Vegetation-Outlook-model

About Model

-------------- github.com/SoulVisionCreations/ndvi-forecasting-advisory -------------- NDVI Forecasting (Prithvi + TFT) is a satellite-driven model that forecasts vegetation greenness (NDVI) about three months ahead at locations across India, expressed as a relative vegetation-condition / drought early-warning signal. TASK Given ~12 months of fortnightly history at a location, predict NDVI (≈ 3 months) ahead. ARCHITECTURE (one self-contained bundle, ~335M params, 757 tensors) • Image encoder: NASA–IBM Prithvi-EO-2.0-300M geospatial foundation model (Vision Transformer, MAE-pretrained on HLS imagery), FROZEN (330M params), adapted with LoRA (rank 16, α 32) on the q/k/v projections of all 24 encoder blocks. • A pre-projection LayerNorm + linear 1024→32 projector compresses the image embedding into a 32-d "image context". • Time-series forecaster: a Temporal Fusion Transformer (ProTFT_Elite) over a 24-fortnight window, fused with the image context and static features. Only ~4.0M params are trainable (TFT 2.40M + LoRA 1.57M + projector 0.035M). INPUTS Dynamic NDVI + weather/land (precipitation, temperature, radiation, evapotranspiration, runoff, soil moisture); static per-MWS profile (seasonal NDVI baselines, SPEI-3 sensitivity) + administrative categoricals (state / district / tehsil); one quarterly HLS composite tile. At inference these are fetched live per (lat, lon, date) from Google Earth Engine + CoreStack (admin) + Open-Meteo (forecast weather). OUTPUT forecast_ndvi at the target date, plus a relative vegetation condition — a z-score versus the location's own history for that time of year, mapped to three classes: Below / Near / Above normal. SELF-CONTAINED The frozen Prithvi backbone is baked into the bundle (byte-identical to the public base) — no separate base-model download or Hugging Face fetch at inference. Bundle files: tft_temporal_production_ft.pt, standard_scaler_*.pkl, label_encoders_*.pkl, train_config.json, mws_static_lookup_UNSCALED.tsv, prithvi_mae.py, config.json. USAGE A CLI and a FastAPI server share one inference path (point both at a weights/ folder). INTENDED USE Research and decision-support for vegetation & drought early-warning in India at the micro-watershed scale, ~3-month horizon; and backtesting past forecasts. OUT OF SCOPE: future dates as input, regions outside India, absolute greenness thresholds, sub-fortnight / single-pixel precision, or use as the sole basis for high-stakes decisions without human review. LICENSE & ATTRIBUTION Apache-2.0. Derivative of Prithvi-EO-2.0-300M (NASA & IBM, Apache-2.0; arXiv:2412.02732) — base weights frozen, not redistributed separately. Data / services: MODIS, HLS, SMAP (NASA); ERA5-Land (ECMWF / Copernicus); CoreStack; Open-Meteo; Google Earth Engine. See NOTICE for full attribution. RELATED ASSETS • Dataset: NDVI / SPEI fortnightly panel (final_spei_output.csv + location manifest), CC BY 4.0. • Toolkit / code: https://github.com/SoulVisionCreations/ndvi-forecasting-advisory • Use-Case: the Vegetation Outlook advisory — a farmer-facing layer that turns this forecast into an area-relative vegetation-condition opinion + stress-mitigation levers.