Denver Urban Tree Classification: Mapping a Neighborhood's Canopy from Free Imagery

When I started, I thought species was the right target, so that is what I tried first. With a simple pixel-by-pixel approach across ten common species my accuracy sat near 17 percent, barely above guessing. NAIP, the US Department of Agriculture's aerial photography, carries only four color bands (red, green, blue, and near-infrared), and at the level of single pixels many species reflect those four bands almost identically. The fuller crown model built below recovers a good deal more of the species signal, as I come back to later, but for what a city actually does the species label is still not the target worth chasing.

I changed the question to match what the data can support and what a forester actually needs. Crews are not dispatched by Latin name. They act on a few practical categories: ash at risk from the Emerald Ash Borer, elm at risk from Dutch Elm Disease, water-demanding maples, drought-tolerant conifers, and everything else. By collapsing ten species into those five management classes, I turned a problem the imagery cannot solve into one it can.

The classes also separate over the seasons, which a single photograph cannot capture but a satellite that revisits the same spot can. That is why I built a five-date series from Sentinel-2, the European Space Agency's free satellite, across the 2025 growing season and measured greenness with NDVI (a standard index where higher values mean more living leaf). Each class moves through the year a little differently. Conifers hold their color into November while the deciduous classes leaf out and fade, and that motion gives my model a signal a midsummer snapshot would miss.

To classify a tree I first have to find it. Instead of labeling individual pixels I delineated whole tree crowns. I chose crowns because they keep each tree as one object, and because that follows how Cross (2019) approached tree mapping in Costa Rican forest. I found the crowns in a canopy height model, a raster of how tall the vegetation stands, which I built from DRCOG's 2020 airborne LiDAR (laser scanning that measures height directly). A watershed segmentation, masked to vegetation, split the canopy into 18,272 individual crowns.

For each crown I measured 109 features: color and near-infrared statistics from NAIP, the five-date Sentinel-2 series and its seasonal changes, several vegetation indices including the red-edge and water indices Cross found most useful for separating species, and height and shape from the LiDAR. I attached a known species to a crown wherever it fell within three meters of a tree in the city's arborist inventory, which gave 3,764 labeled crowns to learn from. I compared four classifiers by cross-validation and kept the gradient-boosted model because it came out ahead.

Before trusting a trained model I wanted to see whether the trees group on their own. I plotted every crown by greenness and height and let a clustering algorithm sort them with no labels at all, and the plot shows real structure in the data. It also shows why the job is hard. When I color the same plot by the true management classes, maple sits almost on top of the catch-all deciduous group, which is exactly where my trained model later makes most of its mistakes.

My model reaches 69.0 percent overall accuracy across the five classes, up from 38.3 percent for my earlier version that worked pixel by pixel, a gain of about 31 points. I also checked balanced accuracy, which averages the per-class scores so the large groups do not dominate, and it is 62.8 percent. The classes a forester cares most about score highest. Conifers and ash are recovered well, while maple is the weakest because it shares so much spectral and structural ground with the other deciduous trees.

My model leans most on color and season rather than on structure. The Sentinel-2 time series and the NAIP bands together carry close to nine tenths of the decision, and the LiDAR height about a tenth. So the result is mostly a question of reflectance, of how each canopy reflects light across the bands and across the year. That is the same premise Cross relied on with commercial satellite imagery in Costa Rica, and I carried it here to free public data and to an urban canopy.

I did not want to trust a single accuracy number from a random split, because it tends to flatter a map like this. Neighboring crowns are photographed under the same light and often belong to the same planting, so if some land in training and their neighbors in the test set, the model can look better than it really is. To get an honest figure I held out whole spatial blocks at a time, 200 and 400 meters across, so entire stretches of the neighborhood stayed out of training together.

Under that stricter test the accuracy settles around 66 percent, roughly two points below the random-split figure and steady across block sizes. The small gap tells me the map generalizes across the neighborhood rather than memorizing it. The one class that slips is maple, for the same reason it was weakest to begin with. This follows a habit I took from NASA and NOAA's remote-sensing training: give the model an honest grade, and do not assume a model trained on one place transfers to another.

Having built the model, I went back to the harder question and asked it to name the species directly. Across the twenty-one species with enough labeled trees to learn from, it was right about fifty-four percent of the time. That is a long way past the seventeen percent my first pixel-based attempt managed, and far above the roughly five percent of pure guessing. The distinctive species come back well, blue spruce, American elm, honeylocust and ash among them, while the look-alikes blur, the lindens with each other and the oaks and ornamental maples into the broadleaf crowd.

So species turned out not to be hopeless from free imagery once I worked crown by crown rather than pixel by pixel. It is simply not reliable enough to stand in for a field survey, and a city does not need it to be. I kept the five management classes as the headline because they are what a crew acts on, but the species result shows how much signal the crown approach pulls out of the same free data.

To me the accuracy figure matters less than the work the map saves. The Emerald Ash Borer kills untreated ash and is moving through the Front Range, and a crew can only inspect so many trees in a season. The most useful thing my model produces is a short, ranked list of where the ash most likely are, so the first trucks go to the right blocks. About sixteen hundred crowns are flagged as likely ash for a crew to confirm and treat. It points the field work rather than replacing it. In a later test I added an October image, when ash turns color on its own schedule, and it improved how reliably the model finds ash by about seven points, the single most useful addition for this borer-triage job.

The crown map is a single snapshot. To watch the canopy change I needed a second date and a sensor that revisits often, so I turned to the Harmonized Landsat-Sentinel product, or HLS, a free NASA dataset that puts the Landsat and Sentinel-2 satellites on one 30 meter grid and delivers calibrated surface reflectance every few days. Surface reflectance is the analysis-ready standard the training I took emphasized, and it is the one thing my crown model above does not use, since that model still learns from raw NAIP brightness values while HLS measures light corrected to the ground.

I built a cloud-free summer composite of Hale for 2016 and for 2025, measured greenness in each, and compared them. The neighborhood canopy turned out to be broadly stable. Average greenness barely moved, and the patches that grew greener and browner roughly cancel, with about two percent of the area losing vegetation and two and a half percent gaining it. I am careful not to over-read the rest, because two summer snapshots a few years apart differ a little just from sun angle and the exact day they were captured, so the stable picture is the trustworthy part rather than the sign of every small change.

The trade is resolution. At 30 meters a single HLS pixel covers several trees plus the street and yards around them, so HLS cannot pick out one crown the way my 1 meter model does. I use the two together. The crown model says what each tree is, and HLS says how the whole neighborhood's greenness is holding up over time.

The newest approach in this field does not hand-build features at all. Geospatial foundation models are large neural networks pretrained on enormous amounts of satellite imagery, in the way a language model is pretrained on text, though these read imagery and return numbers rather than words. The same pretrained model can then be reused for many tasks. To see where my project sits next to that frontier, I ran Prithvi, a foundation model from NASA and IBM trained on Harmonized Landsat-Sentinel imagery, over central Denver. Its weights are published openly, so it downloads once and runs on the same desktop as everything else here. I handed it a six-band HLS image and, with no labels, let it turn the scene into its own learned features, then grouped those into land-cover types.

What comes back is honest about the promise and the limit at once. The model organizes the scene on its own, but Prithvi reads the image in roughly 480 meter patches, so it sees central Denver as a largely uniform residential canopy with only coarse distinctions. That is the same lesson from a new direction. The variation that matters for urban forestry lives at the scale of a single tree, which is exactly where my one meter crown model works, and where a foundation model would need fine-tuning on a graphics card to follow. I include it as a marker of the road ahead, and because the project now spans the full range the field uses, from a random forest to a foundation model.

The newest approach lets a model learn its own features instead of using hand-built ones, so I tried that at the scale this project works at, one tree at a time, on a graphics card here at my desk. I started with the simplest version, a general vision foundation model called DINOv2, and gave it a small aerial chip of each tree. With only the red, green and blue of a single photo to work from it reached about a third of the crowns, far behind the engineered model.

Then I built a model that sees everything the engineered features see: a small network that takes the near-infrared and the laser-measured height alongside the color image and reads the six-date Sentinel-2 series for how each tree changes through the year. That closed almost all of the gap. It matched the engineered model on balanced accuracy, the figure that weights every class equally, and it was the better of the two on the hardest disease classes, lifting elm recall from 0.57 to 0.82 and maple from 0.39 to 0.55. It came in lower on overall accuracy, because it gives up some of the large catch-all class to do better on the rare ones.

I take the pattern across the three models as the honest lesson of the whole project. A model is only as good as the data it is shown. The general model with color alone trails badly, the same idea fed the full multispectral, seasonal and height data pulls level with years of hand-built features, and the gradient boosting still edges ahead on the headline number because, with only a few thousand labeled trees, hand-built features are hard to beat outright. That last point is why I see pretraining as the real next step.

My project now runs the full span the field uses, on a single neighborhood and on free data: an object-based crown model that maps management classes and reaches into species, an honest spatial check on what that accuracy really means, a surface-reflectance look at how the canopy changes over time, and a stack of learned models, from two foundation models to a purpose-built deep net, run over the same ground. The real frontier from here, and the step my deep net points straight at, is a foundation model pretrained on exactly the kind of multispectral and seasonal data this problem depends on, at fine resolution. I see that as a research undertaking rather than a missing afternoon's work.

I chose only free, openly licensed inputs, and the whole pipeline runs from a single script, so the result is reproducible and the dataset is easy to share. I adapted the methodology from Cross (2019), who classified tree species in Costa Rican forest from WorldView-3 satellite imagery. My contribution is carrying that object-based, reflectance-driven approach to free imagery and to the practical problem of managing an urban canopy.