Unusual grids

Most model output is on lat-lon, Gaussian, Lambert or polar stereographic grids, and fastgrib handles those like any GRIB reader. This page covers the less common ones, where fastgrib goes further than wgrib2 or ecCodes.

Spherical harmonics

Global spectral models (ECMWF's IFS, ARPEGE) store some fields as spherical-harmonic coefficients: GRIB2 templates 3.50 to 3.53 with packing 5.50 or 5.51, and GRIB1's equivalents.

coeffs = f.read(0)                                    # the coefficients
t = f.regrid((np.arange(-90, 90.1, 0.5), np.arange(0, 360, 0.5)), 0)
point = f.at(51.5, 0.0, 0)

read() returns the coefficients exactly as ecCodes and pygrib do. regrid() and at() go further and synthesise the field at any points, which those tools don't. They use ECMWF's normalisation and agree with pyshtools to 1e-5 K. A T63 field onto a global 0.5° grid takes 12 ms. Rotated and stretched fields (ARPEGE's kind) are synthesised in their own frame. encode() packs coefficients back to the same bits. There is an example in the gallery.

Satellite images

Space-view grids (template 3.90), such as Meteosat and Himawari images in GRIB, are placed using the CGMS navigation on the field's own ellipsoid. The points match wgrib2's, and PROJ's geos projection to 1e-10 degrees. Points beyond the Earth's edge are NaN, where wgrib2 prints a sentinel and ecCodes gives (0, 0). A full Meteosat disc of 13.8 million points takes 0.04 s.

Views that neither wgrib2 nor ecCodes can place (a satellite off the equator or over a pole, a rotated image, orthographic views) are placed the way PROJ's tilted perspective and orthographic projections do, to 1e-8 degrees. regrid() and at() work from these like any other grid.

Staggered grids

Grids whose alternate rows are offset by half a step (scanning-mode flags 0x08, 0x04, 0x02 and 0x01) are placed as wgrib2 places them.

NCEP's rotated E-grid (3.32768) and B-grid (3.32769, NAM's native output, which wgrib2 can't place) are placed as rotated grids about their centre. NAM's points land where NCEP's own NLAT and ELON fields put them, to those fields' packing precision.

Points listed one by one

Template 3.130 lists each point's position. Its fields read as flat values arrays with each point's latitude and longitude, and at() and regrid() use the nearest point.

Grids defined elsewhere

ICON's unstructured grid (3.101) and curvilinear grids (3.204) don't include their points in the message. Their fields read as flat values, and keys() names the grid file to use (uuidOfHGrid, numberOfGridUsed).