Geography
14 cities, 1 country
Every frame was captured in China. A model trained only on it has never seen another country’s street layout, rooflines, vehicle mix or road markings.
Blind spot
Aerial models fail on the field, region, season or sensor they never trained on. Sparse finds those blind spots in your training data, has a licensed local operator fly exactly them, and delivers imagery with a training licence signed into every file. The report is free. You pay for the capture that closes the gap.
Get a free reportSee a sampleSend where and when your training imagery was captured, even a rough description. We map what your model has seen against where it has to work: geography, biome, season, altitude, sensor. And we check the licence on everything you trained on.
A licensed operator in the missing region flies your spec: resolution, sensor, overlap, revisit dates. Not a generic basemap. The scenes your model fails on, chosen with you.
Every file arrives with a signed provenance record: who flew it, under what credential, and the licence granted. Your compliance team, and your customer’s, can verify it offline. Try the verifier.
Recent papers describe one failure in different words: the model does not travel. These are paraphrases; each links to the paper.
Weed detectors trained on one field, crop or season perform poorly elsewhere, and recovering takes labelled data from the new field.
Existing infrared solar datasets are small and lack geographic diversity.
There is no globally distributed, multi-biome dataset for joint tree-cover and mortality segmentation.
Segmentation models drift across herbicide trials run at different sites and times.
Real thermal UAV training data is too scarce and too uniform, so teams synthesise it from RGB.
Few datasets cover sparse-canopy forest, and manual crown masks are costly.
This is what a team training on VisDrone would get from us. Everything in it is public and linked.
Every frame was captured in China. A model trained only on it has never seen another country’s street layout, rooflines, vehicle mix or road markings.
Blind spot
Non-commercial and ShareAlike, so a model trained on it carries that question into any product. Seven Hugging Face mirrors re-tag it with permissive licences, so a licence filter on the Hub will not catch it.
Cannot ship as-is
Captured over urban and suburban areas, all on one national road system. Farmland, rural roads, and Latin American, African and South Asian traffic and settlements are absent.
Narrow
No signed statement of who flew each frame, under what credential, or with what rights to grant. Diligence has nothing to check.
Unverifiable
From your metadata (place, date, altitude, sensor) and from the pixels themselves, embedded with a general vision model, we map where your training data is dense and where it is empty.
Short pilot flights, about 20 minutes each, in candidate regions. We run your own model on them. Where it breaks is the real blind spot, not the one anyone guessed.
The capture budget goes to the conditions that failed. One flight protocol, one variable changed at a time, so you can see what fixed the model. Twenty-five targeted images can beat thousands of the same field.
Every delivery reports how much new coverage it added to your data, alongside the signed licence record. Diversity becomes a number, not a promise.
| Option | Finds the gaps in your training data | Flies a new region to your spec | Training licence, provable per file |
|---|---|---|---|
| Open aerial datasets | No | No | Mostly non-commercial at fine resolution |
| Crewed aerial and satellite vendors | No | On their own collection plan | Contracts written for viewing and analysis |
| Drone capture networks | Not offered publicly | Yes | Not published |
| Sparse | Yes, free report | Yes, licensed local operators | Yes, signed record per delivery |
“Not offered publicly” and “not published” mean we could not find it in public material as of October 2026. If we have this wrong for anyone, tell us and we will correct it.
lat,lon,captured,source 52.374,4.889,2025-06-14,own flights 40.416,-3.703,2025-07-02,own flights -,-,-,VisDrone2019-DET -,-,-,"vendor orthos, NL 2024" needs to work in: Colombia, Brazil, Spain task: roof and solar panel detection
An example of a footprint that is enough. Nothing you send is shared; reports are written by a person.