This release contains the code utilized within the methodology described in the WRI technical note "Spatial Characterization of Urban Land Use through Machine Learning: Mapping Urban Land Use in India and Mexico" (not yet published). Please see release branch README for full project description.
Some distinguishing features of this iteration of the methodology:
- Python 3
- Training imagery downloaded locally; application imagery processed in memory
- Localewise division of training & validation tranches
- Very large training/validation datasets (>10 million samples) made practicable via Keras
*_generatorfunctionality - On-the-fly construction of training/validation/application sets using sample catalog
- Mapping executed locally or on highly parallelized cloud computing infrastructure
- Automated imagery selection for model application
- "Ensemble" final LULC maps derived from multiple model outputs