Longlist of new ideas

[jsmariegaard]

• ModelResult should be able to mikeio.Dataset/DataArray
• All filename arguments should take pathlib Paths
• Renaming of model and observation should be easier
• sel() method which can return new comparer with subset - then remove all similar arguments in scatter etc...
• save/load comparer to/from file (e.g. .nc)
• ntroduce lead_time and possibly other "categories"
• extract multiple points at-once
• Allow band-pass filtering (e.g. slow dynamics vs fast dynamics (peaks) )
• xcorr if UTC-offsets
• plots
  • colors: it would be nice to re-use plot colors across the package
  • ecdf plot
  • boxplot
  • Scatter with statistics
• lin_slope based on quantiles (more weight to extremes) (PDG suggestion)
• return residual quantiles (they are currently calculated in scatter but the numbers are not returned)
• Scale multi-variate score by std_dev (or similar) of obs - also relevant for tidal signals
• accept observations that are slightly outside domain (tolerance=xx) -> Modelresults
• Outlier detection using the difference to model
• rolling_skill()
• fix skill style
• more measurement types: raster, vertical profile, ...
• directions (circular statistics, suitable ticks on plots)

[Hendrik1987]

• Review naming logic for multi-variable observations (Is it needed and intuitive to also include the variable in observation name, e.g., HKNA_Hm0 here; Or is the name and 'ID'? Do we enforce that it has to be unique?)
• Attributes container
• Optional use of attributes for filtering
• Optional use of attributes in plots, e.g., station name, water depth etc. in 1-location-scatter plot (a la m_compare scatter plot)
• accommodate quality flags
• Observations from other file formats? (E.g., review workflow for CMEMS point observation. The .nc contains all meta data to create point observation.)
• collection of observations should be able to plot coverage and locations without having a model result
• collection of observations can be saved as .nc (?) without having a model result
• collections (with or without model results) can be reloaded and new model results easily be added.
  • I don't think the current con.add(Connector(o1, m1)) works as o1 will not be available anymore when reloading from .nc.
  • Suggest matching observation and model based on (t, x, y, z, variable_name).
• consider xr.dataset instead of df as container, e.g.:
  • Dimensions: record, model
  • Coordinates:
    • time(record)
    • x(record)
    • y(record)
    • z(record)
    • variable_name(record)
    • observation_name(record)
    • model_name(model)
  • Data vars:
    • observation_value(record)
    • observation_quality(record)
    • model_value(record, model)
  • conversion from/to the current df should be relatively simple
  • good attribute containers are available in the ds

[clemenscremer]

• allow unit correction (e.g. cm to m)
• add datum correction (offsets)

[ecomodeller]

Take a look at https://github.com/arviz-devs/arviz/blob/0a91fafca89237af681980aa994576873ef03283/arviz/data/inference_data.py#L90 which also uses xarray to store data. Relevant for ComparerCollection.