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

A pixel classifier across ten common species reached about 17% accuracy. NAIP provides four bands, and many species look alike at that scale. Whole crowns recover more signal, but species labels remain too uncertain to replace field identification.

I regrouped the labels into ash, elm, maple, conifer, and other deciduous: categories tied to disease risk, water demand, and field treatment.

Five Sentinel-2 dates add seasonal behavior. Conifers stay green into November while deciduous crowns leaf out and fade, giving the classifier information a single summer image cannot provide.

A vegetation-masked watershed splits the DRCOG LiDAR canopy-height raster into 18,272 crowns. Each crown stays one object through feature extraction and classification.

Each crown has 109 features from NAIP color and near-infrared, Sentinel-2 seasonality and vegetation indices, plus LiDAR height and shape. Matching crowns to Denver's arborist inventory produced 3,764 labels. Gradient boosting won a four-model comparison.

Unsupervised clustering shows structure in greenness and height, along with the main ambiguity: maple overlaps heavily with other deciduous crowns. The supervised model makes most of its errors there.

Overall accuracy is 69.0%, up from 38.3% for the pixel baseline. Balanced accuracy is 62.8%. Conifer and ash perform best; maple remains weakest because it overlaps other deciduous crowns.

NAIP color and Sentinel-2 seasonality account for nearly nine tenths of feature importance; LiDAR height contributes about a tenth. Most separation comes from how crowns reflect light across bands and seasons.

Random splits can flatter spatial models because nearby crowns share lighting and planting patterns. I therefore withheld complete 200 m and 400 m blocks during testing.

Accuracy holds at 66.4% for both block sizes, about two points below the random split. That supports use within Hale; it does not establish transfer to another neighborhood.

Crown-level classification across 21 sufficiently labeled species reached about 54% accuracy, compared with 17% for the pixel baseline. Blue spruce, American elm, honeylocust, and ash separate best; lindens, oaks, and ornamental maples often merge with similar broadleaf species.

That is useful for exploration, but not reliable enough for field identification. The five management classes remain the operational output.

The model flags about 1,600 crowns as likely ash for field confirmation and treatment planning. Adding an October image improved ash recall by about seven points, making seasonal timing the most useful tested change for Emerald Ash Borer screening.

For change analysis I used NASA's Harmonized Landsat-Sentinel surface reflectance, which places both sensors on a common 30 m grid.

Cloud-free summer composites show a broadly stable neighborhood: about 2% vegetation loss and 2.5% gain. Small differences can also come from acquisition date and sun angle, so I do not interpret each changed pixel as a tree-level event.

A 30 m HLS pixel covers several trees and surrounding ground. It measures neighborhood change, while the one-meter workflow maps individual crowns.

I ran NASA and IBM's Prithvi-EO model over six-band HLS imagery for central Denver, then clustered its learned features without labels. The published weights ran locally.

At roughly 480 m per patch, Prithvi separates broad land-cover patterns but misses tree-level variation. It is a useful coarse baseline, not a substitute for the one-meter crown workflow.

DINOv2 embeddings from one RGB aerial chip per crown reached 32.9% accuracy, well below the engineered feature model.

A purpose-built network combines RGB, near-infrared, LiDAR height, and the Sentinel-2 time series. It matches the engineered model on balanced accuracy and raises elm recall from 0.57 to 0.82 and maple from 0.39 to 0.55, while lowering overall accuracy.

With 3,764 labels, gradient boosting still leads overall. The multimodal network performs better on rare classes; the RGB-only foundation model lacks the spectral, seasonal, and height inputs the task needs.

Every input is free and openly licensed. One script runs the pipeline. The crown-based method adapts Cross (2019) from WorldView-3 tropical forest imagery to public Denver data and urban-forestry classes.