lekejr/geospatial-flood-risk-ml

Machine learning for flood risk prediction in Lagos, Nigeria using open geospatial data # Geospatial Machine Learning for Flood Risk Mapping in Lagos, Nigeria A reproducible geospatial machine learning project that predicts flood risk across Lagos State, Nigeria, using open satellite data and historical flood records. Built entirely with free, open-access datasets and standard laptop computing resources. ## Research Question To what extent can machine learning models, trained on topographic, hydrological, and land-cover-derived geospatial variables, distinguish flood-prone from non-flood-prone areas in Lagos, and which environmental factors contribute most to elevated flood risk? ## Motivation Flooding is a major recurring hazard in Lagos, driven by its low-lying coastal topography, extensive lagoon system, and intense seasonal rainfall. This project explores whether a data-driven, machine learning approach — using only free, open satellite data — can produce a credible, interpretable flood risk assessment as a reproducible complement to traditional flood hazard mapping. ## Study Area Lagos State, Nigeria — selected for its well-documented flood history, strong open-data coverage via Google Earth Engine, and low-lying coastal geography that makes it a scientifically meaningful case study. ## Data | Variable | Source | Resolution | |---|---|---| | Elevation | SRTM 30m DEM (USGS) | 30m | | Slope | Derived from SRTM | 30m | | Rainfall (2023 annual) | CHIRPS Daily (UCSB) | ~5km | | NDVI | Sentinel-2 (ESA Copernicus) | 40m | | NDWI | Sentinel-2 (ESA Copernicus) | 40m | | Land cover | ESA WorldCover v200 | 40m | | Flood occurrence (target) | Global Flood Database (Tellman et al., 2021, *Nature*) | 40m | All data acquired programmatically via the Google Earth Engine Python API. ## Methodology 1. Acquired and clipped all layers to the Lagos boundary via Earth Engine 2. Resampled all predictors onto a common 40m grid 3. Built a full feature table (4,263,810 pixels; ~10% historically flooded) 4. Stratified subsample (80,000 pixels) for laptop-friendl …