Measuring what nobody mapped

The atlas has always been a model of park quality. It reads a park through a Jane Jacobs lens and an ecological-comfort lens and produces a number. But quality and access are different questions, and the second one decides whether the first one matters to a given person. A park can be beautifully framed by mid-rise frontage, richly planted, well connected to three streetcar lines, and still be a place you cannot cross if you use a walker.

So we added a “Getting around” panel: what the walking paths are made of, how far it is to a bench, and how you would arrive by car. Then we spent most of the work deciding when not to show a number.

The gate

Start with seating, because it is the clearest case. We wanted to answer “how far must I walk before I can sit down”, so the pipeline samples the in-park path network every 25 metres and measures the distance from each point to the nearest mapped bench.

The problem is the word mapped. OpenStreetMap holds roughly 9,000 benches for the whole of Toronto. That is not a count of Toronto’s benches, it is a count of the benches somebody bothered to add. A park with two mapped benches would produce a confident median, and that median would describe how much mapping attention a neighbourhood has attracted rather than anything about the park.

So there is a gate. A park has to carry at least five mapped benches and at least one bench per 200 metres of path before we will publish a distance. Of 3,788parks, 3,460 fail that gate, which is 91 percent. They read “not enough mapped seating to assess”. Only 328 parks get a number.

That is a strange thing to ship. Nine out of ten park pages carry a sentence admitting we cannot answer. But the alternative is worse: a number attached to a caveat is read as a number, and the caveat is not read at all. A park that has never been surveyed would look identical to a park that was surveyed and found wanting, and only one of those is a finding about the park.

Elevation, where we published nothing at all

The signal we most wanted was slope. Grade along a path is the measurement that actually decides whether someone using a wheelchair, a walker, or a stroller can cross a park, far more than surface or seating. The plan was to report the share of path length over five percent and over eight percent.

There is no slope number anywhere on this site, because we could not measure it honestly.

The open global elevation models return integer metres. Sample a path every 15 metres and one metre of rounding invents a grade of 6.7 percent, sitting directly on top of the five percent threshold that matters. This is not a theoretical worry. We ran transects: flat ground in Downsview measured a maximum grade of 6.0 percent, produced entirely by rounding. The Christie Pits bowl, a real drop of roughly eight metres that anyone who has walked it can feel in their knees, came back as one metre of total relief, smoothed away. A coarser dataset was worse, putting that same flat Downsview ground at 17.9 percent.

So the cheap data is not merely imprecise, it is wrong in both directions at once: it fabricates slope where the ground is flat and erases it where the ground is steep. Small parks, where terrain matters most to a person on foot, are exactly where it fails hardest.

The right data exists. Natural Resources Canada publishes a high-resolution LiDAR terrain model at one-metre resolution, which would resolve a park bowl comfortably. We could not get values out of its public endpoints, and rather than ship a slope figure derived from a model that reads flat ground as a six percent grade, we shipped nothing. The measurements are written up in the project spec so the next attempt does not start over.

The parking lot that counts twice

One more piece of honesty, this time about a contradiction we chose to keep. A parking lot near a park counts against it in the vitality score, on the Jacobs reading that surface parking deadens a street edge instead of animating it. In the access layer, the same lot counts for the park, because parking is how a great many visits actually begin, especially for anyone who cannot easily walk or take transit.

Both readings are correct. We kept both rather than collapsing them into one number, because which one matters depends on who is asking. The methodology page sets out the thresholds and says so plainly.

What would change the answer

Two different things, and only one of them is technical. Terrain needs LiDAR, and that is a tractable engineering problem someone will solve. Seating needs people: the 91 percent is not a modelling failure but a mapping gap, and it closes one bench at a time, by somebody standing in a park and adding what is in front of them.

Until then the panel says what it knows and admits what it does not. An unmapped bench and an absent bench look identical in the data. The whole discipline of this layer is refusing to let them look identical on the page.