Regridding Differences Between CDMS2 and xCDAT

[manaster]

Question criteria

• I added a descriptive title here.
• I searched the xCDAT GitHub Discussions to find a similar question and didn't find it.
• I searched the xCDAT documentation and/or Xarray documentation.

Describe your question

Recently, I have encountered some differences that arise when NASA's Integrated Multi-satellitE Retrievals for GPM (IMERG) data are resampled with CDMS2 and xCDAT. Specifically, the IMERG data were resampled from 30-minute, 0.1 degree gridded ‘.nc4’ files to 3-hourly, 0.1 degree gridded netCDF4 files and then regridded to 3-hourly, 2 degree gridded netCDF4 files. The resampling and re-gridding are done with both CDMS2 and xCDAT and differences are seen, which are described below. The folders containing the codes and files necessary to examine and recreate this problem are on “gates” at /p/user_pub/PCMDIobs/RSS_tests (see Peter Gleckler for more information).

3-hourly Resampling

Background
In this specific test, IMERG 30-minute, 0.1 gridded ‘.nc4’ files were downloaded for June 1st, 2010 and averaged to 3-hourly maps on the same initial 0.1 degree grid and were written out as netCDF4 files.

• Two different scripts:
  • resampling-code/make_IMERG.3hr.cdms.py
  • resampling-code/make_IMERG.3hr.xcdat.py
• Output files:
  • resampled-imerg-files/IMERG_Final.Run.V06B_0.1deg_3hr_201006-cdms.nc
  • resampled-imerg-files/IMERG_Final.Run.V06B_0.1deg_3hr_201006-xcdat.nc

Comparison of output files
In terms of the output rain rate, the 3-hourly averages are the same when the two files are compared with the exception being the first 3-hourly interval.

• The differences at this time step are quite small and only occur in areas where valid IMERG pixels are next to missing IMERG pixels (often near the sea-ice edge).
• All other time steps are the same between the two files.
• Why this difference occurs is still unknown.

2 Degree Regridding

Background
After the files are resampled to 3-hour averages, the 3-hourly, 0.1 degree gridded netCDF4 files are regridded to a uniform 2-degree latitude/longitude Earth grid.

• Two different scripts:
  • make_IMERG_2deg.cdms.py - CDMS2 horizontal regridder (regrid2.Horizontal)
  • make_IMERG_2deg.xcdat.py - xCDAT horizontal regridder (regridder.regrid2.Regrid2Regridder)
    • According to xCDAT’s website, this xCDAT regridder is a “pure python implementation of the regrid2 horizontal regridder from CDMS2’s regrid2 module”. Therefore, it can be expected that the outputs of these regridders should be nearly identical.
• Output files:
  • resampled-imerg-files/IMERG_Final.Run.V06B_2deg_3hr_201006-cdms.nc
  • resampled-imerg-files/IMERG_Final.Run.V06B_2deg_3hr_201006-xcdat.nc

Comparison of output files
When the 2-degree regridded outputs (are compared, the data between 60S and 60N are identical for both the CDMS2 and xCDAT files. However, at high latitudes, differences appear between the two files.

• Specifically, if there is a missing 3-hourly, 0.1 degree IMERG rain rate value (or values) within a given 2x2 degree grid box. CDMS2 and xCDAT appear to handle the regridding differently.
• CDMS2 will ignore the missing values and instead take the average of all valid data in a given 2x2 degree box.
• xCDAT will set the entire 2x2 degree box to missing if one or more IMERG values are missing within that box.

Hypothesis

As an additional part of this exercise, a 0.1 degree, 3-hourly file was created with xCDAT and regridded to 2 degrees using both the CDMS2 and xCDAT 2-degree regridding codes.

• The resulting output files were identical (code not included, but they are nearly identical to the make_IMERG_2deg.cdms.py and make_IMERG_2deg.xcdat.py codes in the "resampling-code" folder).
• This, along with comparing the output 0.1 degree, 3-hourly files created separately with CDMS2 and xCDAT, suggests that the differences between the 2-degree, 3-hourly files are related to the way that CDMS2 and xCDAT read and write missing values in netCDF4 files (see codes and files on “gates” at /p/user_pub/PCMDIobs/RSS_tests).

How to Recreate the Issue

1. Recreate the files discussed above using the original “.nc4” files located in the “raw-imerg-files” folder and the codes located in the “resampling-code” folder.
2. Change the paths in the appropriate codes based on the machine that they are using.
3. Run the “make_IMERG.3hr.cdms.py” and “make_IMERG.3hr.xcdat.py” codes
4. Run the “make_IMERG_2hr.cdms.py” and “make_IMERG_2hr.xcdat.py” codes in order to appropriately recreate these files.
5. Resulting files should match the files in the “resampled-imerg-files” folder.

Please do not hesitate to reach out with any questions or if clarification is needed on anything.

Are there are any possible answers you came across?

No response

Minimal Complete Verifiable Example (MVCE)

No response

Relevant log output

No response

Environment

xCDAT Version is 0.3.3

xr.show_versions() output listed below:

INSTALLED VERSIONS

commit: None
python: 3.10.6 | packaged by conda-forge | (main, Aug 22 2022, 20:36:39) [GCC 10.4.0]
python-bits: 64
OS: Linux
OS-release: 5.14.0-70.26.1.el9_0.x86_64
machine: x86_64
processor: x86_64
byteorder: little
LC_ALL: None
LANG: en_US.UTF-8
LOCALE: ('en_US', 'UTF-8')
libhdf5: 1.12.2
libnetcdf: 4.8.1

xarray: 2022.11.0
pandas: 1.5.1
numpy: 1.23.3
scipy: 1.9.3
netCDF4: 1.6.1
pydap: None
h5netcdf: None
h5py: None
Nio: None
zarr: None
cftime: 1.6.2
nc_time_axis: None
PseudoNetCDF: None
rasterio: None
cfgrib: None
iris: None
bottleneck: None
dask: 2022.10.2
distributed: 2022.10.2
matplotlib: None
cartopy: None
seaborn: None
numbagg: None
fsspec: 2022.10.0
cupy: None
pint: None
sparse: 0.14.0
flox: None
numpy_groupies: None
setuptools: 65.5.0
pip: 22.3
conda: None
pytest: None
IPython: None
sphinx: None

Anything else we need to know?

No response

[chengzhuzhang]

@manaster thank you very much for the careful write up of this issue! Peter gave a heads-up to the xCDAT team about this issue. And I think @jasonb5 mentioned some hypotheses about the difference in implementation earlier. Jason would you please help on this issue?

[gleckler1]

I believe the first issue raised ("3-hourly Resampling") does not involve use of a horizontal regridder, with the subtle discrepancies only in the first time step, with the other (7) time steps all being identical.

[gleckler1]

plotting code and plots now available at /p/user_pub/PCMDIobs/RSS_tests to help identify discrepancies.

[tomvothecoder]

Hey @gleckler1, just a heads up that I changed the group permissions for /p/user_pub/PCMDIobs/RSS_tests to g=rx using sudo so I can copy them to my home directory for testing.

[tomvothecoder]

Hey @manaster, can you check to see if you still get these differences with the latest version of xcdat (v0.7.1) compared to CDAT? There were enhancements and bug fixes that might have addressed them. Thanks!

[manaster]

Hi @tomvothecoder. Great! Thanks for the update. I will try to check on this soon.