cosmic-crunch

Download JPL GENESIS COSMIC radio-occultation ASCII data files and convert them to netCDF4.

cosmic-crunch crawls the JPL GENESIS data archive, downloads the Level-2 ASCII occultation profiles, and (optionally) converts them into self-describing netCDF4 files.

Problems this solves

Reach for this if you are trying to:

• Bulk-download COSMIC-1 (FORMOSAT-3) GNSS radio-occultation (RO) data from the JPL GENESIS archive without hand-writing a crawler.
• Convert COSMIC / GENESIS Level-2 ASCII occultation profiles to netCDF4 — the profile's fields (temperature, pressure, refractivity, water vapour, …) as a self-describing, xarray-friendly file.
• Resume an interrupted bulk pull — downloads are atomic and skip files already present, so you can just re-run.
• Download only the slice you need — filter by instrument, year, or date (year and date by regex) instead of pulling the whole archive.

Site restructure note (2020 → 2026). JPL restructured the GENESIS site: the old crawl root (/ftp/pub/genesis/glevels) is dead, so v1 of this tool silently "succeeded" while downloading nothing. v2 targets the current root https://genesis.jpl.nasa.gov/ftp/glevels/ and fails loudly — a crawl that finds nothing exits non-zero instead of pretending to succeed.

Mission status. COSMIC-1 (FORMOSAT-3, flight modules FM1–FM6, served here as cosmic1/–cosmic6/) was decommissioned in 2020, so this is a static archive — the data has stopped changing. (COSMIC-2 is a different mission on a different archive and is out of scope.) The COSMIC program's primary archive is UCAR's CDAAC; this tool crawls JPL's independent GENESIS processing of COSMIC-1, which is a separate archive with its own ASCII format.

Installation

pip install cosmic-crunch

This installs the cosmic-crunch command with two subcommands, get and convert. Python 3.10+ is required.

To hack on the package itself, install from a clone instead: pip install -e . (or pip install -e ".[test]" to run the test suite).

Usage

cosmic-crunch has a single entry point with two subcommands:

cosmic-crunch get      # crawl the GENESIS site and download ASCII files
cosmic-crunch convert  # convert downloaded ASCII files to netCDF4

cosmic-crunch get

usage: cosmic-crunch get [-h] [--base-url BASE_URL] [--logfile LOGFILE]
                         [--instrument INSTRUMENT] [--year_regex YEAR_REGEX]
                         [--date_regex DATE_REGEX] [--processes PROCESSES]
                         [--test] [--netcdf4] [--skip_empty] [--compress]
                         [--complevel COMPLEVEL]

Downloads are atomic and resumable: each file is streamed to a .part temporary and renamed into place only once complete, and files already present with a matching size are skipped — so an interrupted bulk pull can simply be re-run.

A successful run resembles:

$ cosmic-crunch get --year_regex=2006 --date_regex=2006-05-02 --netcdf4 --skip_empty --processes=4
Crawling all ./cosmic<#>/postproc: 100%|████████████████████| 6/6 [00:03<00:00,  1.61it/s]
Crawling all ./cosmic<#>/.../<year>: 100%|██████████████████| 6/6 [00:03<00:00,  1.59it/s]
Crawling all ./cosmic<#>/.../<date>: 100%|██████████████████| 3/3 [00:03<00:00,  1.17s/it]
Crawling all ./cosmic<#>/.../L2/<format>: 100%|█████████████| 4/4 [00:04<00:00,  1.09s/it]
Downloading data files: 100%|███████████████████████████████| 20/20 [00:26<00:00,  1.33s/it]
Converting ASCII to netCDF4: 100%|██████████████████████████| 20/20 [00:03<00:00,  6.32it/s]

ASCII to netCDF4 conversion summary:
 - Successful conversions: 17
 - Skipped conversions:    3
 - Conversion errors:      0
 - Total number of files:  20

Downloaded files are written under ./jpl_cosmic/<year>/<date>/txt/.

cosmic-crunch convert

usage: cosmic-crunch convert [-h] [--logfile LOGFILE] [--processes PROCESSES]
                             [--skip_empty] [--compress] [--complevel COMPLEVEL]
                             path [path ...]

Convert one or more ASCII .txt.gz files — or directories of them (crawled recursively) — to netCDF4. netCDF4 files are written into a sibling nc/ directory (mirroring the txt/ layout), or beside the source file otherwise.

cosmic-crunch convert ./jpl_cosmic/2006/ --skip_empty --processes=4

Use from Python

The conversion machinery is importable as a library:

from cosmic_crunch import convert

# Convert a directory tree (or a single file) of ASCII profiles:
convert.crawl_convert(["./jpl_cosmic/2006/"], processes=4)

# Or parse one ASCII file directly into a header dict and a dict of
# pandas DataFrames (one per profile):
header, data, is_empty = convert.read_cosmic_ascii_file(
    "20060501_0632co1_g35_2p6.L2.txt.gz"
)

The netCDF4 output nests each profile in a group (see the structure below), so name the group when reading it back — a bare xarray.open_dataset(path) shows only the global attributes:

import xarray as xr

profile = xr.open_dataset(
    "20060501_0632co1_g35_2p6.L2.nc", group="COSMIC1-Profile"
)

Compression

Output is uncompressed by default. --compress turns on lossless zlib compression (--complevel sets the level, default 7); stored values round-trip bit-identically either way.

Compression is opt-in because it does not reliably shrink these files. A COSMIC profile is many small float64 variables, and HDF5's per-variable chunk + filter overhead outweighs zlib's savings on short profiles — enabling compression can nearly double a small file. It only pays off for long profiles and bulk archival (measured crossover ≈ 1000 levels). Enable it when your profiles are large, and measure on a sample before committing an archive.

netCDF4 output structure

Each ASCII file becomes one netCDF4 file. The ASCII header becomes global attributes; each DataType profile becomes a group whose variables are that profile's columns, indexed by an Index dimension:

netcdf 20060501_0632co1_g35_2p6.L2 {
  // global attributes: ProductCreationTime, ShortName, DataSetID,
  //                     PlatformShortName, Receiver, ... (the ASCII header)

  group: COSMIC1-Profile {
    dimensions:  Index = <n> ;
    variables:   Height, Lat, Lon, Refractivity, Temperature, Pressure, WV Pressure ;
  }
  group: ECMWF-Profile {
    dimensions:  Index = <n> ;
    variables:   Height, Lat, Lon, Refractivity, Temperature, Pressure, WV Pressure ;
  }
}

Missing values in the ASCII data (-9999) are stored as NaN.

Security note

v1 parsed header values with eval(), allowing arbitrary code execution from a malicious or corrupted data file. v2 uses ast.literal_eval with a raw-string fallback — header parsing can no longer execute code.

Development

pip install -e ".[test]"
python -m pytest -q        # offline test suite (never touches the network)
ruff check .

License

MIT — see LICENSE. Built by Erick Shepherd.