conservative regridding with land-sea mask

[markmbaum]

I'm wondering how the weatherbench2 pipeline handles regridding of variables that have land-sea masks, where land points are NaN. For example, when regridding sea surface temperature, conservative regridding expands sea cells into land areas. The plot below shows an example of this.

• the top panel is a slice of SST from gs://weatherbench2/datasets/era5/1959-2023_01_10-6h-240x121_equiangular_with_poles_conservative.zarr
• the middle panel is a conservative regridding to the same 240x121 grid, starting from gs://weatherbench2/datasets/era5/1959-2022-full_37-6h-0p25deg_derived.zarr
• the bottom panel is the difference

The differences are extremely small, but the land-sea pattern is very different. Is there another step in the pipeline that restores a more realistic land-sea mask in the regridded data?

The full script producing that figure is below:

from weatherbench2.evaluation import make_latitude_increasing
from weatherbench2.regridding import Grid, ConservativeRegridder
import xarray as xr
import matplotlib.pyplot as plt
import jax

jax.config.update("jax_enable_x64", True)

hi_res = xr.open_dataset(
    "gs://weatherbench2/datasets/era5/1959-2022-full_37-6h-0p25deg_derived.zarr",
    engine="zarr",
)
hi_res = make_latitude_increasing(hi_res)
hi_res = hi_res["sea_surface_temperature"].sel(time="2020-01-01 00:00:00")
hi_res = hi_res.transpose("longitude", "latitude")

lo_res = xr.open_dataset(
    "gs://weatherbench2/datasets/era5/1959-2023_01_10-6h-240x121_equiangular_with_poles_conservative.zarr",
    engine="zarr",
)
lo_res = make_latitude_increasing(lo_res)
lo_res = lo_res["sea_surface_temperature"].sel(time="2020-01-01 00:00:00")
lo_res = lo_res.transpose("longitude", "latitude")

source_grid = Grid.from_degrees(lon=hi_res.longitude.values, lat=hi_res.latitude.values)
target_grid = Grid.from_degrees(lon=lo_res.longitude.values, lat=lo_res.latitude.values)

fig, axs = plt.subplots(3, 1, figsize=(6, 10))
axs = axs.flatten()

r = axs[0].pcolormesh(lo_res.values.T)
plt.colorbar(r, ax=axs[0])
axs[0].set_title("weatherbench2 240x121 dataset")

regridder = ConservativeRegridder(source_grid, target_grid)
regridded = regridder.regrid_dataset(hi_res)

r = axs[1].pcolormesh(regridded.values.T)
plt.colorbar(r, ax=axs[1])
axs[1].set_title("conservative regridding from 1440x721")

r = axs[2].pcolormesh((regridded - lo_res).values.T)
plt.colorbar(r, ax=axs[2])
axs[2].set_title("difference")

plt.tight_layout()
plt.show()