# Denver Urban Tree Classification

Mapped 18,272 tree crowns from public LiDAR, aerial imagery, and five Sentinel-2 dates. The five-class model reaches 69.0% accuracy and holds at 66.4% when whole spatial blocks are withheld.

- URL: https://tomshanks.dev/projects/denver-urban-tree-classification
- Period: 2026
- Capabilities: Remote sensing and Earth observation; Applied machine learning on sensor data; Geospatial data engineering

## System

- USGS 3DEP LiDAR canopy-height model
- watershed crown segmentation
- NAIP + Sentinel-2 phenology (5 dates)
- per-crown features
- classifier
- spatially held-out evaluation

## Claims and evidence

- **Segmented 18,272 tree crowns in the Hale neighborhood of Denver from public LiDAR and classified them into forestry classes.** _(implemented)_
  - measurement: crowns — 18,272 · classes: ash (EAB risk), elm (DED risk), conifer, maple, other deciduous
  - page: methods and results — https://tomshanks.dev/projects/denver-urban-tree-classification
  - source: strata component, packer and crown field — https://github.com/Tom-Shanks/spatial-fields/tree/master/canopy-strata
- **Accuracy reported with whole city blocks held out, not just a random split, because neighbouring crowns are not independent.** _(implemented)_
  - measurement: accuracy — 69.0% random split · 66.4% spatial block hold-out