MarialRK/suddsight

Interpretable spatiotemporal flood-risk prediction system for the Sudd Wetland, South Sudan — real satellite data (CHIRPS, MODIS, SRTM), six ML models, SHAP explainability, and a full-stack decision-support dashboard for humanitarian early warning. # 🌊 SuddSight **Interpretable Flood-Risk Intelligence for the Sudd Wetland, South Sudan** SuddSight is a full-stack, explainable machine learning system that predicts flood risk across nine flood-prone counties in the Sudd Wetland region of South Sudan. It combines real multi-source satellite data with six trained ML models and SHAP-based explainability to give humanitarian responders transparent, actionable, 14-day early-warning intelligence. Built as a BSc Software Engineering capstone project at African Leadership University. --- ## Why This Exists South Sudan operates roughly 12 functional river gauges for a country of 644,000 km² — leaving most of the flood-prone Sudd Wetland unmonitored. Existing early warning relies on manual field reports that arrive 5–14 days after flooding begins, by which point evacuation and preparedness windows have closed. SuddSight addresses this by turning freely available satellite data into localized, explainable flood-risk forecasts — without requiring any new ground infrastructure. --- ## Key Features - **Real, verified satellite data** — not synthetic: CHIRPS rainfall, MODIS NDVI, SRTM elevation, and OCHA historical flood records, all acquired and cross-validated against known events (e.g. the 2024 South Sudan floods) - **Six trained ML models** compared head-to-head: XGBoost, Random Forest, SVM, LSTM, GRU, and an ARIMA statistical baseline - **SHAP explainability** — every prediction shows exactly which factors drove it and by how much, in plain language - **SuddSight Advisory** — combines model output with real upstream hydrology (Lake Victoria flood-wave transit context) for honest, non-overstated guidance - **Interactive map dashboard** — click any county to see live risk assessment - **Downloadable PDF situation reports** — for offline sharing with field teams - **Live activity log** — tracks assessments made over time, not just a single snapshot --- ## Model Performance All models evaluated with stratified 5-f …