h; accumulation, mm	10-min labels; 10/30/60-min windows and offset-0/offset-8 variants	Jul 2019–Dec 2025
Lightning events and derived count grid	KR	Event fields, including signed intensity in kA; grid counts retained strokes	Event timestamps / 10-min grid intervals	Jul 2019–Jul 2025
DEM	KR, EA	Elevation, m	Static	Source-specific

Calendar endpoints do not establish complete frames, pixels, station operation or detector uptime. Six IMERG monthly paths for KR/EA October–December 2025 are absent. Exact frame availability and source identities belong to the versioned manifest and compact metadata bundle. Counts of frames, station rows, events and task samples are different measures.

Visible GK2A channels and ERA5 are planned. MSWEP arrays and diagnostic HSR reflectivity are excluded from distribution.

Storage, values and missingness

Public gridded observations are prepared as monthly Zarr v3 directories, mirroring product, channel and month paths. The experimental container contract is peak-zarr 0.2.0; original 0.1.0 pilot stores remain readable. Scientific HDF5 v1 values, dtypes, timestamps, coordinates, source identities and processing versions are preserved. Container-conversion provenance is recorded separately. Validated HDF5 remains the internal reference.

Stations, lightning events and catalogues use Parquet. Verified daily station/event views add projected display coordinates while retaining canonical columns and provenance. ZIP is not the primary observation format.

Each gridded store exposes its scientific variable, coordinates and masks at the root, with named dimensions for Xarray. Grids use (time, y, x) or (y, x) for static arrays. Encoded values carry units and scale/offset metadata. PEAK readers return physical values and apply scale/offset once. Source basenames use fixed-width UTF-8 byte arrays with length/dtype metadata.

Use explicit validity information:

• Available frame: its exact native UTC slot is present and valid. A missing frame is not replaced by a nearby frame or a repeated coarse-product label.
• Valid pixel: use the product's stored mask and the task's declared policy. Encoded zero or the container fill value alone does not identify missingness.
• HSP benchmark convention: use frame availability and shared static radar coverage, retaining dry zeros within coverage. The frozen stored HSP mask excludes dry zeros and is not the benchmark evaluation mask. This convention cannot recover unknown raw-source missingness.
• Capped HSR: 655.3 mm/h remains valid capped rainfall. It is neither a missing-data code nor an uncapped estimate of an extreme rate.
• Stations/events: retain per-window QC, valid counts, endpoint variants and event identities. An exported day with no lightning events does not establish detector uptime.

Use UTC to join products. IMERG labels the start of its 30-minute interval; for example, 09:30 represents [09:30, 10:00). Rates are not accumulated depths. The offset-8 AWS measurement window extends beyond its nominal label and requires care in causal forecasting.

The explorer uses prepared availability, distributions, spatial summaries, station QC and task-exclusion statistics. Every summary records units, mask policy and denominators. HSP stored-mask and benchmark-coverage statistics are separate. Source or observation unavailability and unpublished components are reported separately.

Frozen benchmark catalogues

Five catalogue 1.0.0 variants contain 502,053 task samples in total. Samples can overlap within a split; this is not a count of independent weather events. No rainfall threshold selects these catalogues.

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

Catalogue spacing differs from native cadence. Nowcasting reference times are 60 minutes apart in training and 120 minutes apart in validation/test. KR retrieval uses 60 minutes; EA retrieval and IMERG-to-HSP use 30 minutes; coarsened-HSP uses 10 minutes.

Splits are UTC half-open intervals: training [2021-01-01, 2024-01-01), validation [2024-01-01, 2025-01-01) and test [2025-01-01, 2026-01-01). All required inputs and targets stay inside their split. The available 2019–2020 observations are outside these benchmark splits.

The smaller training subset keeps floor(0.10 × eligible samples) within each UTC training month, ranked by SHA-256 of 42:sample_id, with sample ID as the tie-breaker. Validation/test membership is unchanged.

Important task distinctions:

• Retrieval input selections retain the original sample membership and full-input evaluation mask, including masks for omitted channels and DEM.
• The three-HSP-frame mean is a task-defined rate approximation over the common interval, not a native HSP 30-minute accumulation. For this task only, each point-analysis rate represents the following 10 minutes.
• Synthetic HSP coarsening requires at least 24 valid fine cells per coarse cell. It is distinct from downscaling independently measured IMERG.
• AWS, lightning, HSR and GSMaP belong to the internal foundation but are not required by these five initial definitions. Additional inputs require their own named task definition.

Each catalogue table has sample_id, split, reference_time_utc and reduced_train. The Hub configuration's catalogue split displays the complete table; its split column remains the scientific train/validation/test assignment. YAML definitions specify timing, operators, grids and masks. Membership hashes verify regenerated scientific identities; file checksums separately verify the supplied Parquet bytes.

The older HSP experiment lists and catalogue 0.3.1 remain separate. These five catalogues do not establish membership equivalence with previous paper results. Model scores require an explicit sample/evaluation agreement check.

Access and reproduction

These commands use the completed public release. Installing the package currently requires private GitHub access. Use Python 3.12 for the Zarr/release tools.

python -m pip install 'peak-dataset[zarr,catalogues,release,mcp] @ git+https://github.com/ruzcko/PEAK.git@a620c81d2466a4740e562702c223026a8c7647e2'

Set PEAK_RELEASE_MANIFEST to a local completed release.json or an HTTP URL. HF URLs must use a completed 40-character commit, not a mutable branch or tag.

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

# Plan a compact metadata download without fetching data.
peak download --release "$PEAK_RELEASE_MANIFEST" --metadata-only

# Fetch and verify the metadata/static-dependency bundle.
peak download --release "$PEAK_RELEASE_MANIFEST" --metadata-only \
  --output-root ./peak-release --fetch

# Rebuild all five lists without dense observations.
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 selected complete monthly stores and their static dependencies.
peak download --release "$PEAK_RELEASE_MANIFEST" --product HSP --region KR-ROI --months 202308

The compact bundle contains exact frame availability and the static assets needed for reproduction, published under their recorded source terms. Downloading the bundle does not download dense observations. Complete-store downloads resume and verify checksums; online observation loading remains selective.

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)

The loader returns decoded NumPy inputs/targets, source masks, the declared evaluation mask and provenance. The IR105-only example still needs every declared input's mask. Missing or unpublished dependencies fail explicitly. For a local strict-v1 archive, replace release with an explicit data_root.

Custom YAML definitions support explicit sources, dates, temporal windows, strides, splits and deterministic subsets. Give them separate task identities and report their actual parameters. Rainfall-filtered catalogues and model training reproduction follow later.

The local read-only metadata MCP server is optional:

peak mcp serve --release ./peak-release/release.json \
  --definitions ./peak-release/metadata/definitions

Sources, processing and data terms

The source-code repositories use MIT. That license does not cover third-party observations. The card's license: other label is not a grant of data rights. Cleared component records must identify exact source terms and attribution before publication. Unresolved products remain unavailable.

Component	Preserved source identity	Publication record
GK2A, HSP, HSR, AWS, lightning and radar coverage	Product-specific KMA sources and established projected/point lineages	Maintainer confirmed project redistribution approval on 2026-10-10; retain applicable source/project terms and attribution
IMERG	GPM Final 3IMERGH V07B; Version 07 DOI	Eligible under NASA's GPM policy; retain exact-version citation and PEAK processing provenance
GSMaP	Gauge-adjusted hourly GSMaP_Gauge; exact acquired algorithm version unresolved	JAXA terms allow credited distribution subject to their conditions; acquired version and wide online distribution contact record remain pending
EA DEM	Copernicus COP-DEM_GLO-90-DGED, 2021 COG distribution; source DOI	Eligible under GLO-90 terms; retain source notice and exact derived-product attribution
KR DEM	Maintainer-confirmed SRTM source; exact 1 Arc-Second Global v3 product/version inferred from 1,439 mosaic tile names; distinct from EA DEM	Eligible under the USGS SRTM open-distribution record; retain source attribution and version qualification. Dated evidence

The publication decision record records the maintainer's confirmation for the current satellite, radar, station, lightning and derived-mask components. Prior paper redistribution was noted as supporting context. This confirmation does not assign a blanket public license to other products. Separate NASA GPM and Copernicus GLO-90 reviews establish IMERG/EA DEM eligibility under their published terms. The maintainer separately confirmed KR terrain's SRTM source. Together, 12 of 14 prepared product/region groups, comprising 1,373 inputs, are eligible with the recorded attribution and conditions. Only the 156 GSMaP months remain pending. These decisions leave completed conversion identities unchanged. The completed public manifest records the 1,373 uploaded inputs; GSMaP remains unavailable.

Radar v1 preserves the projected monthly lineage; this release is not a new raw-radar reconstruction. GK2A preserves already calibrated/projected IR/WV observations and does not claim a newly verified L1B calibration/navigation run. IMERG and GSMaP use valid-area-normalized conservative remapping, requiring at least 95% valid target-cell overlap and no extrapolation into uncovered areas. They remain separate products.

AWS rates/accumulations derive from valid one-minute RNDAY increments, normalized to nominal windows after minimum-count QC. They are not copies of native RN60m. Lightning retains deterministic event IDs, signed intensity, sensor fields and recorded filtering; only exact duplicate rows are removed.

EA DEM is a digital surface model, average-resampled to the EA grid, with WGS 84 horizontal and EGM2008 vertical references. Valid zero-filled ocean and separate display water masking are retained. Required derived-product notice:

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

See source identities, processing and terms for provider links and audited scope. Upstream DOIs identify the source products, not PEAK. The completed release manifest defines actual distributed components.

Intended use and limitations

PEAK supports research on precipitation modelling, multimodal learning and evaluation with explicit source and task definitions. Product disagreement, geographic coverage, temporal gaps and spatial validity affect interpretation. An available frame is not evidence of valid observations at every pixel.

IMERG is an evaluation reference in the EA retrieval task; agreement with it does not independently establish true-rainfall accuracy. Radar supervision is limited to KR. Station membership changes over time. Exported lightning calendar coverage does not establish detector uptime. Valid capped HSR values limit interpretation of extreme rates. HSP's frozen benchmark convention cannot recover unknown raw-source missingness.

Keep masks, QC, units, temporal windows and split identities when comparing models. Fit normalization and learned preprocessing on training data only. Removing inputs at inference is distinct from retraining a modality ablation. This card and catalogue release establish neither model performance nor operational forecast/warning capability.

Citation and maintenance

Use CITATION.cff and record the immutable data/code revisions, task definition, catalogue membership hash, inputs, operators and evaluation mask used. A PEAK paper link and dataset DOI are pending; they must be added only when assigned. Cite and acknowledge the source products used in the experiment.

Maintainer: Rogelio Ruzcko Tobias. Contact: ruzcko@dilab.kr. Report data/catalogue/viewer issues in PEAK-Dataset and Python reader/converter/download issues in PEAK. Include the release/code revisions, product/view/channel, exact UTC timestamp or sample ID, and expected versus observed behavior.