During my ADDA CDDT-2 program, our team SkyGuard Environmental had three weeks to build something with real-world application. We chose Blantyre's illegal dumping problem. WasteWatch is what we built.

The Starting Point

We did not collect the data ourselves. Our foundation was the work of Mvula et al. (2025) — "Mapping Illegal Dumping Sites in a Low-Resource Region Using GIS and Remote Sensing" — a peer-reviewed study documenting 333 illegal dumpsites across all 21 wards of Blantyre, with GPS accuracy of ±3m and rich waste composition attributes. River proximity was the strongest predictor of illegal dumping. Model accuracy was validated at 89%.

Our job was to turn that research into something actionable.

The Problem We Were Solving

The real question was never "where are the dumpsites?" It was "which ones are most dangerous, and where should collection infrastructure go to have the greatest impact?" That is a prioritization problem, and it requires a system.

Risk Scoring

Health Risk accounts for 50% of the score, covering human waste, diapers, and disease vectors. Environmental risk accounts for 40%, covering river proximity, gullies, and open burning. Volume accounts for the remaining 10%.

Sites scoring above 80 are Critical. Above 60, High. A field officer can open the dashboard and know immediately where to go.

Skip Placement with DBSCAN

For infrastructure planning, we used DBSCAN clustering to identify zones where multiple dumpsites concentrate within 150m of each other. The centroid of each cluster becomes a recommended skip placement location. The cluster radius and minimum site count are adjustable, so planners can adapt recommendations to real budget constraints.

The Build

I handled the full technical architecture. The data pipeline runs in Python using GeoPandas and Scikit-learn, processing raw CSV survey data, converting UTM coordinates to WGS84, computing risk scores, and running clustering. The backend is FastAPI deployed on Render. The frontend is Next.js with Leaflet for the map layer, mobile-optimized for field use.

What I Took Away

Problem framing matters more than algorithm choice. DBSCAN is straightforward. The risk formula is not novel. What made WasteWatch useful was defining the right question before touching any code.

Academic research and engineering complement each other well. Mvula et al. provided the scientific foundation. We built the operational layer. That combination is something I want to keep doing.

Live Demo →