Denver Urban Tree Classification

I started with the most ambitious version of the question, which is telling one species of tree from another from the air. This is called species-level classification, and the appeal is obvious: if I could reliably tell an ash tree from an elm in an aerial image, a city could plan its field work from a desk instead of sending crews to look at every tree.

It did not really work. Across 10 species it reached only 17.0% accuracy, which is nowhere near reliable enough to map species or to stand in for a proper inventory.

So I changed the question. Instead of asking the imagery to name the exact species, I grouped the trees by how a forestry crew would actually treat them and asked it to sort trees into those groups. I put the ash species together because they all matter for emerald ash borer (EAB) planning, the elms together for Dutch elm disease risk, the maples together because of how much water they need, and the conifers together because they handle drought and heat differently from the broadleaf trees around them.

Asked that easier question, the same imagery did much better, reaching 38.3% accuracy. That is a 2.26× improvement over the species-level attempt, and enough of a jump to change what the result could actually be used for.

To be sure the improvement was real and not luck, I checked the confidence intervals, the ranges the true accuracy almost certainly falls within. They did not overlap:

• Species-level classification: 15.8–18.2%
• Management-class classification: 36.7–39.9%

Effect size: Cohen's κ = 0.293

I do not think this failure was something I could have fixed by tuning the model. I tried several different methods and saw the same confusion every time, on two separate imagery sources (NAIP aerial photos and Sentinel-2 satellite images). The real limit is the imagery itself. NAIP records only four bands of light (red, green, blue, and near-infrared), and from those four numbers closely related species simply look too much alike, overlapping by roughly 88–94%. There was not enough information in the pixels to tell them apart.

That is why the answer was not to keep hunting for a cleverer classifier. The useful move was to change what I asked the imagery to do, down to something it could actually deliver.

I would not use this output to sign off on treating, removing, or replacing any tree. What it is good for is shortening the list, so crews check a focused set of likely trees instead of walking the whole city.

That narrowing down is the one job I trust it to do.

Of the five groups, the ash result is the one I care about most, because ash is the group most exposed to a lethal pest.

Its recall was 49.1%, which means the model found about half of the actual ash trees. As a finished inventory that is weak, but for triage it is still worth having, because the two kinds of mistake do not cost the same. If the model wrongly flags a tree, someone wastes a short field visit. If it misses one, a mature ash can go unnoticed until it dies and becomes an expensive removal and replacement.

Because those two mistakes are so lopsided, I judged the model by what it would cost and save in the field, not by its accuracy score alone.

To show why a shorter field list matters, I put rough costs on it. Published work by Nowak and colleagues puts a ground survey at about $7 per tree. Denver has an estimated 300,000-tree inventory, so surveying every tree would cost roughly:

300,000 × $7 = $2,100,000

A targeted approach would instead start from the trees my classifier flags and send crews only to those, which changes the math:

Estimated survey savings: 75.7%

Benefit-cost ratio: 228:1

These are back-of-envelope numbers, meant to show the scale of the saving rather than to serve as a real city budget. The point I am making with them is small and specific: free imagery can shrink the field list before anyone spends time and money checking trees on the ground.

In the area I studied, ash made up 24.3% of the study canopy. That is a lot of trees resting on the fate of a single pest.

Emerald ash borer is a beetle that kills ash trees, and it has been spreading west. Colorado first confirmed it in Boulder County in 2013, and it reached the Denver metro area by 2022. In states where it arrived earlier, it has killed such a high share of the ash that a neighborhood with a lot of ash should be treated as carrying real risk.

If the ash share I measured in Hale held across Denver's estimated 300,000 trees, that would work out to:

24.3% × 300,000 trees = ~72,900 ash trees

I want to be clear that stretching one neighborhood's number across the whole city is rough. Hale is a single area, not a random sample of Denver, so I treat this as a screening figure, not a real count.

Even so, the number is big enough that it clearly warrants a closer look.

Losing trees is not only about the trees. It changes how hot a place gets, how much energy people spend on cooling, and how long the canopy takes to recover.

To put a number on that, I used the cautious low-end value from McPherson and colleagues for dry western cities:

$31/tree/year in energy-related cooling value

Applied to the ash trees at risk, that works out to:

• Hale study area: 751 ash trees × $31 = $23,281/year
• Citywide extrapolation: ~72,900 ash trees × $31 = $2.26M/year

At the high-end value of $98/tree/year, the citywide cooling value at risk rises to:

~$7.14M/year

These are not exact predictions for Denver. They are published low and high estimates applied to the ash share I measured, and I include them because they turn an abstract risk into a dollar range that planners and operations teams can actually reason about.

Replacing a mature tree is not cheap. At about $5,000 per tree, replacing the estimated ~72,900 ash trees would represent:

~$364.5M in replacement asset value

That is not a bill anyone expects to actually pay. I include it for the sense of scale, because mature canopy is costly to replace, slow to grow back, and spread unevenly across neighborhoods.

When the asset at stake is that large, catching problems early and prioritizing field visits is worth doing even with a model that is far from perfect.

Ash trees can be protected with a preventive injection called TreeAzin, usually cited at around $40–$80 per tree.

At its 49.1% ash recall, the workflow would point to roughly:

~35,762 ash trees

Estimated preventive treatment cost:

$1.4M–$2.9M

Estimated replacement cost if those trees were lost:

~$178.8M

Approximate treatment ROI versus replacement:

62:1 to 124:1

I would not let the model decide who gets treated. What it produces is a sensible inspection queue, and once those trees are checked in person the treatment decision goes back to the normal forestry process.

In this study, free four-band imagery was not good enough to identify urban tree species reliably.

The same imagery became useful once I asked it for management groups instead of species.

That shift gave a 2.26× improvement in accuracy, and it produced a triage workflow a city forestry team could realistically use.

Main takeaways:

1. Species-level classification from free multispectral imagery should be treated cautiously.
2. Management groups are a better fit for low-resolution public imagery.
3. Remote sensing can help prioritize field work.
4. Field verification is still required.
5. The useful product is a review queue for field crews rather than an automated inventory.