Count the evidence,
not only the insects.
A three-city evidence atlas that carries visual-AI detections through sampling effort, model uncertainty and urban habitat context, towards a decision a city can defend.
Open the evidence lensPublic evidence confirms the cities and challenge, but not the workshop schema, sites, model or operational accuracy.
A detection becomes meaningful in relation.
An insect, a flowering plant, a time, a weather condition and a valid observation interval. The atlas must preserve that chain before it compares cities.
One pattern. One trust test. One decision.
The visual story is selected by available evidence, not by chart novelty. These three linked views produce the strongest defensible contribution in a three-day residency.
- 01Effort required
The observed pulse
Phenology ribbon · site map · camera-hour ledgerShow when and where unique confirmed encounters change per valid camera-hour, never raw frames as abundance.
- 02Labels required
Can we trust it?
Data-loss flow · blockwise error matrix · risk-coverage curveDecompose the pattern into uptime, filtering, abstention and human-reviewed performance across city, site, device and time.
- 03Delivery target
Where next?
Coverage gaps · habitat novelty · expected uncertainty reductionRecommend the next expert review, camera or conventional survey that is most informative under a visible objective.
Move from a point on a map to a claim worth making.
The visible points below are recent, research-grade iNaturalist context with CC0 or CC BY observation licences and open public coordinates. They are candidate-taxon priors and story material, not Numbering Nature detections or independent validation.
Detect. Locate. Compare. Act.
B++ makes insects visible. This contribution makes the resulting evidence comparable, inspectable and useful to a city team.
- 01
Detect
Retain confirmed, uncertain and rejected tracks with their model and configuration.
- 02
Locate
Join cameras to deployment, habitat and environmental context without exposing sensitive coordinates.
- 03
Compare
Normalise unique encounters by active sensor-hours, then block validation by city, site, device and time.
- 04
Act
Identify where another camera, human review or habitat intervention would reduce uncertainty most.
Recommend the next measurement, not the loudest pattern.
The first honest recommendation is already clear: obtain active sensor intervals, a pinned model bundle and a human-labelled holdout before making a cross-city ecological comparison. Once those gates pass, the same atlas can rank where extra sampling or habitat action reduces uncertainty most.