Make a gmt xarray accessor to store metadata from grdinfo

[weiji14]

Description of the desired feature

Metadata (data about data) is important, and there are specific kinds of metadata used by GMT when it decides how to plot a grid, such as:

• Registration type: Gridline/Pixel
• Coordinate system: Cartesian/Geographic
• Region: xmin/xmax, ymin/ymax, zmin/zmax
• Default Colormap (For GMT 6.1.0+ see Add a CPT column to the gmt_data_server.txt file gmtserver-admin#62)

Instead of explicitly wrapping gmt info or gmt grdinfo as in #147, why not allow the user to just access it via grid.gmt.someattribute? This would make use of xarray accessors and call grdinfo under the hood to retrieve those metadata properties.

Pros:

• We can design it so that the attributes (e.g. z-range) update dynamically when we cut/transform/resample/etc the grid, unlike in Indexing preserves outdated attrs which cause trouble downstream pydata/xarray#2247 where the attributes become out of date or are dropped with each operation
• Will be useful for autodetecting pixel/gridline registration in Properly allow for either pixel or gridline registered grids #476
• Won't need More informative metadata when loading earth_relief grids #494 so much anymore.
• Could store the region/projection information inside, and have fig.basemap() et al. use it automatically
• Tab-completion!!
• Avoids clashing with namespace of other libraries for things like crs and whatnot.

Cons:

• We would be designing yet another standard, as per https://xkcd.com/927/
• Would involve some development effort
• Might not be as discoverable as simply using grid.attrs["someattribute"] or pygmt.grdinfo(grid), but that's just a documentation thing.

In reality, this gmt accessor can do more than just hold metadata. We could extend it to do grid.gmt.plot and more. But let's start with having it hold the metadata we want, and then build on top of it from there.

References:

• Return values of gmt.info #147 (comment)
• Properly allow for either pixel or gridline registered grids #476 (comment)
• Indexing preserves outdated attrs which cause trouble downstream pydata/xarray#2247

Other xarray accessor examples:

• rioxarray - https://corteva.github.io/rioxarray/stable/examples/crs_management.html
• cf-xarray (Get CF attributes)
• xgeo - https://geosynopsis.github.io/xgeo/docs/html/examples.html
• salem - https://salem.readthedocs.io/en/v0.2.3/xarray_acc.html

Are you willing to help implement and maintain this feature? Yes

[seisman]

GMTDataArrayAccessor is powerful but it's important to know that xarray accessors have some limitations (https://docs.xarray.dev/en/stable/internals/extending-xarray.html):

Accessors are created once per DataArray and Dataset instance. New instances, like those created from arithmetic operations or when accessing a DataArray from a Dataset (ex. ds[var_name]), will have new accessors created.

More in-depth discussions are available at pydata/xarray#3205 and pydata/xarray#3268.

To understand these limitations, here is a simplified version of the GMTDataArrayAccessor xarray accessor. Some print statements are added to show how the xarray accessor works:

import xarray as xr

@xr.register_dataarray_accessor("gmt")
class GMTDataArrayAccessor:
    def __init__(self, obj):
        print("Inialize the gmt accessor")
        self._obj = obj
        self._registration = 0
        self._gtype = 0

    @property
    def registration(self):
        print("Returning registration")
        return self._registration

    @registration.setter
    def registration(self, value):
        print("Setting registration")
        self._registration = value

    @property
    def gtype(self):
        print("Returning gtype")
        return self._gtype

    @gtype.setter
    def gtype(self, value):
        print("Setting gtype")
        self._gtype = value

[seisman]

The following test shows that accessors are created/initialized after any grid operations (like slice and arithmetic operations):

import numpy as np
import xarray as xp

interval = 2.5
lat = np.arange(90, -90 - interval, -interval)
lon = np.arange(0, 360 + interval, interval)
longrid, latgrid = np.meshgrid(lon, lat)
data = np.sin(np.deg2rad(longrid)) * np.cos(np.deg2rad(latgrid))
grid = xr.DataArray(data, coords=[("latitude", lat), ("longitude", lon)])

print(grid.gmt.registration, grid.gmt.gtype)
grid.gmt.registration = 1
grid.gmt.gtype = 1
print(grid.gmt.registration, grid.gmt.gtype)

sliced_grid = grid[0:30, 0:30]
print(sliced_grid.gmt.registration, sliced_grid.gmt.gtype)

grid *= 2.0                                                                     
print(grid.gmt.registration, grid.gmt.gtype)

grid2 = grid * 2.0
print(grid2.gmt.registration, grid2.gmt.gtype)

The output of the script is:

Inialize the gmt accessor
Returning registration
Returning gtype
0 0
Setting registration
Setting gtype
Returning registration
Returning gtype
1 1
Inialize the gmt accessor
Returning registration
Returning gtype
0 0
Returning registration
Returning gtype
1 1
Inialize the gmt accessor
Returning registration
Returning gtype
0 0

[seisman]

The following test shows that accesors are created/initialized when accessing a DataArray from a Dataset. (The test is modified from pydata/xarray#3205 (comment))

import xarray as xr
import numpy as np

ds = xr.Dataset({'a': xr.DataArray(np.array([1, 2, 3])), 'b': xr.DataArray(np.array([4, 5, 6]))})

print(ds["a"].gmt.registration, ds["a"].gmt.gtype)

print(ds["a"].gmt.registration, ds["a"].gmt.gtype)

var = ds["a"]
print(var.gmt.registration, var.gmt.gtype)

print(var.gmt.registration, var.gmt.gtype)

The output is:

Inialize the gmt accessor
Returning registration
Inialize the gmt accessor
Returning gtype
0 0
Inialize the gmt accessor
Returning registration
Inialize the gmt accessor
Returning gtype
0 0
Inialize the gmt accessor
Returning registration
Returning gtype
0 0
Returning registration
Returning gtype
0 0