accord-rosetta 0.1.2

Canonical data acquisition for seasonal climate forecasting

Rosetta

An API for fetching climate, environmental, and contextual datasets

Rosetta is ACCORD's data adapter layer. A single fetch() call retrieves data from many different providers (Copernicus CDS, the ECMWF Data Store, OPeNDAP, HTTP, S3, and others) and returns it as a normalized xarray dataset with canonical names, units, and coordinates. Data stays at the source; Rosetta does not host a central copy of anything. Adding a new dataset or provider is a catalog and adapter change, not a rewrite of your workflow.

Installation

pip install accord-rosetta

The distribution is published as accord-rosetta; the import name is rosetta. Rosetta requires Python 3.12 or newer.

Most CDS-based products (c3s/*, obs/era5) need CDS or ECMWF Data Store credentials; see CDS / ECDS setup.

Core API

import rosetta

ds = rosetta.fetch(
    product="nmme/cfsv2",
    variable="precip",
    init="2025-02",
    target="MAM",
    region=[-12, 6, 28, 42],
    hindcast=(1993, 2016),
    verbose=True,
    progress=True,
)

You pass a canonical product and variable plus optional temporal and spatial selectors. You get back a CF-aligned xarray dataset, ready for analysis or modeling. Results are returned in memory by default, or written to a local path or S3.

Verify an install and list what is available:

import rosetta
rosetta.check_all_products()          # config checks for every product
rosetta.catalog.list_products()       # list product ids

Normalized coordinates

Every dataset comes back with canonical coordinate names regardless of the upstream source:

Coordinate	Applies to	Description
lat, lon	all products	Spatial axes (latitude ascending)
time	observations / reanalysis	Monthly time axis (datetime64)
init_time	forecasts	Initialization time (datetime64)
lead_time	forecasts	Lead time (numeric, units vary by source)
member	forecasts	Ensemble member index

Numeric time encodings (for example "months since 1960-01-01") are decoded to datetime64 automatically.

Region input

region accepts three forms:

Form	Example	Behaviour
bbox	region=[-12, 6, 28, 42]	[lat_s, lat_n, lon_w, lon_e]
shapefile	region="kenya.shp"	bounding box slices upstream; result masked to the polygon
geometry	region=gdf.geometry	shapely geometry or geopandas GeoSeries, same masking

Shapefile and geometry inputs need the geo extra (pip install 'accord-rosetta[geo]'). The polygon is reprojected to EPSG:4326 and dissolved, so multi-feature files (for example an archipelago) clip correctly. Cells outside the polygon come back as NaN.

import rosetta

# Clip to a country boundary; values outside Kenya become NaN.
ds = rosetta.fetch(
    product="nmme/cfsv2",
    variable="precip",
    init="2025-02",
    target="MAM",
    region="kenya.shp",
    hindcast=(1993, 2016),
)

By default a grid cell is included only if its centre lies inside the region (boundary="center", the xarray/CDO/rasterio convention, and unbiased for area means). Pass boundary="cover" to keep every cell the region touches (matches rasterio's all_touched=True), which is useful for display or coarse grids where center-based selection can drop a country's thin tips. This applies to bbox and shapefile/geometry inputs alike.

ds = rosetta.fetch(..., region="kenya.shp", boundary="cover")

Note: polygons crossing the plus/minus 180 degree antimeridian are not yet handled (the derived bounding box spans the full longitude range). Split such geometries at the antimeridian before passing them in.

Multiple predictor domains

Seasonal forecasting workflows often use two predictor domains from the same model, for example a large sea-surface-temperature domain and a smaller regional precipitation domain. Make one fetch() call per domain; there is no combined helper, because the domains differ in extent and variable.

import rosetta

# SST predictor: large tropical domain
sst_predictor = rosetta.fetch(
    product="nmme/geoss2s", variable="sst",
    init="2025-02", target="MAM",
    region=[-20, 20, 30, 180], hindcast=(1993, 2016),
)

# Precipitation predictor: regional domain
prcp_predictor = rosetta.fetch(
    product="nmme/geoss2s", variable="precip",
    init="2025-02", target="MAM",
    region=[-20, 20, 10, 75], hindcast=(1993, 2016),
)

# Predictand: observations
predictand = rosetta.fetch(
    product="obs/chirps-v2-monthly", variable="precip",
    target="MAM", region=[-12, 15, 22, 52], hindcast=(1993, 2016),
)

Health checks

import rosetta

rosetta.check_product("nmme/cfsv2")             # one product, config only
rosetta.check_all_products()                    # all products, config only
rosetta.check_all_products(probe_remote=True)   # also probe the live source

Each result includes product, adapter, healthy, kind, message, and checked_at.

Available products

How to read the tables. Hindcast is each model's fixed reforecast period, not a fetch cap: real-time forecasts run past it to the present. Forecast is the live-verified real-time availability, shown as year–present (ongoing) or start–end (retired) when the pinned system version was superseded (hindcasts still fetch, but no new forecasts issue). Members (F/H) are the real-time-forecast and reforecast ensemble sizes, which differ. † marks a deprecated access route. Full field conventions are documented at the top of src/rosetta/catalog.yaml.

Several C3S entries pin a system version whose real-time stream has ended (a 2026 forecast init returns no data); they still fetch hindcasts. CMCC's live stream has moved to c3s/cmcc-sps4. JMA and UKMO do not yet have an active-forecast entry on their current systems (JMA CPS4, UKMO 605).

Seasonal forecast, NMME

Product	Model / system	Host	Adapter	Variables	Cadence	Members (F/H)	Hindcast	Forecast
nmme/cansipsic4	ECCC CanSIPS-IC4	Columbia CCSR	ccsr	precip, temp, sst	monthly	40 / 40	1990–2024	2024–present
nmme/ccsm4	NCAR/COLA CCSM4	Columbia CCSR	ccsr	precip, temp, sst	monthly	10 / 10	1982–2026	2014–present
nmme/cesm1	NCAR CESM1	Columbia CCSR	ccsr	precip, temp, sst	monthly	10 / 10	1982–2026	2017–present
nmme/geoss2s	NASA GEOS-S2S	Columbia CCSR	ccsr	precip, temp, sst	monthly	10 / 4	1981–2017	2019–present
nmme/spear	GFDL SPEAR	Columbia CCSR	ccsr	precip, temp, sst	monthly	30 / 15	1991–2020	2021–present
nmme/spearb	GFDL SPEARb	Columbia CCSR	ccsr	sst	monthly	30 / 15	1991–2020	2021–present
nmme/cfsv2 †	NCEP CFSv2	IRI Data Library	opendap	precip, temp, sst	monthly	28 / 28	1982–2010	2011–present

Seasonal forecast, C3S

Product	Model / system	Host	Adapter	Variables	Cadence	Members (F/H)	Hindcast	Forecast
c3s/cmcc	CMCC SPSv3.5 (sys 35)	Copernicus CDS	cds	precip, temp, sst	monthly	50 / 40	1993–2016	2020–2025 (retired)
c3s/cmcc-daily	CMCC SPSv3.5 (sys 35)	Copernicus CDS	cds	precip, temp, sst	daily	50 / 40	1993–2016	2020–2025 (retired)
c3s/cmcc-sps4	CMCC SPS4 (sys 4)	Copernicus CDS	cds	precip, temp, sst	monthly	50 / 30	1993–2024	2025–present
c3s/cmcc-sps4-daily	CMCC SPS4 (sys 4)	Copernicus CDS	cds	precip, temp, sst	daily	50 / 30	1993–2024	2025–present
c3s/dwd	DWD GCFS2.2 (sys 22)	Copernicus CDS	cds	precip, temp, sst	monthly	50 / 30	1993–2023	2023–present
c3s/dwd-daily	DWD GCFS2.2 (sys 22)	Copernicus CDS	cds	precip, temp, sst	daily	50 / 30	1993–2023	2023–present
c3s/dwd-gcfs21	DWD GCFS2.1 (sys 21)	Copernicus CDS	cds	precip, temp, sst	monthly	50 / 30	1993–2019	2020–2025 (retired)
c3s/eccc-cansips	ECCC GEM5-NEMO (sys 3)	Copernicus CDS	cds	precip, temp, sst	monthly	10 / 10	1990–2020	2021–2024 (retired)
c3s/eccc-cansipsv3	ECCC CanESM5.1 (sys 4)	Copernicus CDS	cds	precip, temp, sst	monthly	20 / 20	1980–2023	2024–present
c3s/eccc-daily	ECCC CanESM5.1 (sys 4)	Copernicus CDS	cds	precip, temp, sst	daily	20 / 20	1980–2023	2024–present
c3s/ecmwf	ECMWF SEAS5 (sys 51)	Copernicus CDS	cds	precip, temp, sst	monthly	51 / 25	1981–2016	2017–present
c3s/ecmwf-monthly	ECMWF SEAS5 (sys 51)	Copernicus CDS	cds	precip, temp, sst	monthly	51 / 25	1981–2016	2017–present
c3s/jma	JMA CPS3 (sys 3)	Copernicus CDS	cds	precip, temp, sst	monthly	155 / 10	1991–2020	2022–2026 (retired)
c3s/jma-cps2	JMA CPS2 (sys 2)	Copernicus CDS	cds	precip, temp, sst	monthly	13 / 10	1981–2016	2015–2022 (retired)
c3s/meteofrance	Météo-France Sys 9	Copernicus CDS	cds	precip, temp, sst	monthly	51 / 31	1993–2024	2025–present
c3s/meteofrance-daily	Météo-France Sys 9	Copernicus CDS	cds	precip, temp, sst	daily	51 / 31	1993–2024	2025–present
c3s/ukmo	UKMO GloSea6 GC3.2 (sys 604)	Copernicus CDS	cds	precip, temp, sst	monthly	62 / 28	1993–2016	2025–2026 (retired)
c3s/ukmo-daily	UKMO GloSea6 GC3.2 (sys 604)	Copernicus CDS	cds	precip, temp, sst	daily	7 / 7	1993–2016	2025–2026 (retired)
c3s/ecmwf-seas51c †	-	IRI Data Library	iridl	precip, sst	monthly	51 / 25	1993–2016	-

Sub-seasonal forecast

Product	Model / system	Host	Adapter	Variables	Cadence	Members (F/H)	Hindcast	Forecast
c3s/ecmwf-s2s	ECMWF S2S	ECMWF Data Store	cds	precip, sst	twice weekly	50 / 11	-	on-the-fly

c3s/ecmwf-s2s is date-keyed: call it with init="YYYY-MM-DD" (the issuance date). Its reforecasts are generated on the fly, so there is no fixed hindcast window. It uses the ECMWF Data Store (ecds.ecmwf.int), a separate service from the Copernicus CDS; see ECMWF Data Store (ECDS) setup.

Reanalysis

Product	Model / system	Host	Adapter	Variables	Cadence	Members (F/H)	Hindcast	Forecast
obs/era5	-	Copernicus CDS	cds	temp, precip, sst	monthly	-	1940–2025	-
obs/era5-land-monthly	-	Copernicus CDS	cds	precip, temp	monthly	-	1950–2025	-

Observation

Product	Model / system	Host	Adapter	Variables	Cadence	Members (F/H)	Hindcast	Forecast
obs/chirps-live-rhiza	-	Rhiza/Sheerwater	sheerwater	precip	daily	-	-	-
obs/chirps-v2-annual	-	UCSB CHC	http	precip	annual	-	1981–2024	-
obs/chirps-v2-daily	-	UCSB CHC	http	precip	daily	-	1981–2025	-
obs/chirps-v2-dekad	-	UCSB CHC	http	precip	dekad	-	1981–2025	-
obs/chirps-v2-dekadal-rhiza	-	Rhiza/Sheerwater	sheerwater	precip	dekadal-rolling	-	1997–2024	-
obs/chirps-v2-monthly	-	UCSB CHC	http	precip	monthly	-	1981–2025	-
obs/chirps-v2-pentad	-	UCSB CHC	http	precip	pentad	-	1981–2025	-
obs/chirps-v3-annual	-	UCSB CHC	http	precip	annual	-	1981–2025	-
obs/chirps-v3-daily	-	UCSB CHC	http	precip	daily	-	1998–2025	-
obs/chirps-v3-daily-rhiza	-	Rhiza/Sheerwater	sheerwater	precip	daily	-	2000–2024	-
obs/chirps-v3-dekad	-	UCSB CHC	http	precip	dekad	-	1981–2025	-
obs/chirps-v3-monthly	-	UCSB CHC	http	precip	monthly	-	1981–2025	-
obs/chirps-v3-pentad	-	UCSB CHC	http	precip	pentad	-	1981–2025	-
obs/ghcn	-	Rhiza/Sheerwater	sheerwater	precip	monthly	-	1981–2024	-
obs/imerg	-	Rhiza/Sheerwater	sheerwater	precip	monthly	-	2000–2024	-

For nmme/cfsv2 †, the model is still an active NMME member (live on CPC FTP); only its IRI Data Library access route is deprecated (IRIDL shutdown around October 2026), with no successor yet.

Output is always NetCDF (extensible to Zarr and GeoTIFF).

CDS / ECDS setup

Rosetta's cds adapter talks to two distinct ECMWF endpoints, which are separate services with separate accounts, API keys, and licence-acceptance flows:

• Copernicus Climate Data Store (cds.climate.copernicus.eu): most c3s/* products and obs/era5.
• ECMWF Data Store (ecds.ecmwf.int): the newer service, currently used by c3s/ecmwf-s2s.

Each product's catalog entry selects its endpoint. You need credentials and accepted licences for each endpoint you fetch from.

Copernicus CDS setup

For Copernicus CDS products (all c3s/* except c3s/ecmwf-s2s, plus obs/era5):

1. Create a CDS account and API key at https://cds.climate.copernicus.eu.

2. Add credentials to ~/.cdsapirc:

   cat > ~/.cdsapirc << 'EOF'
   url: https://cds.climate.copernicus.eu/api
   key: <YOUR-CDS-API-KEY>
   EOF
   

3. Accept the required dataset licences in the CDS web UI before your first download. Each dataset page has a "Terms of use" section you tick once per account. If rosetta.fetch() fails with a 403, the error names the dataset(s) still missing acceptance.

ECMWF Data Store (ECDS) setup

Required for c3s/ecmwf-s2s (and any future product whose catalog entry uses cds_url: "https://ecds.ecmwf.int/api"). ECDS is a separate service; you cannot reuse Copernicus CDS credentials.

1. Create an ECMWF account at https://www.ecmwf.int/, log in to https://ecds.ecmwf.int/, and copy your API key from your profile settings.

2. Point cdsapi at the ECDS endpoint. If you only use ECDS, the simplest setup is to replace ~/.cdsapirc:

   cat > ~/.cdsapirc << 'EOF'
   url: https://ecds.ecmwf.int/api
   key: <YOUR-ECDS-API-KEY>
   EOF
   

   If you also use the Copernicus CDS, keep one endpoint in ~/.cdsapirc and pass the other via the CDSAPI_URL / CDSAPI_KEY environment variables per session.

3. Accept both layers of ECDS licences (ECDS returns 403 until both are accepted):

   • The site-wide Terms of Use, accepted once per account from the ECDS profile settings.
   • The dataset licence, ticked on the dataset's own page. For S2S: https://ecds.ecmwf.int/datasets/s2s-forecasts?tab=download#manage-licences. Other datasets follow the same https://ecds.ecmwf.int/datasets/<dataset-id>?tab=download#manage-licences pattern.

   As with the Copernicus CDS, a skipped licence produces a 403 that names exactly which licence is missing.

4. Test the connection:

   uv run pytest tests/test_integration.py::test_fetch_c3s_ecmwf_s2s_precip -v
   

Cache configuration

Rosetta caches adapter downloads locally using Nuthatch. Cache files live in ~/.nuthatch/rosetta by default (configured in pyproject.toml).

rosetta cache list                        # inspect the cache
rosetta cache clear                       # clear everything
rosetta cache clear --product nmme/cfsv2  # clear one product (with confirmation)

Development setup

git clone https://github.com/accord-research/rosetta.git
cd rosetta
uv sync

Relationship to DeepScale

Rosetta handles ingestion and normalization; DeepScale handles downscaling and skill evaluation. Their interface is standardized xarray.