Research · Track 02
SAAF — aquatic floatables
Measuring the waste that reaches the sea, so a clean-up can be aimed and then checked.
Why now
Context
Waste entering the ocean is almost never measured. Clean-ups get planned on impressions — a photograph, a volunteer's estimate, a small area counted by hand and extrapolated to a coastline — and without a baseline nobody can say whether the problem is growing or shrinking, or whether last year's effort did anything. Drone imagery and detection models are now cheap enough that a survey needs a consumer 4K drone rather than a research budget, which is what makes a repeatable count possible at all. This is also the clearest test of the whole IIDA argument — a problem too long for a startup to fund, too applied for a university to prioritise, and worth a great deal to whoever measures it first.
This is the furthest along of any track here, and the one that least resembles a research plan. SAAF already flies surveys, already has a model, and already publishes its numbers: 99.2% mAP@50 on unseen imagery, 68.2% at the strict mAP@50-95 threshold, about a second per 4K frame, on objects with a median size of sixty pixels. Six classes — glass, plastic bottles, cans, takeaway cups, retort pouches, food containers.
The gap between those two scores is the honest part, and it is where the research actually sits. The everyday score says the model finds the litter. The strict score says how well it agrees with a human about the edges of each piece, and every mass estimate depends on that agreement, because mass is derived from box dimensions converted through ground sample distance and matched to reference weights. SAAF’s own page says it plainly: it is an estimate, not a weighing. Closing that gap, and quantifying what remains, is the work.
What makes this worth running as a track rather than only as a product is the second deliverable. A benchmark and a method note are useful to everyone surveying a coastline, including people who will never use SAAF — and aerial litter detection has no public benchmark worth the name. That is an artefact the organisation can be judged on, which is the test every track here has to pass.
Sub-themes
Scope
- Detection and classification of litter in aerial imagery, at the sizes and altitudes a survey actually flies
- Mass and footprint estimation from box dimensions and ground sample distance — and the error bars that belong on both
- Survey design across tides and seasons, so repeat flights are comparable rather than merely repeated
- Hotspot and trend analysis that tells a municipality where to send a truck
- What a measurement has to look like before a donor will accept it as verification
Year one deliverables
Output
- A published detection benchmark on aerial litter imagery, with the evaluation protocol
- An open method note on mass estimation from aerial imagery, stating what the estimate is not
- Survey results from coastal, creek and inland river sites, repeated across at least two seasons
- A written account of where the model fails, at what altitude and in what water conditions
If the strict score stays where it is and mass estimates cannot be stated with honest error bars, this stops being measurement and becomes a counting exercise with a decimal point, and it closes. It also closes if surveys are flown but nobody — municipality, NGO or donor — changes a decision because of one. A number nobody acts on is not an outcome.