API for multi-dimensional resampling/regridding

[shoyer]

This notebook by @kegl shows a nice example of how to use pyresample with xray:
https://www.lri.fr/~kegl/Ramps/edaElNino.html#Downsampling

It would nice to build a wrapper for this machinery directly into xray in some way.

xref #475

cc @jhamman @rabernat

[rabernat]

Pyresample is probably overkill for that case. Aggregating / decimating regular lat-lon grids could probably be done much more simply. For example

N = 10
fac = 2
x = np.arange(N, dtype=np.float64)
np.add.reduceat(x, np.arange(0,N,fac)) / fac

This gives array([ 0.5, 2.5, 4.5, 6.5, 8.5])

This type of resampling has the advantage of preserving certain integral invariants, as opposed to the nearest neighbor resampling in the example above. (Imagine if there had been lots of spatial variance below the 5 degree scale in that example--it would have been aliased horribly. That was only avoided because the original field was very smooth.) It is also very fast.

Pyresample seems best for complicated transformations from one map projection to another. I'm not sure I fully understand how xray handles grids where the coordinates are themselves 2d fields, as in this example.

[shoyer]

Indeed, SciPy's ndimage.interpolation.zoom would probably be more appropriate (and faster) for regular grids.

Xray currently doesn't have any built-in support for handling projected data, but basic selection and regridding (from explicit arrays of 2D coordinates) seems in scope for the project.

[jhamman]

I'd be interested in helping build a wrapper / api that supports the following types of regridding operations. I don't think pyresample handles all of these:

1. Bilinear interpolation
2. Bicubic interpolation
3. Distance-weighted average remapping
4. Nearest neighbor remapping
5. Conservative (area) remapping
6. Largest area fraction remapping

I would also like to see some better support for 2D coordinate variables. These are specifically outlined in the cf-conventions in section: 5.2. Two-Dimensional Latitude, Longitude, Coordinate Variables. Maybe we can open an additional issue for that...

[forman]

@jhamman: any progress on this? Our team would be happy to contribute as we have similar requirements in our project.

[rabernat]

@forman I am starting to suspect that this might be possible to implement through #818

[jhamman]

@forman - no progress on my end and no plans to work on this in the next 6 months. If your team is interested in working on this, we can discuss the api further and I'm happy to provide input and guidance along the way. One option is to wrap pyresample and possibly fill in some of the missing pieces in their api. I'd like to see a user interface that looks something like this:

da.resample_like(other, kind='bilinear', dims=('lat', 'lon'))

As @rabernat mentions, for some remapping/resampling, the multi-dimensional groupby will help with some of these applications.

[rabernat]

I feel like the biggest application of the multi-dimensional groupby will be with "conservative resampling" and coarse-graining, where you want to make sure to conserve certain integrals (e.g. total heat content) while changing coordinates.

pyresample will be more useful for fine-graining and interpolating missing data.

[godfrey4000]

I have an immediate need in this area. My objective is to create a tool that will enable arithmetic on variables defined on lattices whose points don't coincide. Through my attempts thus far, it has become clear that I need data structures that incorporate spacial indexing and lattice indexing.

Since I have to tackle this issue to proceed, I thought I should follow the thinking discussed in this forum, so that it may be useful to others.

[rabernat]

@godfrey4000: lots of us in the climate community would like xarray-backed regridding. It is a hard problem, however. It wold be great if you wanted to work on it.

At a recent workshop, a group of xarray users developed a draft design document for a regridding package.
https://aospy.hackpad.com/Regridding-Design-Document-tENJARIeg83

Your comments on this would be very welcome. An open question is whether the regridding belongs within xarray or in a standalone package (which would of course have xarray as a dependency).

[jhamman]

@rabernat - I've more or less settled that this belongs in a separate package that uses xarray's accessor features.

I've setup a little project (xmap) to get this started that I hope we can garner some volunteers to help push forward (nudge @godfrey4000 and @forman): https://github.com/jhamman/xmap

There will likely be some overlap between features that xmap and xarray. In those cases, we can try to push relevant features toward xarray.

[godfrey4000]

I'm ready to start working on this project. I already have a prototype regridding class that I developed as part of another project. Working on that, I discovered these points:

• regridding takes a long time because the lattices can be huge
• the design should accomodate parallel processing on a cluster
• data needs to be normalized first (deal with missing values, etc.)
• the user will want choices

Some of these choices are:

• the destination lattice
• the interpolation algorithm
• subset of the dimension space
   
  As the first step in a strategy to achieve this with a sequence of realizable goals, I plan to implement a regridding of just the latitude and longitude dimensions.

Is there a style guide that I can/should follow? Something like this: https://google.github.io/styleguide/pyguide.html? Does it or something else define naming conventions?

[shoyer]

Is there a style guide that I can/should follow? Something like this: https://google.github.io/styleguide/pyguide.html? Does it or something else define naming conventions?

It's pretty standard to follow Python's PEP8 with NumPy-style docstrings.

I generally like the Google style guide, too, but it leans towards being overly strict -- sometimes it's OK to be more relaxed (e.g., it's rules for valid import statements). Also, the public facing version has gotten out of date (I think there are plans to update it).

[duncanwp]

@jhamman @godfrey4000 - I'm not sure of the status of this, but I'm the lead developer on a package called CIS which might be useful/relevant.

It was designed as a command line tool to allow easy collocation (resampling) between different model and observation datasets, but is now also a Python library. We spent a fair amount of time thinking about the various permutations and you can see some of the details in our paper here. Internally we currently use Iris Cube-like objects but it would be pretty easy to operate on xarray Datasets since they share a similar design.

The basic syntax is:

from CIS import read_data
X = read_data('some_obs_data.nc')
Y = read_data('some_other_data.nc')
X.sampled_from(Y)
# or...
Y.collocated_onto(X)

Happy to discuss further here, or in xmap.