Data Analysis

298,800 Photos, 5,106 Cleanup Requests: What Two Years of Aerial Detection Data Reveals

The most comprehensive drone-based illegal dumping dataset ever assembled in a US city — and what it tells us about how cities actually clean up.

March 2026  ·  Aerbits Research

Between March 2022 and April 2024, a single drone operator flew more than 700 flights over five San Francisco neighborhoods. The drone — a DJI Mavic 3 Cine equipped with a Hasselblad camera — captured 298,800 aerial photographs at nadir (straight down), each one analyzed by a custom AI model trained to detect illegal dumping.

The system identified 9,590 confirmed dumpsites. It filed 5,106 cleanup requests through the city's 311 system. It tracked every case from detection to resolution.

This article examines the complete dataset: 13 operational database tables covering detections, cleanup requests, service requests, training data, flight metadata, and community engagement. It is, to our knowledge, the most comprehensive public accounting of a drone-based illegal dumping detection program ever published in the United States.

---

The Pipeline: From Sky to Clean Street

Every detected dumpsite followed the same pipeline: the drone captured the image, the AI model classified it, a human reviewer confirmed the detection and filed a 311 cleanup request, and the city dispatched a crew. By linking detection records to cleanup requests to service requests, we can measure every stage of this pipeline.

Of the 5,106 cleanup requests filed, 3,846 have complete end-to-end data — from the timestamp the photo was captured to the timestamp the city marked the case closed. These 3,846 chains represent the most granular measurement of detection-to-cleanup performance available for any aerial monitoring program.

Pipeline Stage	Median Time	Performance
Detection → Report Filed	10.1 hours	93% under 24 hours
Report Filed → City Cleanup	26.7 hours	67% under 48 hours
Total: Detection → Clean	37.6 hours	72% under 72 hours

The pipeline performed consistently across the program's duration. Monthly median total-pipeline times ranged from 20 to 90 hours, with the longest delays occurring during high-volume months when the single operator was processing hundreds of detections simultaneously.

The system detected faster than one person could file reports. With automation, the constraint shifts to the city's cleanup capacity — where it belongs.

---

What Happened to Every Case

Of the 4,455 service requests tracked in the 311 system:

Outcome	Count	Share
Case Resolved	1,844	41.4%
Pickup Completed	455	10.2%
Duplicate (already reported)	189	4.2%
Unable to Locate	173	3.9%
Other closed statuses	1,220	27.4%
Still open	574	12.9%

The 189 "Duplicate" cases (4.2%) reflect the system's triage working as intended. The reporting criteria required that a site not have been reported within the prior 48 hours — but some sites were reported by residents through other channels between detection and filing.

---

The Recurrence Problem

This is perhaps the most important finding in the entire dataset.

Across two years of monitoring, the AI system identified trash at 661 unique locations (defined as distinct lat/lng coordinates rounded to three decimal places, representing approximately 111-meter zones). Of those:

62% of all dumpsite locations were detected on more than one day. Not the same piece of trash photographed twice — but the same location, cleaned up, and then dumped on again. And again. And again.

The worst location — near the intersection of Quesada Avenue and 3rd Street in Bayview-Hunters Point — was detected with fresh illegal dumping on 77 different days across the two-year monitoring period. The city cleaned it every time. And every time, it came back.

Days Detected	Number of Sites	Description
1 day only	252	One-time incidents (38%)
2–3 days	93	Occasional recurrence
4–5 days	52	Regular dumping
6–10 days	59	Chronic sites
11–20 days	59	Severe chronic sites
21–40 days	62	Persistent hotspots
41+ days	84	Entrenched dumping corridors

Cleanup without detection is a treadmill. The city was spending millions cleaning the same sites over and over — without the data to prove it was happening.

---

What Gets Dumped

Every 311 report included waste classification data from the AI model, enabling DPW to dispatch the appropriate crew and equipment. The breakdown reveals the composition of illegal dumping in San Francisco:

Waste Type	Reports	Share	Median Response
Loose garbage, debris & yard waste	3,160	62%	28.7h
Bagged/boxed/contained garbage	822	16%	17.9h
Overflowing city bins	210	4.1%	28.7h
Furniture	175	3.4%	22.0h
Mattresses	75	1.5%	19.0h
Blighted lots	36	0.7%	49.2h
Blighted buildings	32	0.6%	21.2h
Appliances	30	0.6%	16.0h
Oil, paint & liquid spills	30	0.6%	106.8h
Electronics	10	0.2%	—
Needles	4	0.1%	—

---

How the AI Learned

The detection model was not static. It improved continuously through a human-in-the-loop training pipeline: detections were reviewed, corrections were labeled, and updated training data was fed back into the model for retraining.

3,522 training datapoints were hand-labeled over the program's lifetime, with a 93% approval rate — meaning the AI's detections were overwhelmingly correct, and only 3 were explicitly rejected as false positives. The remaining 241 were pending review at the time of data export.

Training activity peaked in two cycles: September 2022 (802 datapoints) and November 2022 (908 datapoints). These correspond to major model retraining events that produced measurable improvements in detection accuracy.

The result was dramatic. Early flights (March–May 2022) detected trash in 1–3% of photos. By mid-2023, the mature model was detecting trash in 30–52% of targeted flight photos — a 15× improvement. This wasn't just the model getting better; it was also the flight paths being refined to focus on confirmed dumping corridors. The combination of a smarter model flying smarter routes is the core of the system's advantage.

---

Not Every Detection Became a Report

Of 9,590 confirmed trash detections, 5,106 became 311 cleanup requests — a conversion rate of 53%. This was by design. Not every detection warranted a city response. The reporting criteria were deliberate:

1. Size threshold: The site had to be large enough to justify dispatching a cleanup crew. Small litter that would be swept in routine maintenance was excluded.
2. 48-hour deduplication: If the same location had been reported within the prior 48 hours, no new report was filed. This prevented 311 spam and duplicate dispatch.
3. Public right-of-way only: Dumping on private property or touching building structures was excluded — those require different enforcement channels (code enforcement, not DPW cleanup).

---

The Hardware

The program operated with consumer and prosumer drone hardware — not military-grade or custom-built systems. This is significant: it demonstrates that effective aerial detection doesn't require specialized equipment.

Six drones were used over the program's lifetime. The primary aircraft — a DJI Mavic 3 Cine with the Hasselblad L2D-20c sensor — captured 107,765 of the 147,611 original photos (73%). The 4/3" CMOS sensor with its larger pixel pitch provided superior low-light performance and dynamic range compared to smaller-sensor alternatives, which proved important for early morning and overcast flights.

The optimal altitude band was 30–50 meters (100–165 feet), balancing ground sample distance (~0.6 cm/pixel at 30m, sufficient to identify a trash bag) against coverage area per photo (~32m × 24m footprint at 30m). At the average ground speed of 31 mph, each flight could cover approximately 1.5–2 linear miles of roadway.

---

Five Neighborhoods, One Operator

The program covered five San Francisco neighborhoods, with wildly different detection profiles:

Neighborhood	Photos	Trash Detected	Detection Rate
Bayview / Hunters Point	134,880	7,695	5.7%
Mission	74,506	662	0.9%
Hunters Point (south)	20,046	1,403	7.0%
Dogpatch	10,540	30	0.3%
Excelsior	3,995	43	1.1%

Bayview-Hunters Point accounted for 95% of all trash detections despite representing 52% of total photos. The contrast is stark: Dogpatch, just blocks away, had a detection rate of 0.3%. This isn't a citywide uniform problem — it is overwhelmingly concentrated in environmental justice communities.

---

Community Response

The program generated measurable community engagement beyond the detection data:

• 78 newsletter subscribers — 49 signed up in a single month (June 2022), corresponding to a community signature campaign supporting the program during the city's budget cycle.
• 75 ZIP code availability checks — residents checking whether aerial monitoring covered their area. 94124 (Bayview) led with 16 checks.
• 19 address-specific checks — residents entering their street address to see if their block was monitored.
• 9 drone pilot signups — licensed Part 107 pilots from across the country (Florida, Texas, Massachusetts, North Dakota, North Carolina, California) expressing interest in operating for the program.

The subscriber spike in June 2022 aligns with the community signature campaign that ultimately contributed to the program's inclusion in public policy discussions. The pilot interest from six different states signals market demand beyond San Francisco.

---

What the Data Proves

Two years. One operator. Five neighborhoods. 700+ flights. Here is what this dataset — the largest of its kind — actually proves:

1. The pipeline works. Median 37.6 hours from drone detection to clean streets, with 72% of cases resolved within 72 hours. This is faster than complaint-driven 311, which relies on residents reporting problems after they've already festered.
2. The city responds. SF DPW's median cleanup time of 26.7 hours after receiving a report demonstrates that the bottleneck isn't willingness to clean — it's knowing where to send the crews.
3. The same sites keep getting dumped on. 62% recurrence rate. 205 chronic sites. This is the strongest argument for persistent monitoring: one cleanup doesn't solve the problem.
4. Classification matters. Different waste types require different equipment and response protocols. AI classification at the point of detection enables smarter dispatch.
5. The AI improves with real data. 3,522 hand-labeled training points drove a 15× improvement in detection rate. Local data trains local models.
6. It scales. One person with a consumer drone covered five neighborhoods. A municipal contract with dedicated operators could cover an entire city.

The problem was never that cities don't care about illegal dumping. The problem was that they couldn't see it fast enough. Now they can.