Prepare the package for internal release

[yifatdzigan]

Including documentation, code refinement and unit tests

[fnattino]

Hi @rogerkuou, looks great!

I have run the tests, the notebook and the python scripts, it is nicely commented so it has been easy to follow the workflows. The package seems to me very well structured, and it seems easy to expand the NN models implemented.

Just let you here all my comments, but overall very nice work!

• [x] Very minor point: in setup.py there is a comma missing after geopandas in the dependency list, making installation with pip fail. I have also noticed that the dependencies are listed both in the setup.py file and in the requirements.txt . Maybe it would be nice to keep dependencies in only one place, so that one needs to update a single list whenever a new package needs to be added to the dependencies?
• [x] Also concerning installation: I have noticed that on Mac OS installing Cartopy via pip requires many of its dependencies to be installed ‘manually’ beforehand. The easiest way for me was to install Cartopy via conda, then install the repo via pip - is this an issue for other OSs as well? If this is the case, maybe some extra details should be mentioned in the README?
• [x] About the input data in the examples, and I think : most of the files are missing the extension. E.g. for the pickle data, maybe use .pickle to make clear the format?
• [ ] As a general comment, I have noticed that the GPI data are stored as pandas.Dataframe s. I think one could also work with geopandas’s GeoDataFrames, making explicit use of the lat/lon information that these data frames include. These could then be saved e.g. in GeoJSON format (which is also human-readable) instead of pickle binaries, even though maybe the latter is less space consuming. Using GeoDataFrames would make some of the geo-tasks easier, e.g. the intersection with shape files.
• [ ] In the plotting routines, scatter plots are employed to represent the data. Depending on the size of the figure and of the markers, some of the points might end up overlapping with each others, with the risk that the data corresponding to some small areas might result invisible. One might think to use something like imshow for matrix/raster representations, which might be particularly suitable if the data always lays on a regular grid.

[rogerkuou]

Merged all minor fixes to master. The initial release is available at: https://github.com/VegeWaterDynamics/ml_lsmodel_ascat