Temporal and spatial resampling

[forman]

Is your feature request related to a problem? Please describe.

Need a function or set of functions that can resample, reproject, and coregister data cubes from different sources.

Before:​ Any number of datasets Di each having any their own CRS, spatio-temporal resolution and coverage and number variables Vi,j​ After:​ A single dataset with a specified subset of Vi,j all resampled to the same grid, same CRS, same spatio-temporal resolution, same coverage.

Describe the solution you'd like

Input (Note, this may be split into individual spatial and temporal resampling functions): ​

• Datasets D_i and names of variables in each to be resampled​
• Mapping of variable names to up- and down-sampling method​s
• Target spatial CRS: EPSG code or CRS URI​
• Target spatial coverage (in CRS units): bbox or geometry​
• Target temporal coverage: time range​
• Target resolutions (in CRS units): Δx, Δy, Δt​
• Order of dimensions: first x,y then t, first t then x,y*​

Output:​

• Dataset with resampled variables and adjusted metadata*

To be considered:

• Allow resampling to a coverage subset that may not be aligned with the grids of the input dataset.​
• Focus is on spatio-temporal resampling and subsetting​
• Spatial resampling should be separate from temporal resampling because of the different dynamics of δV/δt and δV/δx. TBD!​
• Perform spatial resampling after/before temporal resampling. TBD! *​
• But ideally, we’d like to do a n-D resampling on any number n of a dataset’s dimensions.
• Allow specifying various up-sampling (interpolation or selection) and down-sampling (aggregation or selection or interpolation) methods per variable. ​
  • Up-sampling and down-sampling methods may be xcube extensions​
  • Up-sampling and down-sampling parameters should be method IDs or functions*​
  • Can we have the same interface for up-sampling and down-sampling?​
• Other
  • Unite common transformations: benefit from fact that some or all variables may be transformed in the same way and transformation coefficients/data can be precomputed before the actual transformation. ​
  • Allow for distributed, out-of-core processing using Dask. ​
  • Use performance shortcuts whenever possible, e.g. block resampling*​

Describe alternatives you've considered

Various libraries exists but none provides the generality needed here.

[forman]

Implementation Plan

This plan also addresses Allow for other common spatial CRS #112.

• [x] Provide a low level nd-image resampling function resample_ndimage that can do up- and downsampling for a given affine transformation matrx.
• [x] Introduce is_rectified: bool property in GeoCoding class.
• [x] The spatial resampling function should, given that source and target CRS are not the same, transform source coordinate arrays to target CRS, hence create a non-rectified source geo-coding.
• [x] The spatial resampling function should, given that the source geo-coding is not rectified. call xcube.core.rectify.rectify_dataset. To make rectify_dataset applicable, source spatial resolution should be lower or equal target. If not, apply resample_ndimage to source first.
• [x] Introduce crs: pyproj.crs.CRS property in ImageGeomand GeoCoding classes. Derive crs in ImageGeom.from_dataset(ds) and GeoCoding.from_dataset(ds) by using pyproj.crs.CRS.from_cf(ds.attrs). Deprecate is_geo_crs: bool argument in ImageGeomand GeoCoding constructors.
• [x] ImageGeom.from_dataset(ds) should also work if ds is a rectified dataset and hence x,y coordinates are 1D arrays, e.g. lon,lat.
• [x] Make use of the GeoCoding.is_rectified property in ImageGeom.from_dataset().
• [ ] After rectifying / reprojecting a dataset, set CRS CF attributes using ds.attrs.update(crs.to_cf()).
• [ ] When writing data cubes and CRS is known, make sure we encode CRS, i.e. use ds.attrs.update(crs.to_cf()).
• [x] Introduce ImageGeom.x_res and ImageGeom.y_res properties and use them where ImageGeom.xy_res is used. Let ImageGeom.xy_res return a tuple. Introduce ImageGeom.mean_xy_res and use it where old ImageGeom.xy_res was used and cannot be replaced by ImageGeom.x_res and ImageGeom.y_res.
• [ ] Revert @pont-us's fix from 03.06.2020 of #330 in xcube/core/imgeom.py. See comment "Reduce height if it takes the maximum latitude over 90° (see Issue #303).". Fix is not general enough, as y is not necessarily a latitude.
• [ ] Address Allow for other common spatial CRS #112 at the same time. -> No longer ask for 'lon' and 'lat' coordinate variables, instead use ImageGeom class.
• [ ] Introduce CRS in various xcube.core.extract functions so coordinates from different CRS can be used.
• [ ] No longer ask for 'lon' and 'lat' coordinate variables directly, instead use ImageGeom class. (#112)
• [x] Ask dask-image project to do a release 0.5 including new dask_image.ndinterp package.
• [x] Let data stores return CRS names in open parameters that are parsable by pyproj.crs.from_string() only
  • [x] xcube-cci data store
  • [x] xcube-cds data store
  • [x] xcube-sh data store
• [ ] Let data stores add CF conformat CRS grid mapping information into datasets
  • [ ] xcube-cci data store
  • [ ] xcube-cds data store
  • [ ] xcube-sh data store
• [ ] Document xcube resampling, especially the spatial resampling requires a clear description of the algorithm