Why street-level scenes, and not just land use
Most urban heat island research works from land-use categories and satellite thermal imagery. That combination explains the coarse pattern well — downtown is hotter than the park — but it says very little about what a street actually looks like at eye level: how much canopy hangs over it, how tall the walls are, what occupies the ground floor.
Those are precisely the things a planner can change. So the question we set out to answer was whether micro-scale scene characteristics, read directly off street-view imagery, carry real explanatory weight for surface temperature — and whether they can be combined with social vulnerability rather than analysed apart from it.
What we asked
- Can NYC's streets be sorted into meaningful urban scene types using street-view imagery?
- How do urban morphology, points of interest, and socio-economic features relate to urban heat island effects?
- Is heat exposure distributed unevenly across socio-economic contexts?
The models are deliberately interpretable. The goal was never to predict temperature as accurately as possible — it was to be able to say which attributes of a street are doing the work, in units a planner can act on.
Three findings worth remembering
Canopy has a threshold, not a slope. Once tree canopy coverage passes roughly 28%, land-surface temperature drops consistently. Below that, added greenery does much less. This is the kind of nonlinearity a linear model flattens away entirely — and it matters, because it turns "plant more trees" into a concrete target.
Height helps, and helps more with greenery. Each additional floor is associated with roughly 0.40 °C lower land-surface temperature, and the effect is strongest where built form and green elements occur together. Shade and vegetation are not competing strategies; they compound.
The burden is not shared evenly. High-poverty neighbourhoods with large minority populations face elevated heat exposure, driven by low vegetation cover and inadequate infrastructure. Heat risk in New York is not only a physical geography — it tracks the city's social geography as well.
The practical upshot: mitigation should be targeted rather than uniform. Greening investment aimed at underserved neighbourhoods — and pushed past the canopy threshold rather than sprinkled thinly — does more for both temperature and equity than the same budget spread evenly across the city.
What I'd take into the next study
Two things stand out. First, interpretability paid for itself: the 28% threshold is only visible because the model was built to be read, not just scored. Second, treating social vulnerability as a covariate rather than a separate follow-up analysis is what let the inequity result fall out of the same model — the two questions really are one question.