NOAA CoastWatch Ocean Heat Content — Icechunk

Icechunk stores of the full NOAA CoastWatch Ocean Heat Content (OHC) Product Suite archive, spanning 2020 to present, for three regions:

Each region is a separate repository because the three regions use different lat/lon grids and cannot share a virtual Zarr array.

These repositories do not copy the science data. Icechunk stores only metadata and byte-range references back to the original NetCDF/HDF5 files hosted at NOAA CoastWatch; reads are streamed directly from coastwatch.noaa.gov over HTTP range requests ("virtual chunks"). See How to open it for the one extra step this requires.

How these were built

The notebooks that built these stores live in the GitHub repository https://github.com/fish-pace/icechunks (directory coastwatch-heat-content/), and a copy of each sits alongside this README on Source Cooperative:

They are provided to document how the archive was assembled and as a starting point for anyone building something similar. The write notebooks import shared helpers from icechunk_utils.py (included here); running them additionally requires Source Cooperative write credentials, so for most readers they are read-along references.

About the data

The Satellite Ocean Heat Content Suite (produced by USDOC/NOAA/NESDIS/OSPO with the University of Miami/Rosenstiel School) blends altimeter sea-surface-height anomalies with GeoPolar blended SST and the SMARTS climatology to estimate upper-ocean heat content and related fields. It is widely used in hurricane-intensification analysis.

• Resolution: 0.25° (~25 km), daily
• Projection: geographic latitude/longitude (WGS84-style; see the crs variable)
• Conventions: CF-1.6
• Source product: http://www.ospo.noaa.gov/Products/ocean/ocean_heat.html
• Source archive: https://coastwatch.noaa.gov/pub/socd2/coastwatch/ocean_heat/

Each region covers a different domain (e.g. the North Atlantic grid is 0°–60°N, 100°W–0°, 241 × 401). Open a repo and inspect its latitude/longitude coordinates for the exact extent.

Variables

Coordinates are time, latitude (degrees_north), and longitude (degrees_east).

Missing data is represented as -999 in the source files. A CF missing_value attribute has been added where the source omitted it, so xarray masks -999 to NaN automatically on read — no manual masking needed.

Groups inside each repository

Every region repo has the same three groups. They exist because the archive is not homogeneous, and a single virtual Zarr array cannot span files that differ in variable set or byte-level encoding:

1. daily vs. 14day* — these are two generations of the product (source dirs {region} and {region}14) with different variable sets. The 14day* files add the SST quality_flag variable (and 14day_v1 additionally carries a scalar quality_information summary). Keeping them separate preserves each generation's variables without forcing empty columns.
2. 14day_v1 vs. 14day — partway through 2025 the source files switched from NetCDF-3 (big-endian) to HDF5 (little-endian). Because Icechunk stores a single codec pipeline per array and references the original bytes, big-endian and little-endian chunks cannot live in one virtual array. The split is at day-of-year 084/085 of 2025, verified identical for all three regions.

Note on the overlap and gaps. The daily and 14day products overlap by a few days around January 2024 (both cover 2024-01-15…2024-01-18). A handful of source files with corrupt (all-zero) coordinates and data were dropped during construction, so a group may begin one day after its nominal codec boundary.

For a continuous single time series, open the groups separately and concatenate the shared variables at read time (see below).

How to open it

Requires icechunk >= 2.1 and xarray. The science arrays live at CoastWatch, outside these Icechunk stores, so you must authorize the virtual chunk container at open time — this is the one non-standard step.

CoastWatch blocks the default python-requests User-Agent, but Icechunk sends its own (icechunk-rust-x.y.z), so no custom headers are needed for reads.

One continuous time series across groups

Use chunks={} (dask-backed) as shown — with chunks=None the arrays are plain NumPy and this concatenation eagerly materializes several GB into memory.

Credits

• Data: USDOC/NOAA/NESDIS/OSPO; University of Miami / Rosenstiel School of Marine and Atmospheric Science (contact: Lynn K. Shay). Product page: http://www.ospo.noaa.gov/Products/ocean/ocean_heat.html
• Icechunk packaging: built with VirtualiZarr and Icechunk; the original file bytes remain at NOAA CoastWatch.