This isn't a simulation. This is every AI detection from every flight over San Francisco — 700+ flights across five neighborhoods, from December 2021 to April 2024.
The raw detection rate rose over time — but that alone can be misleading. Flight cadence changed materially, and when the gap between flights grows, more dumping can accumulate before the next pass. The more honest view is to compare the raw rate with a gap-normalized rate that accounts for days since the previous flight.
Three things moved at once: the model improved, routes became more targeted, and flight frequency changed. That third factor matters a lot. If a corridor is revisited after a long gap, more trash has had time to accumulate, which can inflate per-flight detection rate even without any model gain. The chart now shows both the raw rate and a gap-normalized rate so viewers can see that distinction directly.
Every detection has a GPS coordinate. 9,590 confirmed trash detections, each one placed on the map of Bayview-Hunters Point. The pattern is unmistakable — the same streets, the same corners, the same blocks.
The densest clusters aren't random. They correspond to specific infrastructure: dead-end streets, highway underpasses, vacant lots adjacent to commercial zones, and corners where DPW trucks have limited access. These aren't sites where people dump once — they're sites where people dump repeatedly, because the conditions that enable dumping persist. Detection identifies the pattern. Enforcement changes the conditions.
Detection without action is surveillance. Detection with action is infrastructure. Every confirmed dumpsite was automatically filed as a 311 cleanup request — no human intervention required.
Not every detection became a report. Only sites meeting three criteria: (1) large enough to warrant a crew dispatch, (2) not already reported in the prior 48 hours, and (3) on the public right-of-way, not touching buildings. Each report included GPS coordinates, timestamped aerial photo, and AI-classified waste type. Intelligent triage — not 311 spam.
Behind every data point is a flight. Some mornings, one quick pass. Some days, three separate flights — morning, midday, evening — covering different routes. The operational tempo tells its own story.
Every flight was planned, executed, and processed by a single operator. The drone launched from neighborhood streets. Photos were analyzed in real time. 311 reports were filed automatically — but only for sites that were large enough, not recently reported, and on the public right-of-way. No city department, no fleet of vehicles, no team of inspectors — one person with a drone and an AI model covered more ground than the city's entire enforcement apparatus.
This dataset is the proof. Not a demo, not a projection, not a pitch deck statistic. 298,800 photos. 700+ flights. 9,590 confirmed dumpsites. 4,801 cleanup requests — each one triaged for size, recency, and location before filing. All from five neighborhoods in one city.
Now imagine an entire city.