Utilities for working geographic data defined on various grids.
Features:
- regridding
- Permissively licensed python healpix utilities
Grids currently supported:
- regular lat lon
- HealPIX
- Documentation: https://nvlabs.github.io/earth2grid/
- GitHub: https://github.com/NVlabs/earth2grid
Pre-requisites:
Note
We recommend using the pytorch docker image on NGC.
Once the pre-requisites are installed you can install earth2grid with
pip install --no-build-isolation https://github.com/NVlabs/earth2grid/archive/main.tar.gz
Currently these are built for CUDA12, x86_64, linux, and pytorch 2.7 only.
VERSION=v2025.7.1
pip install \
-f https://github.com/NVlabs/earth2grid/releases/expanded_assets/$VERSION \
earth2grid
If in an NGC container, you may need to add --no-deps to avoid an error since the included pytorch is often not a released version.
>>> import earth2grid
>>> import torch
... # level is the resolution
... level = 6
... hpx = earth2grid.healpix.Grid(level=level, pixel_order=earth2grid.healpix.XY())
... src = earth2grid.latlon.equiangular_lat_lon_grid(32, 64)
... z_torch = torch.cos(torch.deg2rad(torch.tensor(src.lat)))
... z_torch = z_torch.broadcast_to(src.shape)
>>> regrid = earth2grid.get_regridder(src, hpx)
>>> z_hpx = regrid(z_torch)
>>> z_hpx.shape
torch.Size([49152])
>>> nside = 2**level
... reshaped = z_hpx.reshape(12, nside, nside)
... lat_r = hpx.lat.reshape(12, nside, nside)
... lon_r = hpx.lon.reshape(12, nside, nside)
>>> reshaped.shape
torch.Size([12, 64, 64])