ENH: Optional Read-Only RasterIO backend

[jhamman]

RasterIO is a GDAL based library that provides Fast and direct raster I/O for use with Numpy and SciPy. I've just used it a bit but have been generally impressed with its support for a range of ASCII and binary raster formats. It might be a nice addition to the suite of backends already available in xarray.

I'm envisioning a functionality akin to what we provide in the PyNIO backend, which is to say, read-only support for which ever file types RasterIO supports.

[fmaussion]

👍 Rasterio shines at reading georeferencing metadata out of any file, and I guess it would be no big deal to treat the various info as attributes in an xarray dataset. It is also possible to do lazy reading out of rasterio files.

(example with a geotiff file: https://github.com/fmaussion/salem/blob/master/salem/datasets.py#L263)

[jhamman]

Thanks @fmaussion. This was a helpful illustration of how this could be done. The salem GeoTiff and Grid objects include all and more than I was hoping to implement in xarray. However, after a bit more looking into this, I have mixed feelings about whether this would work in xarray. A brief summary of the pros/cons as I see them now.

Pros:

• Rasterio supports a wide range of raster formats (e.g. GeoTiff, ArcInfo ASCII Grid, etc.)
• Combined with pyproj, coordinate variables can be inferred
• Supports windowed reading (and writing), this would fit in well with the chunking approach already taken by xarray.
  • Supports lazy loading of array values, this would fit in well with the loading policies of the other xarray backends.

Cons:

• Would require xarray to adopt conventions for projecting arrays, naming (coordinates, dimensions, arrays), and handling of raster attributes. I can image ways this could be done but it may take us in down a path we don't want to go.
• The xarray backends generally return Datasets, however, rasterio returns individual arrays that would better be applied to DataArrays.

[fmaussion]

Hi @jhamman , I tend to agree with your doubts. I'll still comment on your cons:

To (1): I also think that xarray should avoid opening the projection can of worms. But the minimum things that xarray could do with rasterio is to read corner coordinates, dx and dy and define the two coordinates "x" and "y" out of it, without taking care of whether these are meters, degrees of arc or whatever. As long at the other rasterio file attributes are available as attribute of the DataArray or DataSet objects, users can do their own mixture

To (2): some geotiffs files also have more than one band. I don't know if these bands are named or have metadata, so maybe xarray will have to take decisions about these names too (most probably 1, 2, 3...).

I'll add a (3): rasterio depends on GDAL, which is huge and every now and then causes trouble on conda. This might also cause troubles to the continuous integration of xarray

Altogether this might be more complicated than worth it, but maybe the rasterio folks have interest in this and might provide more support.

If the idea for xarray accessors is implemented (#706 (comment)) this will allow more specific libraries like mine to do their own rasterio support at low cost.

[jhamman]

@fmaussion -

As for (1), I like your idea of leaving out the projection of the coordinates. That certainly makes things easier from the perspective of the backend.

A band dimension in (2) seems pretty manageable.

I'm not concerned about the GDAL dependency (3). I would love to see more robust conda support for GDAL but that's another issue. This would be an optional backend, similar to Pynio, which isn't broadly available on conda. We could sort out the CI issues.

So, if we took the simplest approach for implementing a rasterio backend, open_dataset would always return a Dataset with a single unprojected DataArray (name to be determined). The other big question is what to call the dimensions, since that is not explicitly provided in all raster formats.

[jhamman]

@fmaussion - Here's an example of the basic functionality I'm thinking of implementing: https://anaconda.org/jhamman/rasterio_to_xarray/notebook

A things to think about:

1. I've given each array the raster name. Does that make since? This allows us to return a Dataset instead of a DataArray.
2. Which attributes do we want to copy over from the rasterio dataset? It is not entirely clear which attributes in the rasterio._io.RasterReader object should become attrs.
3. I have not implemented lazy or windowed reading yet but it should be pretty straightforward using the window argument to src.read().

[fmaussion]

Hi @jhamman , this is close to how I would've done it, but I am maybe not the most qualified (probably the gis specialists from rasterio would be more helpful). But still, a couple of remarks from my side:

• I wouldn't necessarily do the try_to_get_latlon_coords systematically. When the raster coords are lat-lon, the new coords are redundant. And when the coords are x-y, the lat-lon info are only partly useful (since the grid will be unstructured in lat-lon). Furthermore, I am not sure if +init=EPSG:4326 is the only lat-lon proj available (there are surely more - at least if you leave the wgs-84 area)
• as mentioned by perrygeo in your rasterio post, the data model of geotiffs is not always clear. The pixel coordinates are very likely to be at the top-left corner of the pixel (as I assume in my small salem library). Most netcdf datasets we are using in the meteo/climate community are pixel-centered. I don't know if this is something that xarray wants to consider, but this becomes important if you want to make accurate projections. (in practice, the two concepts are equivalent for most applications, but you have to know what is what: in my small library I called those representations center_grid and corner_grid: https://github.com/fmaussion/salem/blob/master/salem/gis.py#L101 )

To your questions:

1. I agree that returning a dataset is a good idea. I don't know if raster is a good name, but I have no other idea right now
2. I don't know. The projection was always enough for me 😳

[jhamman]

As for 1) I'm open to having more discussion on decoding the coordinates. My contention here is that are useful, even in their unstructured format, since it permits visualization out of the box. I'll ping @perrygeo for more on this.

2. I don't really want to get into this because there isn't a standard treatment in geotiffs so it would, at best, be a guess on our end.

[perrygeo]

@fmaussion @jhamman re: projecting coordinates to lat-lng.

If you consider the raster cells as independent points, you can project them independently but they will likely not be regularly spaced. With few exceptions, if you need to maintain a regular grid, transforming data between projections will alter the shape of the array and require resampling (GDAL and rasterio call the process "warping" to reflect this). There are decisions and tradeoffs to be considered with the various resampling methods, selecting new extents and cell sizes, etc so it's typically not something you want to do on-the-fly for analyses.

I think keeping the xarray coordinates as generic cartesian x-y makes sense, at least initially. Even in many GIS tools, analysis is done on a naive 2D plane and it's assumed that the inputs are of the same projection. I'd recommend doing any reprojection outside of xarray as a pre-processing step (with e.g. gdalwarp or rio warp).

[shoyer]

Because each point can be computed separately, we could straightforwardly add latitude/longitude as lazily computed 2D arrays (under "coordinates"), similarly to how we currently handle on-the-fly data rescaling.

[jhamman]

@shoyer - that's what I was thinking too. In fact, that's more or less what I did in this example, although this is a eager implementation: https://anaconda.org/jhamman/rasterio_to_xarray/notebook

[IamJeffG]

@jhamman Small observation from your notebook: It's not good to assume a negative y-step size. Rarely, I will come across a dataset that breaks convention with a positive y coordinate, meaning the first pixel is the lower-left corner, but at least the dataset is self-consistent. Rasterio works beautifully even with these black sheep, so we don't want an xarray reader to force the assumption.

[IamJeffG]

In a past life I made side library that wraps rasterio's API to take and return xarray.DataArrays. It provides IO/clip/warp/rasterize operations on DataArrays, which themselves are annotated with the CRS and affine transforms as attributes.

My most common use case was reading disparate rasters and aligning them to the same grid:

1. Use rasterio to load separate spatial rasters over roughly the same area; let's say one is 30-meter satellite and one is 3-meter agricultural yield. Often I'll immediately wrap them in an xarray.DataArray and persist the CRS and affine transform as attributes.

2. Clip the fine-resolution yield array to my area of interest. I can either use rasterio.read(window=(...)) when reading or xarray.sel(x=slice(...), y=slice(...)) post-hoc.

3. I want to overlay the two arrays in a single xarray.Dataset. They may or may not have the same projection but definitely do not the same grid. I'll leave the fine-resolution data untouched to avoid needless resampling, but instead upsample and re-align the coarse-resolution satellite array to match the affine transform of the clipped fine-resolution array. Behind-the-scenes this looks like:

   # agriculture and satellite are both 2D DataArrays.
   satellite_10m = np.zeros(agriculture.shape, dtype=satellite.dtype)
   rasterio.warp.transform(
       satellite.values,
       satellite_10m,
       src_transform=satellite.attrs['transform'],
       src_crs=satellite.attrs['crs'],
       dst_transform=agriculture.attrs['transform'],
       dst_crs=agriculture.attrs['crs'])
   # This forces realization of the dataset in memory. I don't do much out-of-core.
   

4. Stack them into the same Dataset, sharing the transform and coordinate variables from the fine-resolution array.

5. Now use xarray to do cool compuations on the aligned datasets.

Reprojecting or clipping after reading xarray, like I do, goes against @perrygeo's recommendation. So maybe my example is moot, but I really like being able to do this programmatically in python, not CLI.

Even if xarray's new rasterio backend only provides a reader (and not rasterio.warp or rasterio.features functions), Step 3 shows it's very useful if, in addition to the data, the reader will expose the CRS and affine transform objects to the client.

However, if you both expose the transform and realize the coordinate variables, it's possible for them to diverge as the single source of truth. In my above workflow, anytime I clip (step 2) or warp (step 3) data, my side library needed to manually re-set that DataArray's transform and coordinate variables. (This is surely out of scope for rasterio or xarray!)