3,846 complete chains traced from the moment a drone detected illegal dumping to the moment DPW cleaned it up. Every stage timed. Every outcome tracked.
The median time from drone detection to filing a 311 report was 10.1 hours — that was the human review step. With full automation, that drops to minutes. The city's cleanup response (median 26.7 hours) was actually fast. The total pipeline — detection to clean streets — ran at 37.6 hours median. 57% of cases resolved within 48 hours.
But cleaning up a dumpsite once doesn't mean it stays clean. The real question: how many sites get dumped on again?
The same streets, the same corners, the same blocks — detected on up to 77 different days over two years. The city cleaned them. And they came back. And the city cleaned them again. Without persistent aerial monitoring, there's no way to prove the pattern, prioritize enforcement, or measure whether interventions are working. Detection doesn't just find dumpsites — it proves recurrence.
So what gets dumped? The detection system didn't just find trash — it classified it. Every 311 report included the waste type, enabling DPW to dispatch the right crew with the right equipment.
A mattress needs a flatbed. A paint spill needs hazmat. Loose debris needs a loader. By classifying waste types at the point of detection, dispatch can send the right crew the first time — reducing return trips and wasted labor hours. The data shows bagged waste gets cleaned fastest (17.9h median) while oil and liquid spills take 4× longer (106.8h) due to specialized response requirements.
The AI model improved because a human taught it. Every detection was reviewed. Corrections were fed back into training. The model got smarter with every flight.
That's not a failure — that's a success signal. "Unable to Locate" means DPW dispatched a crew and the trash was already gone — cleaned by community members, property owners, or a prior crew. The detection system is so sensitive it catches dumpsites that get resolved before the city even arrives. It's evidence the community cares.
| Primary Drone | DJI Mavic 3 Cine (L2D-20c) — 107,765 photos |
| Secondary Drone | DJI M3E (Enterprise) — 34,915 photos |
| Average Flight Altitude | 25.6m (84 ft) AGL |
| Optimal Altitude Band | 30–50m (100–165 ft) |
| Average Ground Speed | 31 mph (14 m/s) |
| Maximum Ground Speed | 48 mph (21.6 m/s) |
| Camera Angle | Nadir (−89° to −90°) — 99.9% of photos |
| Image Resolution | 5280 × 3956 pixels (20.9 megapixels) |
| Ground Sample Distance | ~0.6 cm/pixel at 30m altitude |
| Coverage Per Photo | ~32m × 24m at 30m altitude |
This is the complete operational picture. Detection, classification, reporting, response times, recurrence patterns, waste types, AI training, and hardware — from 13 database tables and 298,800 aerial photos.
One person. Two years. Five neighborhoods. Proof that it works.