Source: the NRCan-hosted Nominatim instance

One of this package’s seven data sources — the one source bound but not wired as a tier. vignette("data-sources") is the overview and puts it in context.

osm_geocode() queries https://maps.canada.ca/nominatim/search — the Nominatim instance Natural Resources Canada hosts over OpenStreetMap data on the Federal Geospatial Platform, and the one NRCan’s own aggregator queries internally under its nominatim service key.

It is not nominatim.openstreetmap.org, and that matters: the volunteer instance’s usage policy forbids bulk geocoding outright, so pointing this package at it would make every user of the package a violation. The NRCan instance is keyless, national, and under no such restriction.

osm_geocode() is exported. It is deliberately not a geocode() tier, and the whole of this vignette is about that decision.

library(cangeocode)

Unlike the other source vignettes here, the chunks below are not evaluated. There is no measured result to show, for the reason in the next section but one.

What this adds to the package

A geocoder that refuses, and a second data lineage.

The other three online services always answer. BC returns a locality centroid for an address that does not exist; NRCan’s geolocator returns a confident wrong street; Quebec’s returns a street centroid scoring higher than the correct civic point would. Every one of them needs a floor built on top of it to turn “the service responded” into “the address was found”.

Nominatim returns an empty array. When it has the street but not the number it returns the road itself at place_rank 26, which is a legible found the street, not the address rather than a point pretending to be a house. Both of the confident wrong answers this package uses as cautionary examples — 1 Rue Notre-Dame Ouest, Montreal placed 500 km away by the geolocator, 28 Silver St, Corner Brook placed on a different street — come back correct or refused here.

It is also the only source here that is not, directly or indirectly, a government address register. NAR, RQA and the BC and Quebec services all draw on official registers; where two of them agree, that can be shared lineage rather than confirmation. OSM is compiled differently. That would make it a genuine second opinion — if the licence allowed it to be used as one by default.

Licence: why this is not a tier

OpenStreetMap data is ODbL — attribution plus share-alike, with the obligation attaching to a derived database. NAR and the road network file are the Statistics Canada Open Licence, the BC geocoder and NRCan’s geolocator the Open Government Licence, RQA and Quebec’s geocoder CC-BY. Those compose freely. ODbL does not compose with them the same way.

A default tier would fold a handful of ODbL rows into a result table and change what the caller may do with the whole of it, silently, for the sake of a few rows the other tiers missed. That is not a decision a geocoding package should make on a user’s behalf.

So: the service’s own licence string rides along as osm_licence on every row, and reaching for this source is an explicit call.

osm_geocode("990 Bute St, Vancouver, BC")$osm_licence
#> "Data © OpenStreetMap contributors, ODbL 1.0. http://osm.org/copyright"

If the accuracy probe eventually shows it recovers a useful part of NAR’s tail, the decision to make is still about the licence, and it is the user’s to make per project rather than this package’s to make by default.

`uncertainty_m` is `NA`, and that is deliberate

osm_geocode(c("990 Bute St, Vancouver, BC", "99999 Nowhere Rd, Nowhereville, SK"))

nar_osm_uncertainty_m() returns NA_real_. Every other source here has a number attached to each precision class; this one does not, because it has not been measured against NAR. data-raw/probe_osm.R exists and runs over the same sample as data-raw/probe_geolocator.R, so the two services will be directly comparable — it has not been run at scale. Until it has there is no coverage figure, no p90, and no basis for placing this service anywhere in a method chain.

Inventing a plausible constant to make the row tidy would assert something unmeasured, and would be indistinguishable in the output from the figures that were measured. So the column stays NA.

What the first live runs showed

On a handful of addresses rather than a sample, so read these as observations and not as measurements:

Coverage is uneven by construction. OSM’s Canadian addresses are concentrated in cities that had a municipal open-data import. Downtown Vancouver, Montreal and Toronto all answer at building level; nothing rural has been tried. The number to watch when the probe runs is the answer rate — most of the loss here will be coverage, not rejection.
French word order matters, and is handled. 1 NOTRE-DAME RUE O returns nothing where 1 Rue Notre-Dame Ouest returns the address. This is the second place in the package where a query is spelled for a particular service rather than in NAR’s canonical form; Quebec’s locator is the other.

Two things about the response

The road is parsed and the municipality is not. Nominatim already separates house_number, road, city and ISO3166-2-lvl4, which is a real advantage over the geolocator’s single string. But road is a full street line (Bute Street, Rue Notre-Dame Ouest) and still has to go through normalize_address() to be compared component by component; the municipality fields are already components and go straight through. Parsing display_name instead would mean getting past a building name and two sub-municipal localities — The Berkeley, 990, Bute Street, Davie Village, West End, Vancouver, … — to reach what the service handed over separately.

The municipality is coalesced from city, town, village, municipality and hamlet, and deliberately stops there. suburb, neighbourhood, quarter and city_district sit below the municipality and would match a query’s municipality against something smaller than one.

n_matches counts distinct addresses, not results. 1155 Robson St, Vancouver comes back as two OSM objects — the building, and an office inside it 8 m away — with identical house number, road and city. Counting results would report an ambiguity that does not exist, so survivors are deduplicated on address_key() of the parsed answer.

The floor still exists

Refusing honestly does not remove the need for a check. Two conditions:

place_rank >= 30 and a house_number of its own. Rank 30 alone is not enough — 24 Sussex Dr, Ottawa comes back at rank 30 with no house number at all, so both halves are load-bearing.
The separated fields must agree with the address that was sent, through the same nar_address_agreement() the geolocator’s floor uses.

That comparison is shared code rather than per-service, because its rules — an empty string is absent and cannot contradict; the street name and civic number are what was asked, so a missing one is a failure; the municipality matches by whole-word containment in either direction — are properties of comparing two Canadian addresses, not of any one service. And the re-parse gets no con, so no gazetteer: a gazetteer would rewrite the answer back toward the question and launder the error the floor exists to catch.

Where the measurements live

file.show(system.file("notes", "nrcan-geolocator.md", package = "cangeocode"))

The geolocator note covers this service too — it is the geolocator’s own Nominatim sibling — and the OpenStreetMap section of geocoding-status.md records what remains unmeasured.