NDVI-forecasting-model-dataset

FILES • final_spei_output.csv — the numerical panel, one row per (micro-watershed, 14-day fortnight). ~5.4M rows, 26 columns. • all_mws_locations_dedup.csv — the location manifest: mws_id, lat, lon (~24,800 rows). COLUMN MANIFEST — final_spei_output.csv Identity / time • mws_id — micro-watershed id (join key to the location manifest) • latitude, longitude — MWS centroid coordinates • state, district, tehsil — administrative names (label-encoded categoricals in the model) • date — fortnight date (YYYY-MM-DD) • month — calendar month (seasonal feature) Vegetation (target + input) • NDVI — MODIS MOD13Q1 vegetation index, resampled to the 14-day grid Weather / land — ERA5-Land • total_precipitation_sum — precipitation • temperature_2m — 2 m air temperature • dewpoint_temperature_2m — 2 m dewpoint • u_component_of_wind_10m, v_component_of_wind_10m — 10 m wind components • surface_solar_radiation_downwards_sum — downward solar radiation • surface_net_thermal_radiation_sum — net thermal radiation • total_evaporation_sum — evapotranspiration • runoff_sum — runoff Soil moisture — SMAP • sm_surface — surface soil moisture • sm_rootzone — root-zone soil moisture Drought / derived • water_balance — precipitation − evapotranspiration (input to SPEI) • spei_3 — 3-month Standardized Precipitation-Evapotranspiration Index (≈ 7 fortnights) • spei_3_lag1 — SPEI-3 lagged one fortnight • sensitivity_spei3 — per-MWS NDVI sensitivity to SPEI-3 • baseline_ndvi_mean, baseline_ndvi_std — per-MWS seasonal NDVI baseline (mean / std) COLUMN MANIFEST — all_mws_locations_dedup.csv • mws_id — micro-watershed id • lat, lon — centroid coordinates (~24,800 rows; the model trained on the ~21,900 with usable history.) PROVENANCE • NDVI — MODIS MOD13Q1 (16-day, 250 m), resampled to the 14-day fortnightly grid. • Weather / land — ECMWF ERA5-Land (precipitation is ERA5, not CHIRPS). • Soil moisture — NASA SMAP. • SPEI-3 — computed from the precipitation / evapotranspiration series and joined in. • Numerical history ends in 2023. • Train-vs-serve delta (documented): the table was built with MODIS collection 006 + SMAP 007; live inference reads the current MODIS 061 + SMAP 008. HOW TO USE (train / fine-tune) Set DATA_PATH=final_spei_output.csv and MWS_CSV=all_mws_locations_dedup.csv. Full recipe (env vars + command, plus the two things not shipped — the ~755k Prithvi composite tiles and the base Prithvi-EO-2.0-300M checkpoint) is in docs/TRAINING.md in the toolkit repo: https://github.com/SoulVisionCreations/ndvi-forecasting-advisory LICENSE & ATTRIBUTION CC BY 4.0. Derived from MODIS (NASA), ERA5-Land (ECMWF / Copernicus), and SMAP (NASA); please credit those upstream providers and this project. See NOTICE in the toolkit repo.