README_DATA_PROCESSING.md

Data Processing Runbook

This scaffold defines the intended processing stages before model training.

Stages

1. Build a Sentinel-2 image index with RGB, NIR, NDVI, cloud metadata, and source trace.
2. Build hourly environmental features from ERA5-Land, CHIRPS, SMAP, and optional Open-Meteo cross-checks.
3. Align multimodal samples using Sentinel-2 timestamp t0 as the anchor.
4. Generate traceable proxy stress labels in the [0, 1] range.
5. Create spatiotemporal splits that avoid zone-time leakage.
6. Package artifacts with checksums and validate the dataset contract.

No Training Boundary

Processing stops after dataset validation and versioning. Model training is out of scope for this phase.