A city can plant a thousand trees and still not know whether anyone sits beneath them. Greenery is easy to count; interaction is not. This project reads everyday photographs to recover something planners rarely measure — how people actually relate to urban nature, and how that relationship shifts across a decade.
16,276geo-tagged photos
2014-25twelve years
5nature types
3roles of nature
495gold-standard labels
3VLMs benchmarked
Choosing the model
Three readers, one chosen
Reading interaction reliably matters more than reading it cleverly. We scored three frontier vision-language models against 495 human-annotated images, using accuracy and macro-F1 for nature's role and joint exact-match across all four semantic labels. The most consistent model was selected to label the full decade.
selected
Qwen3-Omni-Flash
role accuracy 0.978 · macro-F1 0.947 · joint exact-match 0.903
GPT-4.1
joint exact-match 0.760 · less consistent on the rare Partner role
Gemini-2.5-Flash
joint exact-match 0.683 · weaker on edge cases
Benchmark: 495 gold-standard images · four semantic tasks. The selected model reached label-wise accuracy 0.972 and joint exact-match 0.903; the benchmark's own inter-rater reliability was substantial (Cohen's κ 0.86 for photographer engagement, 0.85 for nature's role). Errors are directional — Subject and Partner collapse toward Setting (54% / 27% of errors), so the model under-reports the richer relationships rather than inventing them.
How the model reads a scene
From a photograph to a structured reading
Each image is interpreted along a hierarchical framework. First the functional role of nature — is it the Setting for human activity, the Subject being admired, or a Partner in direct physical interaction? Then three finer dimensions: who is present, how active they are, and whether the photograph itself documents the encounter.
↑ Pick a photo to watch HNI-VLM resolve its labels — or launch the live model to try your own.
A decade of interaction · Vancouver 2014-2025
The anatomy of a decade
The full dataset is mapped at the top of this page — colour it by any dimension and drag the timeline through twelve years. Read together, the labels describe not just where people meet nature, but how.
Role of nature all 16,276 photos
Subject
72%
Setting
27%
Partner
1.3%
The photographer's stance all photos
Implied
72%
Documented
28%
Who is present the 4,535 photos with people
Group
58%
Individual
42%
How active the 4,534 photos with people
Moderate
61%
Sedentary
31%
Vigorous
8%
The decade, in numbers
Explore the readings
Over twelve years, the record tilts from inhabiting nature to observing it. Nature-as-subject climbs from about two-thirds of the photos to over four-fifths after 2019, while group outings and active use fall away after 2020 — what these images increasingly record is nature observed rather than inhabited. Switch the lens, switch the view: trace each category as a line across twelve years, or read its yearly composition as stacked bars. Click a legend item to isolate it.
You can see the decade in the explorer above. Whether what you see holds up — that's another question.
Spatial statistics
What shapes how a neighbourhood meets nature
The same readings then pass through spatial regression — at the level of each green feature, and of each neighbourhood (Dissemination Area). The sharper question is no longer where interaction is denser, but what shapes its mode: whether a place meets nature as a setting, a subject, or a partner. The answer is configuration — access, amenities, mixed-use vitality — more than the sheer quantity of greenery.
Which spaces draw interaction? feature-level OLS · ref: street trees in pocket spaces
Bigger, better-connected spaces draw the most interaction — yet hold it least densely, so it is distributed street greenery that concentrates encounters per hectare.
Reference: street trees in pocket spaces · points are standardised OLS coefficients with 95% CI · green = positive, red = negative, grey = not significant.
How much interaction? DA volume models · count & density (Appendix E)
Nature and access explain interaction volume well (count R² ≈ .45, density ≈ .37); the built environment adds a few points, socioeconomics almost nothing. The conventional, access-based result — useful, but it only says how much, not how.
Bars: model R² as predictors are added — Nature → + Built env → Full (Appendix E).
What shapes the mode of engagement? DA hierarchical models · logit shares · N = 585
Setting
adj R² 0.24 · most predictable
Subject
adj R² 0.13
Partner
adj R² 0.13
What a neighbourhood affords
Setting
Subject
Partner
Greenway access
▲▲▲
▲▲▲
–
Drinking fountains
▲▲▲
▲
▲▲
Recreational facilities
▲
–
▲▲▲
Commercial vitality
▲▲▲
–
▲
Mixed land use
–
▲
–
Building density
–
–
▼
Who lives there · modest but equity-relevant
More children
–
▲
–
More seniors
–
–
▲
More car commuting
▼▼
–
–
More non-official-language speakers
–
▼
–
Access beats acreage. The sheer quantity of greenery — street-tree density, food trees, community gardens, dog-off-leash areas — predicted no mode of interaction. What a place affords — greenway access, fountains, recreation, commerce, mixed use — shapes how nature is engaged; who lives there adds modest, equity-relevant signals.
▲ positive · ▼ negative · – not significant · more marks = stronger association · full hierarchical models, N = 585.
What we found
Conclusions
Primary
01
Observed, not inhabited
Across 16,276 photographs, people mainly observe nature rather than physically engage with it: 72% present nature as the main focus (Subject), 27% as a background to activity (Setting), and only 1.3% as direct physical contact (Partner). Since 2019, the observational share has risen above 80%. Hands-on, embodied contact is rare, and becoming rarer.
02
Form shapes interaction, not quantity
How people engage with a green space depends on its form, not the amount of vegetation. Type matters: water, street trees, and parks present nature as the main view (Subject, 71–79%), while sports grounds frame it as a background (Setting, 53% vs 45%). Size matters too: large, active spaces attract the most physical and group interaction but a lower density, while small pocket spaces are the most intensively used and the most varied.
03
Social context follows form
The social setting follows form as well. Solitary engagement concentrates in small, distributed greenery, while group activity gathers in parks, greenways, and large spaces. The form of the space shapes this as much as the visitor does.
Secondary
A vision–language model can read interaction from photographs. Applied to geotagged social-media images, the model labels each scene along the framework with high accuracy (0.978 for nature’s role, validated against a 495-image, two-annotator benchmark). This gives a scalable, repeatable way to measure human–nature interaction without surveys or fieldwork.
Large spaces gather visits; street trees distribute contact. Parks, greenways, and water attract far more visits, but spread across large areas, while street trees, distributed across the city, bring nature close in many places. (Type and size alone explain visit counts at R² 0.38 and density at R² 0.54.)
The observational shift is genuine, not an artefact of the data. The rise in observational photographs after 2019 matches independently documented pandemic changes in outdoor recreation, supporting the finding beyond platform-specific effects.
Errors run in one direction. When the model errs, it labels Subject as Setting in 54.5% of cases and Partner as Setting in 27.3%. It tends to miss the physical and relational modes, so their already small share is, if anything, an underestimate.
Why it matters
Significance
The implication is not that cities need more vegetation, but that the form of a space shapes how people engage with nature. This reframes the task for three groups.
Planning & Policy
Provide the full range
Each form supports mainly one kind of engagement: a district of only street trees offers mostly viewing, and one of only large parks mostly group visits. Planning should therefore combine types and sizes so that each neighbourhood supports the full range — observation, social gathering, physical contact, and solitude — and ensure that lower-income and more diverse areas are not limited to a single mode.
From how much green a place has, to which forms of engagement it supports, and for whom.
Design
Design the form to deliver
Form predicts engagement but does not guarantee it. A large waterfront supports physical contact only where the water is reachable and there is something to do; a pocket park draws its value from closeness and variety rather than size. Each space should be designed to support its intended mode, not merely to match a category on a plan.
Design for the intended mode, not the category on the map.
Management & Community
Activate the mode, with residents
The same space can function as a background, a focus, or a site of physical contact, so its mode is sustained — or changed — through programming, maintenance, and everyday use. Residents should help decide which forms of engagement their spaces support, rather than treating the mode as fixed by form alone.
The mode is managed and chosen, not fixed by form.
The philosophy
Life begins, and celebrates when human and nature interact and fuse.