luforestal/prosopis-dryland-detection

Two-stage CNN pipeline for detecting invasive Prosopis juliflora in African drylands using Sentinel-2 phenology and VHR imagery # Prosopis Dryland Detection Two-stage remote sensing pipeline for mapping invasive *Prosopis juliflora* in African drylands. ## Problem *Prosopis juliflora* is a highly invasive tree species across semi-arid Africa (Ethiopia, Kenya, Sudan, Somalia). It is visually similar to native acacias, making manual identification difficult. However, a key ecological difference enables automated detection: **Prosopis is evergreen while most co-occurring acacias are drought-deciduous**, retaining leaves through the dry season. ## Detection Strategy ### Stage 1 — Phenological Screening (Sentinel-2) Filter candidate areas using dry-season NDVI differential: pixels with high NDVI during peak dry season that drop significantly in the wet season indicate evergreen shrubs unlikely to be native species. Output: probability map at 10 m resolution. ### Stage 2 — CNN Species Discrimination (VHR) Convolutional neural network trained on Very High Resolution imagery (Planet, Maxar, or WorldView) to confirm Prosopis vs. lookalikes (acacia, eucalyptus, irrigated vegetation) based on canopy texture, branching structure, and pod color. ## Repository Structure ``` prosopis-dryland-detection/ ├── data/ │ ├── raw/ # Downloaded occurrences (GBIF, iNaturalist), imagery │ └── processed/ # Filtered points, labeled patches, training sets ├── gee/ # Google Earth Engine scripts (JavaScript / Python API) ├── notebooks/ # Exploration and analysis notebooks ├── src/ # Python modules (data fetching, preprocessing, model) ├── models/ # Saved model weights and configs └── results/ └── figures/ # Maps, plots, validation outputs ``` ## Data Sources | Source | Purpose | |--------|---------| | GBIF / iNaturalist | Occurrence points for label generation | | Sentinel-2 (GEE) | Phenological NDVI time series — Stage 1 | | Planet / Maxar / WorldView | VHR patches for CNN training — Stage 2 | | Government field surveys | Validation ground truth (e …