meteolibre-dev/ERA5_patchify

ERA5 Patchified Dataset

Patches extracted from ECMWF ERA5 reanalysis on a 0.25° global grid, tiled into 128×128 non-overlapping patches with float16 normalized channels. Designed for ML training — use alongside IFS HRES open data at inference time for a train-on-reanalysis / infer-on-forecast workflow.

Data Structure

Files are stored as Parquet, named:

era5_{first_snapshot}_{region}_patches_{group_idx:04d}_{file_idx:04d}.parquet

Columns

Column	Type	Description
ifs_data	bytes	Raw float16 bytes of the (T, C, H, W) patch tensor
ifs_shape	list[int]	Shape tuple, e.g. [3, 77, 128, 128]
ifs_dtype	str	"e" (numpy half / float16)
channel_names	list[str]	Ordered channel names (see below)
channel_offsets	list[float]	Per-channel normalization offset
channel_scales	list[float]	Per-channel normalization scale
elevation_data	bytes	Float16 elevation patch (128, 128)
elevation_shape	list[int]	(128, 128)
elevation_dtype	str	"e" (float16)
epsg	int	CRS, always 4326
lon	float	Center longitude of patch
lat	float	Center latitude of patch
patch_x_idx	int	X index in the regional grid
patch_y_idx	int	Y index in the regional grid
region	str	Region name (e.g. europe, global)
snapshot_labels	list[str]	ISO labels of the T snapshots
time_spacing_hours	int	Hours between snapshots (6)
resolution	float	Grid resolution in degrees (0.25)
patch_size	int	Spatial patch size (128)
source	str	Always "era5"

Recovering the Tensor

import numpy as np
import pyarrow.parquet as pq

table = pq.read_table("era5_2024-06-01T0000Z_europe_patches_0000_0000.parquet")
row = table.slice(0, 1).to_pydict()

# Reconstruct tensor
tensor = np.frombuffer(row["ifs_data"][0], dtype=row["ifs_dtype"][0]).reshape(row["ifs_shape"][0])
# tensor shape: (T, C, 128, 128), float16

# De-normalize
for ci, (offset, scale) in enumerate(zip(row["channel_offsets"][0], row["channel_scales"][0])):
    if scale != 0:
        tensor[:, ci, :, :] = tensor[:, ci, :, :].astype(np.float32) * scale + offset

Channels (77 total)

Surface (13 channels)

#	Name	Description	Unit	Offset	Scale
1	mucape	Convective available potential energy (surface-based)	J kg⁻¹	0	500
2	2t	2m temperature	K	273.15	40
3	2d	2m dewpoint temperature	K	273.15	30
4	10u	10m U wind component	m s⁻¹	0	30
5	10v	10m V wind component	m s⁻¹	0	30
6	100u	100m U wind component	m s⁻¹	0	40
7	100v	100m V wind component	m s⁻¹	0	40
8	tp	Total precipitation	m	0	0.05
9	sp	Surface pressure	Pa	101325	5000
10	msl	Mean sea level pressure	Pa	101325	5000
11	tcwv	Total column water vapour	kg m⁻²	0	50
12	tcc	Total cloud cover	(0–1)	0	1
13	lsm	Land-sea mask	(0–1)	0	1

Pressure Levels × 8 variables = 64 channels

Levels: 1000, 925, 850, 700, 500, 300, 250, 200 hPa

#	Prefix	Description	Unit	Offset	Scale
14–21	t_{level}	Temperature	K	273.15	50
22–29	u_{level}	U wind component	m s⁻¹	0	60
30–37	v_{level}	V wind component	m s⁻¹	0	60
38–45	q_{level}	Specific humidity	kg kg⁻¹	0	0.02
46–53	w_{level}	Vertical velocity	Pa s⁻¹	0	5
54–61	gh_{level}	Geopotential height	m	5000	30000
62–69	vo_{level}	Relative vorticity	s⁻¹	0	5×10⁻⁴
70–77	r_{level}	Relative humidity	%	50	50

Full channel name example: t_850 = temperature at 850 hPa.

Normalization

Values are stored normalized as float16:

normalized = (raw_value - offset) / scale

Recover raw values with:

raw_value = normalized * scale + offset

Normalization constants are identical to the IFS HRES dataset, enabling seamless cross-training (train on ERA5, infer on IFS HRES) without re-normalization.

Temporal Structure

Each patch contains T consecutive analysis snapshots spaced 6 hours apart (cycles 00, 06, 12, 18 UTC). The default is T=3 (18h window).

Consecutive patch groups stride by T×6 hours for continuous temporal coverage with no gaps:

Group 1: 00z → 06z → 12z
Group 2: 18z → 00z(+1d) → 06z(+1d)
Group 3: 12z(+1d) → 18z(+1d) → 00z(+2d)
...

Spatial Coverage

Region	Bounding Box (lon_min, lat_min, lon_max, lat_max)
global	(-180, -90, 180, 90)
europe	(-30, 30, 45, 75)
north_atlantic	(-80, 20, 0, 70)
north_america	(-140, 15, -50, 75)
asia	(50, 0, 160, 75)

Grid: 0.25° × 0.25° regular lat-lon (EPSG:4326). Patches are non-overlapping 128×128 grid cells (≈32° × 32° at 0.25° resolution).

Comparison with IFS HRES Patchified Dataset

	ERA5 (this dataset)	IFS HRES
Type	Reanalysis (best-estimate historical)	Operational analysis (near-real-time)
Temporal range	1940 → present	Rolling 2–3 days only
Latency	~5 days (ERA5T) / ~2 months (final)	Near real-time
Resolution	0.25°	0.25° (open data) / 0.08° (licensed)
Consistency	Reanalysis = physically consistent	Model upgrades cause breaks
CAPE	Surface-based CAPE	Most-unstable CAPE
Channels	77 (no tprate)	78 (includes tprate)
Geopotential	Height (m) after ÷9.80665	Height (m)
Normalization	Same offsets/scales	Same offsets/scales

Recommended workflow: Train on ERA5 (years of consistent data), infer on IFS HRES (real-time availability). The shared normalization and channel naming makes this a drop-in switch.

Elevation

Each patch includes a 128×128 float16 elevation map derived from a global DEM, reprojected to the same 0.25° grid. Elevation is stored raw (meters above sea level), not normalized.

Source

Data downloaded from the Copernicus Climate Data Store (CDS) via the cdsapi Python client.

• Surface: reanalysis-era5-single-levels
• Pressure levels: reanalysis-era5-pressure-levels

License

CC-BY-4.0 — please attribute ECMWF / Copernicus Climate Change Service as the data source. See the CDS terms of use and ERA5 licence.