🚰 Automated GIS-ML System for Water Infrastructure Risk in Kenya

This project builds an automated, end-to-end GIS and machine learning pipeline to analyze water point infrastructure risk across Kenya, with a focused application in Nairobi County. Interact on Kepler . View on ArcGIS online:

---

Map Preview

---

1. Project Objective

The primary goal was to build a scalable, automated geospatial and machine learning workflow to assess water point infrastructure risks across Kenya, identifying patterns of non-functionality, underserved zones, and failure risk nationwide.

Nairobi County was designated as a high-priority focus area for:

• Targeted filtering of high-risk points
• Visualization of priority zones
• Actionable recommendations for urban WASH interventions

Key deliverables:

• Spatial SQL automation for service area coverage and clustering
• Predictive ML model for water point failure risk
• Interactive maps and GIS-ready exports
• Prioritization insights for rehabilitation and policy

2. Data Sources & Ingestion

Datasets:

• WPdx Kenya water points (CSV): 21,953 points with attributes (status_clean, install_year, water_tech_clean, pop_served_500m, is_urban, etc.) Time period of the dataset:01 January 2011 - 01 December 2024. Modified: 14 December 2025
• WorldPop 2020 population raster (ken_ppp_2020.tif): ~100 m gridded population counts
• GADM Kenya Admin Level 1 boundaries (shapefile): 47 counties, including Nairobi (NAME_1 = 'Nairobi')

Ingestion:

• Water points → PostGIS table water_points (GeoPandas + to_postgis)
• Population raster → population_raster (raster2pgsql, tiled, SRID 4326)
• Admin boundaries → admin_boundaries
• Nairobi subset created via ST_Intersects → water_points_nairobi (~11 points initially, expanded with buffer to ~25)

All data processed in PostgreSQL/PostGIS.

3. Spatial Analysis & Automation (Kenya-wide)

Nationwide workflows:

1. Service area coverage
   • PL/pgSQL function get_population_served: computes population within 500 m using ST_Buffer, ST_Clip, ST_SummaryStats
   • Added pop_served_500m to full water_points table

2. Clustering of non-functional points

   • Table non_functional_points (~15,232 rows)
   • ST_ClusterDBSCAN (eps ≈ 1 km, minpoints = 5) → cluster_id
   • Results: 10,273 clustered points; largest cluster = 196 points

3. Cluster summarization

   • Per-cluster: size, avg pop_served_500m, centroid
   • Exported as GeoJSON (clusters_points.geojson, clusters_centroids.geojson)

4. Machine Learning – Failure Risk Prediction (Kenya-wide)

Target: is_non_functional (1 = Non-Functional / Abandoned / dry season; 0 = Functional)

Final model: Hybrid XGBoost (combined features from iterative runs)

Features (30 after encoding):

• pop_served_500m, age (2026 – install_year), in_cluster, cluster_size, cluster_avg_pop, dist_to_functional
• One-hot encoded: water_source_clean, water_tech_clean, is_urban

Performance (test set):

• ROC AUC: 0.9735
• Accuracy: 0.92
• Non-functional class: Precision 0.96, Recall 0.91, F1-score 0.94

Top drivers (feature importance):

1. water_tech_clean_Motorized Pump - Electric (dominant)
2. in_cluster
3. cluster_avg_pop
4. water_tech_clean_Public Tapstand
5. water_source_clean_Protected Well
6. cluster_size
7. dist_to_functional
8. age

Feature Importance

Interpretation:

• Electric motorized pumps and public tapstands are the strongest failure predictors, likely due to power issues, high usage, and maintenance challenges.
• Spatial clustering (large clusters, high cluster population density) strongly elevates risk.
• Distance to nearest functional point and age are meaningful secondary factors.

Risk scoring:

• Failure probability computed for all points
• High-risk threshold (>85%) used nationally; Nairobi-specific filtering applied post-analysis

5. Nairobi County Focus & High-Risk Filtering

While the core analysis was conducted nationwide (to maximize data for clustering and modeling), Nairobi County was prioritized for targeted application:

• Total points in Nairobi approximate bounds (lon 36.60–37.10, lat -1.50 to -1.10): 259
• Risk score distribution in Nairobi: mean ~0.064, max ~0.623 (much lower than national high-risk areas)
• Top risk points in Nairobi: mostly functional motorized pumps (age 17–36 years), with max risk ~0.623

Nairobi Risk

High-risk filtering:

• Nationwide high-risk points (>50% threshold): ~13,014
• In Nairobi bounds (>50%): 2 points
• Final Nairobi high-risk GeoJSON exported (high_risk_nairobi.geojson) for visualization and ArcGIS Online publishing

Nairobi insight:

• Nairobi shows significantly lower predicted failure risk compared to national averages, likely due to better access, newer infrastructure, or urban maintenance advantages.
• The few high-risk points are associated with motorized pumps, consistent with the national top driver.

6. Visualization & Deliverables

Tools: leafmap / Kepler.gl
Key outputs:

• Clustered non-functional points (colored by cluster_id)
• Cluster centroids (size-scaled)
• High-risk points (nationwide + Nairobi subset) in red

Clusters_points&centroids

High-risk Points Nationwide

High-risk Points Nairobi

Exported files:

• clusters_points.geojson & clusters_centroids.geojson
• high_risk_points.geojson (nationwide)
• high_risk_nairobi.geojson (Nairobi-focused)
• master_water_risk_prioritization.csv (full dataset with risk scores)

7. Conclusion

The project delivered a robust national-level analysis of water point risks in Kenya, with Nairobi County highlighted as a high-priority urban case study.

Key findings:

• Electric motorized pumps and public tapstands are the most failure-prone technologies
• Large non-functional clusters (especially in high-population areas) are major underserved zones
• Predictive model (AUC 0.9735) effectively identifies risk drivers
• Nairobi shows lower overall risk (mean ~0.064, max ~0.623) compared to national hotspots, likely reflecting urban advantages