emmanuelamor5/nairobi-risk-mapping

MTL Nairobi is a deep learning model that serves as an evidence-based XAI tool for community groups and city planners to assess housing conditions and access to water and sanitation against UN-Habitat standards in Three major informal settlement regions of Kibera, Mathare and Mukuru in Nairobi. # Multi-Task Deep Learning for Just Urban Governance Parcel-level flood risk micro-zonation and structural deprivation assessment for Nairobi's informal settlements (Kibera, Mathare, Mukuru), combining an Attention U-Net (flood hazard) and GraphSAGE (structural deprivation) through a late-fusion policy decision engine, surfaced via an explainable-AI Folium dashboard. Strathmore University, School of Computing and Engineering Sciences — ICS Project II. Full methodology in the project proposal. ## Status **Sprint 1 (weeks 1-2): Data pipeline** — implemented. Acquires and validates all four data sources (Sentinel-2, Sentinel-1, SRTM/TWI, OSM vectors) for a given settlement and produces a tiled raster stack + a serialised PyTorch Geometric graph. `DataPipeline` is settlement-agnostic (pass `settlement_key="kibera" | "mathare" | "mukuru"`), and `run_sprint1_all_settlements()` runs all three in one call — see notebook section 5. Sprints 2-6 (model heads, fusion engine, dashboard) are not yet implemented — this repo currently covers the data layer only. ## Repo structure ``` src/ config.py # settlement boundaries, date ranges, thresholds data_pipeline.py # DataPipeline class -- Sprint 1 entrypoint utils/ gee_utils.py # Sentinel-2/1, SRTM/TWI acquisition (Google Earth Engine) graph_utils.py # OSM ingestion + proximity graph construction notebooks/ 01_sprint1_data_pipeline.ipynb # Colab notebook, runs the pipeline end-to-end tests/ test_graph_construction.py # unit tests for graph edge cases (no network needed) data/ # gitignored; populated at runtime ``` ## Setup ### Local (for editing / running tests) ```bash python -m venv .venv && source .venv/bin/activate pip install -r requirements.txt pytest tests/ ``` The unit tests run entirely offline against synthetic geometries — they don't need GEE or OSM credentials. ### Google Colab (for actually running the pipeline) Open `notebooks/01_sprint1_da …