Derived Data: Explainable Rice Yield from Sentinel-1 and Sentinel-2 Satellite Data for Food Security

Derived Data comprise the final “modeling tables” produced by the end-to-end pipeline after (i) cleaning heterogeneous administrative/yield tables, (ii) aligning all sources to a shared 16-day temporal grid (bin16), (iii) joining on the composite key (region_id, bin16), and (iv) aggregating the resulting time series into annual features per (region_id, year). The derived feature space is phenology-aware: NDVI time-series are summarized into yearly mean/max and an integral measure, and phenological milestones (onset, peak, senescence) plus season length are extracted from the smoothed NDVI trajectory. Meteorological and reanalysis variables are converted into annual totals/means, with wind direction represented via sine/cosine embeddings to avoid angular discontinuities and precipitation stabilized using a log transform. For robust cross-regional learning, features are standardized per province using training-only statistics (preventing leakage into the 2024 temporal hold-out), a province indicator is included, and residual missing annual values are imputed with median-per-region values after aggregation. These derived tables are the inputs for model training/validation and for producing explainability artifacts and regency-level parity/error diagnostics in the 2024 inference stage.