York Region data projectCan a model guess how tall a Richmond Hill street tree is from its trunk and its age?
York Region publishes a record for each street tree it owns or maintains along its regional roads, including in Richmond Hill. Our students download the Richmond Hill rows and teach a computer to guess each tree’s height band. All figures in this section come from our own code, not from the Region.
Our results on York Region Street Trees, MUNICIPALITY = Richmond Hill (9,505 rows fetched 22 October 2026; 9,398 with a height band; 7,518 trees to learn from, 1,880 held back to test)| Model | How it guesses a test tree’s height band | Share of the 1,880 test trees it gets right |
|---|
| Baseline | Always says “2 - 5 metres”, the most common band in the training trees | 63.9 per cent |
| 1 nearest neighbour | Copies the band of the single most similar training tree by trunk diameter and planting year | 72.3 per cent |
| 15 nearest neighbours | Takes a vote among the 15 most similar training trees | 76.8 per cent |
| 101 nearest neighbours | Takes a vote among the 101 most similar training trees | 74.5 per cent |
The program below is the whole exercise in standard-library Python. It asks the Region’s map service for the Richmond Hill rows 2,000 at a time, keeps the three fields it needs, shuffles the trees with a fixed seed so the result can be repeated, holds back one tree in five for testing, and scores two models. Run on 22 October 2026, it printed the baseline and the 15-neighbour figures in the table:import json, random, urllib.request, urllib.parse
from collections import Counter
URL = ("https://ww8.yorkmaps.ca/arcgis/rest/services/OpenData/"
"Biodiversity/MapServer/0/query?")
ask = {"where": "MUNICIPALITY='Richmond Hill'", "returnGeometry": "false",
"outFields": "CURRENTDBH,YEARPLANTED,TREEHEIGHT", "f": "json",
"orderByFields": "OBJECTID", "resultRecordCount": 2000}
trees, ask["resultOffset"] = [], 0
while True:
reply = json.loads(urllib.request.urlopen(URL + urllib.parse.urlencode(ask)).read())
if not reply["features"]:
break
for f in reply["features"]:
a = f["attributes"]
if a["TREEHEIGHT"]:
trees.append((int(a["CURRENTDBH"]), a["YEARPLANTED"], a["TREEHEIGHT"]))
ask["resultOffset"] += len(reply["features"])
random.seed(2026)
random.shuffle(trees)
cut = len(trees) * 4 // 5
train, test = trees[:cut], trees[cut:]
def guess(dbh, year, k=15):
near = sorted(train, key=lambda t: ((t[0] - dbh) / 10) ** 2 + ((t[1] - year) / 10) ** 2)
return Counter(t[2] for t in near[:k]).most_common(1)[0][0]
common = Counter(t[2] for t in train).most_common(1)[0][0]
base = sum(t[2] == common for t in test) / len(test)
knn = sum(guess(t[0], t[1]) == t[2] for t in test) / len(test)
print(len(trees), "trees,", len(train), "train,", len(test), "test")
print("always", repr(common), "->", round(100 * base, 1), "% right")
print("15 nearest neighbours ->", round(100 * knn, 1), "% right")
1. Clean what arrives as text
The diameter field, CURRENTDBH, is stored as text, so every value comes back as a string like “5” and has to be turned into a number. The height is not a number at all but a band such as “2 - 5 metres”, which makes this a classification problem rather than a regression.
2. Beat the lazy answer
Because nearly two thirds of the test trees sit in one band, a model that always says “2 - 5 metres” already scores 63.9 per cent. Any model worth keeping has to clear that line, and the 15-neighbour vote clears it by about thirteen points.
3. Choose k, then check it
One neighbour copies noise from a single odd record; 101 neighbours blur real differences. Fifteen did best on this split. Students then reshuffle with a new seed to see whether fifteen still wins or whether that was luck.
Where the model goes wrong
Accuracy hides who gets ignored. Of the 57 test trees recorded as “0 - 2 metres”, the 15-neighbour model got only 16 right and called 38 of them “2 - 5 metres”. Of 150 trees in the “10 - 15 metres” band it got 61 right and placed 70 one band lower. Small classes get outvoted by big ones. That is a real lesson in machine learning, not a bug in the student’s code, and it is why we always print the breakdown beside the headline number.
Three more honest notes. A diameter of 5 appears on 2,816 of the 9,505 rows, mostly on trees planted in recent years, and a handful of rows record round values such as 100; the field does not state its unit, so we do not guess one. Planting years run from 1920 to 2026, with 170 rows planted in 2026 itself. And the file covers only the trees the Region owns or maintains: its own description says it holds street trees “within regional road right-of-ways”, with only “partial data available for Regional owned street trees in rural areas”, so it is not a census of every tree in the city. The commonest species in the Richmond Hill rows is the Ivory Silk Lilac (925 rows), ahead of the Kentucky Coffee Tree (673), and Bayview Avenue carries more rows than any other road (1,751). We use it under the York Region Open Data Licence, whose terms say users may “Use the Information commercially”; the Region suggests the credit “Contains public sector information made available under The Regional Municipality of York’s Open Data Licence”. The City of Richmond Hill’s own map layers carry no reuse licence, which is why we used the regional file. Nothing on this page is endorsed by York Region.
