Bmaina/vegetation-lstm-kenya

LSTM pipeline for vegetation forecasting and land degradation detection in Kenya using 6 years of Sentinel-2 NDVI (2019–2024). Covers Nairobi, Mt Kenya and Aberdares. PyTorch · Google Earth Engine · 6M ha classified into Recovering, Stable, Degrading and Drought-stressed. # 🌿 Vegetation Recovery Forecasting — Central Kenya ### LSTM Time Series Model · Sentinel-2 NDVI · Nairobi / Mt Kenya / Aberdares --- ## Overview This project builds an end-to-end deep learning pipeline to **forecast vegetation health and detect land degradation** across central Kenya using 6 years of monthly Sentinel-2 satellite imagery (2019–2024). The model learns the seasonal and multi-year patterns of vegetation from NDVI time series, then forecasts NDVI 3 months ahead for every pixel in the study area. Pixels where vegetation is declining faster than the model expects are flagged as **anomalies** — enabling early warning of drought stress and land degradation. **Study area:** Nairobi metropolitan region, Aberdare Range, and Mt Kenya ecosystem — covering approximately 6 million hectares of mixed urban, agricultural, and forested landscape. **Direct application:** Supports Kenya's national land restoration commitments, the African Forest Landscape Restoration Initiative (AFR100), and CIFOR-ICRAF's Regreening Africa programme. --- ## Why This Matters Kenya has committed to restoring **5.1 million hectares** of degraded land by 2030 under the Bonn Challenge. Tracking where restoration is working — and where land continues to degrade — requires satellite-based monitoring at scale. Traditional approaches compare two snapshots in time. This pipeline goes further by: - Learning the **expected** seasonal vegetation pattern for each pixel - Forecasting what NDVI **should** look like 3 months ahead - Flagging pixels where actual vegetation **diverges from forecast** as early warning signals - Classifying every pixel as **Recovering, Stable, Degrading, or Drought-stressed** This is the same methodology used by INPE (Brazil's deforestation agency) and NASA's SERVIR programme for operational land monitoring. --- ## Pipeline ``` Sentinel-2 Imagery (GEE) ↓ Monthly NDVI Composites 2019–2024 (72 time steps) ↓ Cloud Gap Filling + Savitzky-Golay Smoothing ↓ NDVI Ano …