One of this package’s seven data sources — one of the four online
services. vignette("data-sources") is the overview and puts
it in context.
Quebec’s Ministère des Ressources naturelles et des Forêts runs an
Esri GeocodeServer over the Répertoire québécois des
adresses — the same register rqa_import() loads
offline, served live. Two locators sit behind it:
RQA_Adresse for addresses and RQA_Rue for
streets.
qc_geocode() is the client, "qc" is a
geocode() tier, qc_validate() re-asks it about
an answer you already have, and qc_reverse_geocode() is
the only online reverse geocoder in this package — the
other three services do forward only.
What this adds to the package
Reverse geocoding that is not NAR, and a Quebec fallback that needs no import.
The reverse endpoint is the distinctive one. Everything else here
answers “where is this address”; qc_reverse_geocode()
answers “what address is at this point” against a provincial register,
where reverse_geocode() answers it against NAR. In Quebec
those are close to the same question — see the warning below — but the
service will name an address NAR does not carry.
As a forward fallback it is worth about what NRCan’s geolocator is:
on 600 Quebec corporations addresses it takes coverage from 81.0% to
83.3%, recovering 12.3% of what NAR left unplaced. If you have run
rqa_import(), the offline "rqa" tier does more
than this and does it without the network — see
vignette("source-rqa"). The online tier is what you reach
for when you have not.
Licence
The register is CC-BY 4.0, the same terms
rqa_import() carries, and rqa_attribution() is
the string that discharges it:
rqa_attribution()
#> [1] "Contains information licensed under CC-BY 4.0 from the Repertoire quebecois des adresses, Ministere des Ressources naturelles et des Forets, Gouvernement du Quebec."That composes with the Statistics Canada Open Licence covering NAR,
which is why this can be a default tier where
osm_geocode()’s ODbL data cannot.
Using it
qc_geocode("1 Rue Notre-Dame Ouest, Montreal, QC") |>
select(match_method, uncertainty_m, qc_locator, qc_score, qc_address, lon, lat)
#> match_method uncertainty_m qc_locator qc_score qc_address lon
#> 1 qc_address 20 RQA_Adresse 82.51 1 Rue Notre-Dame Ouest, Montréal H2Y1S5 -73.55596
#> lat
#> 1 45.50617Reverse:
qc_reverse_geocode(-73.5555, 45.5019)
#> qc_address qc_postal qc_city qc_dist_m lon lat
#> 1 12-285 Place D'Youville, Montréal H2Y2A4 H2Y2A4 Montréal 11.71217 -73.55552 45.5018And as a validator against an answer you already have:
g <- geocode(c("1 Rue Notre-Dame Ouest, Montreal, QC",
"455 Rue Sainte-Catherine Est, Montreal, QC",
"1020 Rue Bleury, Montreal, QC",
"2 Cote du Palais, Quebec, QC",
"150 Boulevard Rene-Levesque Est, Quebec, QC"),
method = c("nar", "nar_interpolate"))
qc_validate(g) |>
select(input, match_method, qc_match_method, qc_score, qc_address, qc_dist_m)
#> input match_method qc_match_method qc_score
#> 1 1 Rue Notre-Dame Ouest, Montreal, QC nar_building qc_address 82.51
#> 2 455 Rue Sainte-Catherine Est, Montreal, QC nar_building none 94.12
#> 3 1020 Rue Bleury, Montreal, QC nar_building none 80.48
#> 4 2 Cote du Palais, Quebec, QC none none 94.29
#> 5 150 Boulevard Rene-Levesque Est, Quebec, QC nar_building qc_address 83.88
#> qc_address qc_dist_m
#> 1 1 Rue Notre-Dame Ouest, Montréal H2Y1S5 0.9305597
#> 2 Rue Sainte-Catherine Est, Montréal NA
#> 3 1020 Rue De Bleury, Montréal H2Z0B9 NA
#> 4 Côte du Palais, Québec NA
#> 5 150 Boulevard René-Lévesque Est, Québec G1R2B2 1.0469956Two of the five come back with a usable second reading, and the other
three are worth walking through, because between them they demonstrate
almost everything in the next section. Rows 2 and 4 got street
centroids — note the answer has no civic number in it — and row
3 got Rue De Bleury where Rue Bleury was asked
for, which is the same street under the name the register prefers. All
three were thrown out by the floor rather than by the service, which
reported no problem at all; the third is the floor being stricter than
it needs to be, which is the side to err on when the alternative is
silently accepting a wrong street.
And look at the scores. The two rejected street centroids scored 94.12 and 94.29; the two accepted civic points scored 82.51 and 83.88. The score is not a ranking, and this is what that looks like.
What to watch out for
It is not independent of NAR, so it is a fallback and not a check
The locators are named RQA_Adresse and
RQA_Rue. They serve the Répertoire — which is also what
NAR’s Quebec records are built from. Over 400 NAR Quebec building points
the median disagreement is 0.9 m, p90 13.3 m, nothing
over 500 metres; at scale, 2.5 million paired addresses agree to a
median of 21 cm.
Those numbers look like a spectacular validation and are nothing of
the kind. They are evidence of shared lineage.
qc_validate() says so in its own documentation, and the
small qc_dist_m in the example above is that effect, not a
confirmation. There is no second opinion available for Quebec
inside this package — British Columbia is the only province
where one exists.
The query has to be spelled French-canonical
This is the reason the tier is worth having at all, and it is not
cosmetic. The locator’s reference strings are
Rue Notre-Dame Ouest — type first, direction spelled out.
NAR’s own canonical form is NOTRE-DAME RUE O. Sent that,
the service does not degrade gracefully; it stops matching. On 400
Quebec addresses, by share resolved to a civic point:
| rendering | civic | street only | unmatched |
|---|---|---|---|
NOTRE-DAME RUE O — NAR’s canonical form |
31.5% | 4.0% | 64.5% |
| NAR order, direction spelled out | 58.0% | 3.0% | 39.0% |
| French order, direction spelled out | 58.8% | 14.2% | 27.0% |
| NAR order, type and direction spelled out | 95.0% | 3.5% | 1.5% |
| French order, type and direction — what the package sends | 95.5% | 3.5% | 1.0% |
The abbreviations are what break it; the word order barely matters. Spelling out the direction is worth 26 points, the street type another 37, the order under one.
The failure mode is the worst available:
1 RUE NOTRE-DAME O, MONTREAL comes back as a street
centroid scoring 92.4, where the correct civic point scores 82.5.
The abbreviated form does not lose the address — it replaces it with a
confident wrong answer several hundred metres away.
qc_geocode() re-renders every query for you, so you never
send the NAR form by accident.
Score is not a precision ranking
Correlation between the score and how far the answer landed from NAR’s building point is Spearman 0.018 — none. Worse, street-only answers score higher than civic ones: civic matches run 75.9–86.2 with a median of 83.0, street centroids 75.8–95.2 with a median of 87.0. The score measures how much of the string you sent was consumed, which a short street name consumes more completely than a long address.
So min_score defaults to 0 here and should
stay there. A threshold removes correct addresses before it removes
street centroids — the opposite of what min_score does for
the BC geocoder.
Loc_name is the precision field.
RQA_Adresse is an address, RQA_Rue is a
street, and match_method reports that as
qc_address or qc_street.
Addr_type is Feature for both and separates
nothing.
uncertainty_m here is not measured
20 m for qc_address, 500 m for qc_street.
Like the BC tiers’ figures, these are precision semantics turned into
deliberately pessimistic order-of-magnitude numbers, not error bars
comparable to the NAR tiers’.
The floor re-parses the answer, without the gazetteer
The service always returns something, so the package compares the returned title against what was asked for, component by component. That re-parse deliberately runs without the gazetteer — a gazetteer would resolve the returned string back toward the query and launder the very error the check exists to catch.
It earns its keep. 12 RUE SAINT-JEAN, GATINEAU comes
back as 12 Rue Saint-Jean-Bosco, Gatineau under
RQA_Adresse: right municipality, right civic number, wrong
street, and nothing in the service’s own fields says so.
Where the measurements live
file.show(system.file("notes", "quebec-addresses.md", package = "cangeocode"))data-raw/probe_qc.R reproduces the rendering and score
tables above.