AzadiGathui/environmental-degradation-monitor

Using machine learning to identify encroachment onto protected areas in Kenya # Environmental Degradation Monitor ## Project Overview A change-detection tool that uses satellite imagery and semantic segmentation to identify encroachment onto protected areas in Kenya. The system compares segmentation masks from two points in time, intersects detected changes with known protected area boundaries, and surfaces alerts where natural land has transitioned to built-up land within a protected zone. The initial scope targets a single county (Kajiado or Nairobi TBD). ## Scope **In scope (MVP)** - Binary or multi-class semantic segmentation of satellite imagery (built-up, vegetation, water, bare soil, cropland) - Change detection between timestamps - Intersection of detected change with protected area and riparian zone shapefiles - Visual output showing before/after masks and flagged encroachment zones - A simple browser-based interface for region selection **Out of scope (MVP)** - Real-time or automated satellite data ingestion - Integration with external workflows **Potential future extensions** - Temporal trend analysis across multiple timestamps - Confidence scoring and uncertainty quantification per alert - All-weather detection - Automated retraining pipeline to handle model drift - API integration ## Data Sources | Layer | Source | Notes | |---|---|---| | Satellite imagery | Copernicus / Sentinel-2 | Multispectral bands; two timestamps per target area | | Labeled training data | Radiant Earth LandCoverNet, ESA CCI Land Cover (Africa), RCMRD Geoportal | Kenya-specific chips preferred to reduce domain shift | | Protected areas | Existing shapefiles | Starting point; riparian zones approximated from water body proximity | | Riparian zones | Derived from river shapefiles + buffer | Approximate band on either side of water bodies | Training data selection must account for geographic leakage — chips from the target test county should be held out of training and validation entirely. ### Sources (Updated): - downloaded 4 lulc shapefiles from t …