[POC]: Classification notebook

[adebardo]

Context

The purpose of this ticket is to set up a notebook allowing demcompare users to perform metric measurements, such as mean or NMAD, in specific areas using classification. This notebook can be included in examples and used as a tutorial notebook during the practical session.

Setup

For the tutorial, it would be helpful to use demcompare data along with the associated masks.

• Proposed data:
  • DEM
  • Mask

Information

• The masks will be treated as rasters using the GeoUtils tool. mask = gu.Raster('xdem_gironde_mask.tif')
• The masks are binary masks stacked in different bands --> TODO: Convert the demcompare mask into this format.

Notebook Workflow

1. Opening the DEM
2. Opening the mask
3. Converting the mask into 4 Boolean masks
4. Adding a function to select one of the 4 masks and apply it to the DEM dem.set_mask(mask_single_band_boolean)
5. Providing a set of plots (slope, hillshade, etc.) to highlight the use of classification
6. Providing a set of metrics (NMAD, mean, etc.) to highlight the use of classification

Extras parts

In demcompare, it is possible to measure metrics based on grouped information; it would be useful to demonstrate in the notebook how to do this using the xdem API.

Examples:

• Calculating the NMAD for different slope intervals (here, consider dividing the slope into masks according to the intervals)
• Merging a segmentation mask with slope intervals to measure metrics

/estimation 4d

[rhugonnet]

Here's the discussion I was mentioning about scalable classif (with raster-vector) extraction: https://discourse.pangeo.io/t/advice-for-scalable-raster-vector-extraction/4129.
If scalability becomes important, we probably want to either let geocube figure this out and use it in GeoUtils (via Xarray accessor), or use/inspire ourselves from rasterstats.

[rhugonnet]

On the edited description: Sounds good!

One remark on the extras:

In demcompare, it is possible to measure metrics based on grouped information; it would be useful to demonstrate in the notebook how to do this using the xdem API.

By "grouped information", you mean performing data binning? (I see the translation is "groupement par classes" in French, I didn't know!)
We do this a lot. We have a helper function that is used both for uncertainty analysis and coregistration: nd_binning (see https://xdem-rhugonnet.readthedocs.io/en/towards_0.1/gen_modules/xdem.spatialstats.nd_binning.html#xdem.spatialstats.nd_binning), mostly a wrapper of https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.binned_statistic_dd.html#scipy.stats.binned_statistic_dd returning the results in a pd.DataFrame with properly defined intervals, and allowing to do this in N-D. We also have plotting for 1D/2D: https://xdem-rhugonnet.readthedocs.io/en/towards_0.1/gen_modules/xdem.spatialstats.plot_1d_binning.html#xdem.spatialstats.plot_1d_binning.

On future improvements:

The limitation right now is that those functions expect arrays, and not raster. For easier use in the future, we were thinking of moving these tools to a geoutils.stats module, and adding a method to Raster for performing nd_binning on a raster, with a list of other variables that could also be rasters (which would be inherited by the DEM class; or Xarray accessor). See #378.

Once we have the Xarray accessors, we could also potentially rely on groupby through flox for computational performance, see: https://xarray.dev/.blog/flox.