This summer I built howsteep.org, a survey that measures road climb gradients from public LiDAR — every gradient stated with the distance it was measured over. It exists because of OpenStreetMap: every one of its ~2,900 published climbs, and the tens of thousands of roads measured behind them, is measured along an OSM centreline. No other dataset could have supplied that, in two countries, for free. So first: thank you.
This week the survey audited its whole New England corpus — 10,672 mapped routes — with an instrument that asks the USGS 3DEP point cloud a simple question: is there a physically graded road bed along this line? (A real road is a bench — a strip that stays cross-level while the hillside it’s cut into tilts. Nature doesn’t build those.)
Two results, and the bigger one is the positive:
The map is overwhelmingly real. About 4,600 roads came back physically confirmed — an independent, instrument-grade corroboration of OSM’s rural New England coverage — and thousands more were confirmed visually in NAIP imagery. Total phantom findings: 204, under 2% of what we audited. For a region famous for its TIGER-import legacy, that’s a data-quality result worth having on the record.
The give-back for the rest: a MapRoulette challenge — New England: mapped roads with no road on the ground — with 78 tasks so far, each carrying its evidence: Tier A (26) were verified by human review of laser + aerial evidence; Tier B (52) show open ground in NAIP with plainly no road corridor. Most look like paper roads, import artifacts, or driveways promoted to residential ways. If you check one and there IS a road, please mark it Not an Issue — that feedback flows back into the survey and improves the instrument. A third tier (126 point-cloud-only findings under canopy) is ready if the community finds the first two useful.
Workflow transparency: this is an AI-enabled project, and I want that on the table. The measurement pipeline itself is ordinary deterministic code — plane fits over ground returns, window maxima — but it was built and operated in collaboration with an AI assistant (Anthropic’s Claude), and one evidence tier uses a vision model to read aerial imagery panels (each task states its evidence class). No AI has edited OSM in this work and none will: every finding goes through human review — mine first, and now yours, which is exactly what this challenge asks. And every change to the survey itself traces to a human decision, made one of two ways: some rows I reviewed by hand, one at a time; many more fell under categorical policies I set after reviewing the evidence for the whole class — which the tooling then applied mechanically. The machine applies rules; it doesn’t make them.
Method, error model and per-climb open data (ODbL): howsteep.org/methodology.html. Corrections and scrutiny welcome — the survey retires its own published numbers in public when they fail review (26 pages went down yesterday), so we’re not precious about being wrong.