Some important changes:
- Now seasons are not hardcoded, instead defined in a json file referenced in SEASONS_FILE environment variable.
- It is possible now to use any number of seasons or any date interval for using landcoverpy
- Added seasons example files
- Improved RAM usage when creating the tile dataframe, now pd.concat is not used.
- Beta windowing read-write in prediction is now available! Right now it has only been tested splitting the rasters in (5,5) slices, since there are problems with windowing different spatial resolution images, i.e. 20m 10m sentinel2-bands.
- An example for calling the predict workflow using windowing is
workflow(execution_mode=ExecutionMode.LAND_COVER_PREDICTION, window_slices=(5,5)). This reduces maximum RAM usage from 64GB to 8GB. - It is pending to improve the windowing allowing different window slices OR allow using use_block_windows.
- Workflow main script splitted in train predict to improve readibility