kellenmurerwa/flood-risk-hmm-rwanda

# Geospatial Flood-Risk Modelling using Climate Data and Hidden Markov Models in Rwanda **Capstone project — BSc Software Engineering, African Leadership University** Author: Kellen Murerwa · Supervisor: Emmanuel Adjei A reproducible machine-learning + Hidden Markov Model framework that estimates **daily Low / Moderate / High flood-pressure states** for the **Nyabugogo–Nyabarongo corridor in Kigali, Rwanda**, by fusing real rainfall, terrain, hydrology and urban-exposure data, and validates the result against the official flood-risk polygons. --- ## What it does 1. Builds a **250 m grid** over the study corridor (729 cells). 2. Pulls **real, free, keyless data** per cell-day, 2018–2024: - **Rainfall** (1/3/7/14-day totals) — Open-Meteo **ERA5** archive - **Elevation + slope** — Open-Meteo **SRTM** elevation - **Distance-to-river, road & building density** — **OpenStreetMap** (Overpass) - **Official flood-risk polygons** — Rwanda GeoPortal / **geodata.rw** (Ministry of Environment) 3. Derives **flood-pressure labels** (rainfall trigger × terrain susceptibility, with an unobserved drainage latent + noise so the task is not trivially circular). 4. Trains and benchmarks: rainfall-threshold & static-polygon baselines, Logistic Regression, Decision Tree, **Random Forest**, **XGBoost** (+ SHAP), and an **HMM** temporal layer — evaluated on a **temporal train/validation/test split** (train 2018–2022, validation 2023, test 2024). 5. Validates spatially against the official polygons and ships a **Streamlit dashboard**. ## Headline results (2024 test hold-out) | Model | Macro-F1 | High-recall | |---|---|---| | XGBoost (deployed) | **0.813** | 0.843 | | Random Forest | 0.813 | 0.849 | | Rainfall-threshold baseline | 0.626 | — | | Static-polygon baseline | 0.324 | — | Train/validation/test macro-F1 are near-identical for both ensembles (XGBoost 0.804 / 0.805 / 0.813; RF 0.813 / 0.806 / 0.813) — no over-fitting. Spatial validation vs official polygons: containment 0.30, * …