GFS 2 m temperature — NWP quickstart¶

Fetch a recent GFS run's 2 m temperature over a small bbox and read the resulting Cloud-Optimized GeoTIFF.

Requirements (live download):

pip install earthlens[nwp]
conda install -c conda-forge eccodes libgdal-grib   # binary libs

herbie-data downloads only the .idx byte range for the requested variable (>99 % bandwidth saving); pyramids.grib.open_grib reads it with GDAL's GRIB driver, and earthlens crops it to the bbox.

Setup¶

A couple of imports: datetime to pick a recent cycle date, and the unified EarthLens entry point that drives the NWP backend.

In [ ]:

import datetime as dt

from earthlens.earthlens import EarthLens

import datetime as dt from earthlens.earthlens import EarthLens

Pick a recent cycle¶

GFS keeps roughly the last 10 days of runs on NODD, so we target yesterday's 00Z cycle to be safe.

In [ ]:

# A recent 00Z cycle (GFS keeps ~10 days on NODD).
cycle_day = (dt.datetime.now(dt.UTC) - dt.timedelta(days=1)).strftime('%Y-%m-%d')
cycle_day

# A recent 00Z cycle (GFS keeps ~10 days on NODD). cycle_day = (dt.datetime.now(dt.UTC) - dt.timedelta(days=1)).strftime('%Y-%m-%d') cycle_day

Build the download request¶

Describe the request to EarthLens: the nwp backend, the gfs dataset, the 2 m temperature variable, the date window, a small bbox over the US Northeast, the output directory, lead step 0, and the aws mirror.

In [ ]:

lens = EarthLens(
    data_source='nwp',
    dataset='gfs',
    variables=['temperature_2m'],
    start=cycle_day,
    end=cycle_day,
    aoi=[-80, 40, -75, 45],
    path='out/gfs',
    steps=[0],
    mirror='aws',
)

lens = EarthLens( data_source='nwp', dataset='gfs', variables=['temperature_2m'], start=cycle_day, end=cycle_day, aoi=[-80, 40, -75, 45], path='out/gfs', steps=[0], mirror='aws', )

Download the COGs¶

download() fetches each cycle/step as a Cloud-Optimized GeoTIFF cropped to the bbox and returns the written paths. We list their file names to confirm what landed on disk.

In [ ]:

paths = lens.download(progress_bar=False)
[p.name for p in paths]

paths = lens.download(progress_bar=False) [p.name for p in paths]

Read one output¶

The output is a standard COG, so we open the first one with pyramids.grib.open_grib and inspect its grid shape (bands, rows, columns).

In [ ]:

from pyramids.grib import open_grib

ds = open_grib(str(paths[0]))
ds.shape

from pyramids.grib import open_grib ds = open_grib(str(paths[0])) ds.shape

Next steps¶

To request several lead times, pass steps=[0, 6, 12, 24] (or horizon=48); each (cycle, step) yields one COG. Add aggregate=AggregationConfig(freq="1D", op="mean") to reduce the stack to daily means.