This folder contains notebooks and scripts to extract training data, train an ML model and apply to some tiles to to produce a cultivated (A11) product

The functions to extract data for a polygon from ODC are within the dea-notebooks repo.

1. Collect training data as shapefile
   • Labels cover LCCS classes e.g. 111, 112, 228
2. Sample shapefile using data_sampler.ipynb to balance classes
3. Extract training data for each feature in shapefile - extract_data_for_shp_custom.py & deaclassification_tools
   • Per pixel should give a broader distribution instead of median per feature.
   • Currently using 6 geomedian bands and 3 mads bands and some indices
   • Can include [[phenology]] as a feature.
4. OPTIONAL: Join multiple training extraction outputs using:
   1. pr -mts' ' *.txt > training_data.txt
   2. Manually edit the hashtags out of the file
   3. awk '{$8="";print $0}' training_data.txt | sed 's/ / /' > training_data_trim.txt
      • awk removes the 8th column and sed deletes the double spaces left behind.
5. Visualise this data using data_visualiser.ipynb

Model training is conducted using train_ml_model.py. This saves a dictionary containing the model and features used to train. The parameters in this script were partially tuned using a grid search cv approach.

The notebook apply_ml_model_dc.ipynb loads the pickled model and applies to a number of study sites. The sites are defined in au_test_sites.yaml.

• Transfer model to sandbox: scp [email protected]:/g/data/r78/LCCS_Aberystwyth/training_data/cultivated/2015_merged_sample/model.pickle /home/jovyan/development/livingearth_australia/models/cultivated_sklearn_model.pickle
• Remember to export paths correctly
  • python3 setup.py install
  • export LE_LCCS_PLUGINS_PATH=/home/jovyan/development/livingearth_australia/le_plugins
  • export PYTHONPATH=/home/jovyan/dev/dea-notebooks/Scripts:/home/jovyan/development/livingearth_australia/
• Run python3 ../../livingearth_lccs_development_tests/scripts/run_test_sites.py -o master_branches --year 2015 --level 4