meteolibre-dev/global_sat_metar

Global GMGSI + METAR Patches (v1)

Dense global satellite imagery (GMGSI, 4 channels, 0.1° / ~9 km, hourly) paired with sparse global METAR station observations rasterized onto the same 3600×1800 grid, sliced into 128×128 spatial patches with a 7-frame hourly temporal context. Designed as a self-supervised / supervised pre-training corpus for weather foundation models that need to jointly see geostationary satellite fields and ground-truth in-situ observations.

This dataset is generated by src/generate/generate_satellite_metar_dataset_v1.py in the meteolibre_datasetgen repository. The default 3-month chunked pipeline that produced the public release is run_satellite_metar_pipeline.sh.

---

---

Dataset summary

Each row of each *.parquet file is not a single hourly frame — it is a spatio-temporal sample:

• Space: a 128×128 patch cropped from the global 3600×1800 GMGSI / METAR grid (≈ 12.8° × 12.8° at the equator, half-overlapping along longitude at high latitudes due to the equirectangular projection).
• Time: a window of T = back_step_hours + 1 + forward_step_hours hourly frames centred on the reference timestamp, by default T = 7 (5 hours past → reference → 1 hour future), 1-hour cadence.
• Sensors: a dense 4-channel geostationary satellite field (sat_data) co-registered with a sparse 7-channel rasterized METAR field (metar_data, NaN where no station reported in the time window), and a static 1-channel elevation map (elevation_data).

The METAR field is intentionally mostly NaN: the model is meant to treat in-situ observations as a sparse sensor overlay on top of the dense satellite field rather than as a fully sampled target.

The pipeline stores binary blobs of contiguous arrays together with shape/dtype metadata, so each row is fully self-describing and the file can be loaded with a 4-line snippet (see Data instance).

Visual overview

Supported tasks

This corpus is intentionally multi-task. The two main intended uses are:

• Sat → ground nowcasting / imputation: given the dense satellite field at past timesteps, predict the sparse METAR field at the reference time (and optionally the next hour). Loss is naturally masked to valid observation pixels.
• Weather foundation model pre-training: jointly modelling satellite
  • ground sensors with masked / contrastive / diffusion objectives. The METAR sparsity makes it a natural test bed for in-painting-style objectives.

Secondary uses that the same data enables:

• 4-band → 7-band sensor translation (satellite to gridded surface analysis, also called "station-based downscaling" in the literature).
• Pre-training for downstream MTG / Meteosat fine-tunes (GMGSI is the natural global proxy for the FCI bands, with a similar longwave-IR / visible / water-vapour / shortwave-IR breakdown).
• METAR station data quality control by training a model to flag outliers vs. what the satellite field would predict.

Dataset structure

The dataset is shipped as a flat folder of *.parquet files, named:

<YYYY-MM-DD_HH-MM>_global_rn<8-hex-rand>_patches_<NNNN>.parquet

• Each file groups up to patches_per_file = 64 patches that share the same reference timestamp.
• The hex suffix is a per-run random tag used to avoid filename collisions between reruns.
• patches_<NNNN> enumerates split files when the per-series output exceeds patches_per_file.

The HF datasets library can load the whole release with:

from datasets import load_dataset
ds = load_dataset("meteolibre-dev/global_sat_metar", split="train")

Or locally:

import pyarrow.parquet as pq
table = pq.read_table("data_satellite_metar_v1/2021-07-14_00-00_global_rnXXXXXXXX_patches_0000.parquet")

Data fields

Field	Type	Shape	Dtype	Description
sat_data	bytes	—	—	Raw bytes of the satellite patch; reshape with sat_shape and sat_dtype.
sat_shape	list[int64]	(4,)	—	Array shape (T, 4, 128, 128) — T hourly frames × 4 GMGSI channels.
sat_dtype	string	—	—	NumPy dtype string, typically '<f2' (float16).
metar_data	bytes	—	—	Raw bytes of the METAR patch; reshape with metar_shape and metar_dtype.
metar_shape	list[int64]	(4,)	—	Array shape (T, N_FEATURES, 128, 128) with N_FEATURES = 7.
metar_dtype	string	—	—	NumPy dtype string, always '<f4' (float32).
epsg	int64	scalar	—	EPSG code of the patch CRS; always 4326.
x_coord	float64	scalar	—	Patch-centre projected x in CRS units (here, longitude in degrees, EPSG:4326).
y_coord	float64	scalar	—	Patch-centre projected y in CRS units (here, latitude in degrees, EPSG:4326).
lon	float64	scalar	—	Alias of x_coord (duplicate, kept for downstream convenience).
lat	float64	scalar	—	Alias of y_coord.
patch_x_idx	int64	scalar	—	Patch x-index in the global patch grid (stride 64 px).
patch_y_idx	int64	scalar	—	Patch y-index in the global patch grid.
date	string	scalar	—	Reference timestamp of the temporal window, format YYYY-MM-DD HH:MM:SS (UTC).
elevation_data	bytes (optional)	—	—	Raw bytes of the static elevation patch.
elevation_shape	list[int64]	—	—	(128, 128).
elevation_dtype	string	—	—	NumPy dtype string, '<f4' (float32). Nodata = -9999.0.

GMGSI satellite channels (sat_data)

Channels are stacked along axis 1 in this fixed order:

Axis	Name	GMGSI channel	Physical quantity
0	lw_ir	GLOBCOMPLIR	Longwave IR brightness temperature (K).
1	vis	GLOBCOMPVIS	Visible reflectance (0–1, dimensionless).
2	wv	GLOBCOMPWV	Water-vapour channel brightness temperature (K).
3	sw_ir	GLOBCOMPSIR	Shortwave IR brightness temperature (K).

METAR ground-station features (metar_data)

Channels are stacked along axis 1 in this fixed order:

Axis	Name	Unit	Definition / encoding
0	tmpc	°C	Air temperature, 2 m.
1	dwpc	°C	Dew-point temperature, 2 m.
2	mslp	hPa	Mean sea-level pressure.
3	cloud_cover	fraction ∈ [0, 1]	Encoded from METAR skyc token: CLR/SKC/NCD/NSC → 0.0, FEW → 0.2, SCT → 0.4, BKN → 0.7, OVC/VV → 1.0, unknown → NaN.
4	p01m	mm	1-hour accumulated precipitation (NaN treated as 0 for rasterization).
5	wind_u	m/s	Eastward wind component, u = -wspd * sin(wdir).
6	wind_v	m/s	Northward wind component, v = -wspd * cos(wdir).

The METAR timestamp window used to fill one frame is [reference_time - metar_window_sec, reference_time], default metar_window_sec = 3600 s (i.e. the 1-hour running window ending at the reference). When several stations fall on the same grid cell, the most recent observation in the window is kept.

Elevation (elevation_data)

• Source: data/ELE.tif (a global DEM in EPSG:4326 at 0.0083° resolution from upstream gs://eumetsat_mtg_preprocess/assets/ELE.tif).
• Reprojected bilinearly to the GMGSI 0.1° grid and patched identically to the satellite/METAR fields.
• Nodata sentinel: -9999.0 (typically appears over ocean and over the polar caps where the source DEM does not extend). May be dropped per patch by filtering on (ele == -9999.0).mean() < threshold.

Data splits

This release is not pre-split. Patches are organised by reference timestamp, so a natural split is:

• Chronological split (recommended): pick a cutoff date and assign patches by date. This avoids temporal leakage and matches the autoregressive nature of any model trained on it.
• Spatial split (for OOD evaluation): hold out specific continental regions. Note that station density is highly inhomogeneous (see Discussion of biases).

A typical split is 80 % train / 10 % val / 10 % test on the reference timestamp, with a one-month gap on each side of the val/test windows to reduce correlation.

Temporal layout and overlap

A single parquet row contains T hourly frames centred on date and spans

[date - back_step_hours, date + forward_step_hours]

at 1-hour cadence. With the defaults (5 back / 1 forward) that is T = 7 frames per row.

Consecutive reference timestamps are placed every --cadence_hours hours (default 1 h). The number of frames that overlap between two neighbouring rows is T - cadence_hours:

cadence_hours	Window	Overlap per row
1	7 frames	6 / 7
3	7 frames	4 / 7
4	7 frames	3 / 7
7	7 frames	0 / 7 (no overlap)

For pre-training the dense overlap is a feature (heavy augmentation of the same physical state). For clean train/val/test splits, prefer --cadence_hours 7 and use --ref_hours_utc 0,3,6,9,12,15,18,21 together with --cadence_hours 24 to pick the synoptic hours.

The reference archive for GMGSI starts on 2021-07-14; rows anchored before that date are silently skipped.

Data instance

import numpy as np
import pyarrow.parquet as pq

table = pq.read_table(
    "data_satellite_metar_v1/2021-07-14_00-00_global_rnXXXXXXXX_patches_0000.parquet"
)
row = table.to_pylist()[0]

sat = np.frombuffer(row["sat_data"], dtype=row["sat_dtype"]).reshape(row["sat_shape"])
metar = np.frombuffer(row["metar_data"], dtype=row["metar_dtype"]).reshape(row["metar_shape"])
ele = np.frombuffer(row["elevation_data"], dtype=row["elevation_dtype"]).reshape(row["elevation_shape"])

print("sat   :", sat.shape, sat.dtype, "valid frac =", float(np.isfinite(sat).mean()))
print("metar :", metar.shape, metar.dtype, "valid frac =", float(np.isfinite(metar).mean()))
print("elev  :", ele.shape, ele.dtype, "nodata frac =", float((ele == -9999.0).mean()))
print("centre:", row["lon"], row["lat"], "at", row["date"])

Expected output (typical mid-latitude patch):

sat   : (7, 4, 128, 128) float16  valid frac ≈ 0.99
metar : (7, 7, 128, 128) float32  valid frac ≈ 0.03
elev  : (128, 128) float32        nodata frac ≈ 0.0
centre: -90.4 38.8 at 2021-07-14 00:00:00

Data modalities

• Image (grayscale multi-channel): sat_data — 4 channels at 0.1° resolution, observed hourly.
• Tabular (gridded): metar_data — 7 channels at the same 0.1° resolution, observed approximately hourly but only at station locations. The temporal axis makes the modalities jointly a 4-D tensor.
• Static image (grayscale): elevation_data — 1 channel, no temporal axis.
• Tabular metadata: epsg, x_coord, y_coord, lon, lat, patch_x_idx, patch_y_idx, date.

Dataset creation

Pipeline

The data is built by src/generate/generate_satellite_metar_dataset_v1.py. The chunked production pipeline is run_satellite_metar_pipeline.sh, which:

1. Iterates over the [start_date, end_date) range in 3-month chunks.
2. For each chunk, calls the generator with --cadence_hours 7 (the default 7-frame window then becomes non-overlapping, so each physical hour appears in exactly one parquet row).
3. Stages the resulting parquets into a per-chunk subfolder and pushes them to the HF dataset repo via hf upload-large-folder.
4. Deletes the local staging folder before continuing, so the host disk is never asked to hold more than one chunk at a time.

Per-reference-timestamp recipe

For each reference timestamp t:

1. Satellite stack: download the 7 GMGSI GeoTIFFs gmgsi_global_8km/YYYYMMDD_HH00.tif for t-5h, t-4h, …, t, t+1h from gs://gmgsi_global_8km, cache the most recent 7 in memory, stack them along the time axis.
2. METAR stack: for each day touched by the window, load the metar_global_1h parquet chunks from gs://metar_global_1h that might overlap that day (filename-based filter, then day-clipped after load), cache the most recent 8 days in memory. For each hourly frame, select observations whose timestamp falls in [t-3600s, t], rasterize them onto the 3600×1800 grid, encode the 7 METAR features, and stack along the time axis. Per-cell deduplication keeps the latest observation.
3. Elevation: load data/ELE.tif once per process, reproject bilinearly to the GMGSI 0.1° grid, reuse for every patch.
4. Patch extraction: tile the global 3600×1800 grid into 128×128 patches with a 64-pixel stride. By default --patches_fraction 0.5 keeps a random 50 % of patches per reference timestamp to limit storage. Patches that are ≥ --max_nan_fraction (default 0.95) NaN in the satellite field are dropped — typically high-latitude patches outside the geostationary field of view.
5. Serialization: each kept patch is serialized with its shape and dtype metadata and written to a patches_<NNNN>.parquet file holding up to 64 rows.

Versioning

Version	Generator	Notes
v1	generate_satellite_metar_dataset_v1.py	First public release: GMGSI 4-ch + METAR 7-ch + optional elevation, 128×128 patches, 7-frame window.

Source data

NOAA GMGSI (satellite)

• Provenance: NOAA Global Mosaic of Geostationary Satellite Imagery, hosted on the AWS Open Data program (s3://noaa-gmgsi-pds/).
• License: U.S. Government work, public domain (17 U.S.C. § 105).
• Acquisition: src/gmgsi/download_gmgsi.py downloads the four GLOBCOMPLIR / GLOBCOMPVIS / GLOBCOMPWV / GLOBCOMPSIR channels anonymously from S3, reprojects them to a common 0.1° global EPSG:4326 grid (-180..180, -90..90, 3600×1800), and uploads the resulting 4-band GeoTIFFs to gs://gmgsi_global_8km/gmgsi_global_8km/.
• Temporal coverage: ~2021-07-14 to present, hourly.
• Spatial coverage: ~73°N to ~73°S (geostationary satellite field of view); polar regions are NaN in the source data.

IEM ASOS / METAR (ground stations)

• Provenance: Iowa Environmental Mesonet (IEM) ASOS / AWOS / METAR archive.
• License: IEM distributes the data under a non-commercial attribution clause; treat as attribution required, no warranty. This dataset is a derived work — please cite IEM and the upstream MADIS / NWS feeds in any publication that uses it.
• Acquisition: src/metar/download_metar.py queries the IEM bulk download endpoint, deduplicates to one record per station per timestamp, keeps hourly observations in the global IEM network, filters to |lat| ≤ 73° to match the GMGSI field of view, and uploads to gs://metar_global_1h/.
• Typical density: ~5,400 stations globally, ~200,000 records/day.
• Cadence: 1 hour, all 24 hours UTC.

Elevation DEM (ELE.tif)

• Provenance: bundled at data/ELE.tif in the repo; downloaded from gs://eumetsat_mtg_preprocess/assets/ELE.tif by src/generate/elevation.py if absent. Bounds are -180..180, -60..65 in EPSG:4326 at 0.0083° resolution (no nodata value set in the source). License follows the upstream asset; if redistributing, verify the DEM source (likely a derived SRTM / ETOPO1 / GEBCO product).
• Use: bilinear reprojection to the 0.1° GMGSI grid.

Curation choices

• Reference archive cutoff: START_DATE_GLOBAL = 2021-07-14 (first available GMGSI timestamp).
• Time window: default 5-back / 1-forward frames (7 hours total) at 1-hour cadence. --back_step_hours and --forward_step_hours are configurable.
• Patch stride: 64 pixels (50 % overlap between adjacent patches).
• Patch sampling: random 50 % of patches per reference timestamp by default (--patches_fraction 0.5); patches with ≥ 95 % NaN satellite pixels are dropped.
• METAR dedup: latest observation per (station, timestamp) before rasterization, then per-cell latest observation within the rolling window. Stations outside |lat| ≤ 73° are dropped.
• METAR NaN handling: missing tmpc / dwpc / mslp / wind are kept as NaN in the rasterized grid. Missing p01m is rasterized as 0 (matches the operational convention "no report = no accumulation"). Missing skyc becomes NaN cloud_cover.

Annotations

The METAR channel is an annotation in the learning sense (it is what the satellite field is meant to predict or impute), but it is not a human annotation: every cell value is derived directly from the corresponding IEM METAR observation. There is no annotation process involving human labelers. The sky-condition encoding is a deterministic lookup table (see METAR ground-station features).

Personal and sensitive information

The dataset contains no personal data. The finest-grained geographic unit is a 0.1° grid cell (≈ 11 km × 11 km at the equator), which is too coarse to identify individuals. Station codes are publicly listed ICAO-style identifiers; lat/lon are the official station coordinates published by IEM.

Considerations for using the data

• NaN handling is part of the contract. The METAR grid is mostly NaN (typically > 95 % per patch). Any model that ignores this will learn to predict NaN. Loss masking, masked-token objectives and weighted RMSE are the natural choices.
• GMGSI is also mostly NaN near the poles. The same patch may have 0 %–100 % valid satellite pixels depending on latitude; patches with more than --max_nan_fraction valid pixels missing have already been filtered out at generation time.
• Channel semantics differ. Three of the four satellite channels are brightness temperatures (K) and one (vis) is a reflectance (dimensionless 0–1). Normalise per channel before training.
• METAR observations are point measurements, not grid-cell averages. A station at the edge of a 0.1° cell is rasterized into that cell, so neighbouring cells can carry the same observation. Treat the field as sparse samples, not as a gridded analysis.
• p01m is rasterized as 0 when missing. If you need to distinguish "no report" from "0 mm", fetch the raw METAR parquets directly.
• Cadence / overlap is configurable but baked in. If you train on the default 1-hour cadence with a 7-frame window, neighbouring parquet rows are heavily correlated. Either re-generate with a larger --cadence_hours or build your sampler to drop near-duplicate windows.
• Re-using the global elevation for evaluation is a leak. The DEM is static; treating it as an input is fine, but it must not appear in any "future" channel when evaluating a forecasting model.

Social impact

• Positive: the dataset is intended to enable open, reproducible research on global weather nowcasting and foundation-model pre-training, with explicit coverage of the entire globe (not just the CONUS / Europe bias of most public ML weather corpora). It is released under CC-BY-4.0.
• Operational: any model trained on this data is a research artifact, not an operational forecast system. It must not be used for safety-critical decisions without a full evaluation by a National Meteorological or Hydrological Service.
• Environmental: re-distributing NOAA GMGSI derivatives is permitted (U.S. Government work, public domain); the IEM METAR data carries the IEM attribution clause — please cite IEM in any publication. The bundled DEM is a derived product; check the upstream license before re-distributing the dataset outside CC-BY.

Discussion of biases

• METAR station density is highly inhomogeneous. Western Europe, the CONUS, Japan and Australia are very densely sampled; the Southern Hemisphere oceans, Africa, central Asia and the polar regions are extremely sparse. A model that ignores spatial position will learn an "average ocean-with-no-stations" prior for most of the globe.
• The geostationary satellite field of view caps coverage at ~73° in both hemispheres. Polar weather is structurally absent from the satellite field and very sparse on the ground side; do not expect good performance there.
• METAR reporting is uneven in time. Night-time METAR cycles are less frequent than daytime cycles at many stations (especially auto-ASOS), and special reports inflate the cadence during severe weather. Models that treat the 1-hour window as a uniform sample will over-weight convective regimes.
• METAR skyc is a categorical code reduced to an opaque fraction. The encoding is monotonic with cover but loses the height / multi-layer information present in the raw observation.
• The DEM is bounded (-60..65 latitude) and is the only static channel, so a model that has learned to use it implicitly assumes that ocean cells have elevation = -9999.0. Pre-train on elevation-masked inputs if you want the model to be robust to ocean-only settings.
• The default patches_fraction = 0.5 is a uniform random subsample, not a balanced one. Continental coverage still dominates the resulting patch counts.

Additional information

Licensing

• Code that generates the dataset: Apache-2.0 (see the repo LICENSE).
• The dataset itself is released under Creative Commons Attribution 4.0 (CC-BY-4.0). Please cite IEM and the NOAA GMGSI programme when publishing results that use the METAR or satellite channels.
• The bundled DEM (data/ELE.tif) is a derived product: verify the upstream license before redistributing the dataset outside CC-BY.

Citation

If you use this dataset in academic work, please cite:

MeteoLibre. Global GMGSI + METAR Patches (v1). Hugging Face dataset, 2025. URL: https://huggingface.co/datasets/meteolibre-dev/global_sat_metar

And the upstream data providers:

NOAA Global Mosaic of Geostationary Satellite Imagery (GMGSI). NOAA Open Data Dissemination Program on AWS. URL: https://registry.opendata.aws/noaa-gmgsi/

IEM ASOS / AWOS / METAR archive. Iowa Environmental Mesonet, Iowa State University. URL: https://mesonet.agron.iastate.edu/

Repository

• Dataset generator: src/generate/generate_satellite_metar_dataset_v1.py
• Production pipeline: run_satellite_metar_pipeline.sh
• Companion datasets and downloads: src/gmgsi/, src/metar/

Contact

Issues and questions: open a ticket in the meteolibre_datasetgen GitHub repository.

---

Last updated: 2025.