How Aerial Images and Satellite Data Sorted Hidden Patterns of America's City Trees

Researchers at Oregon State University explored how much of that information could be recovered from observations collected above the ground.

A tree outside your window can cool the pavement, catch rainfall and provide a welcome patch of green on a crowded street. Its species influences how much shade it offers, which pests threaten it and whether its pollen makes spring uncomfortable for nearby residents. Keeping track of those differences across millions of city trees takes time, money and records that many municipalities struggle to maintain.

Researchers at Oregon State University explored how much of that information could be recovered from observations collected above the ground. Their study, "Multimodal Classification of Urban Trees into Remote-Sensing-Derived Taxonomic Groups Across the United States," appeared in the journal Remote Sensing. Andrew Larkin, Matthew Bozigar and Perry Hystad combined several kinds of imagery and environmental data to investigate patterns across urban forests.

The team classified trees into six broad groups with 79.3% accuracy on a separate test dataset. These groups reflect similarities in canopy appearance, height, seasonal changes and environmental surroundings, rather than exact species identities. The research offers a way to describe urban forests across large areas, with clear limits on what those descriptions can tell us.

Building a clearer picture from millions of records

The researchers began with approximately 5.43 million tree records from 60 American cities. Checks for unreliable coordinates, missing observations, unsuitable labels and other problems reduced the usable dataset to about 1.85 million records across 36 cities. Those records represented 579 taxa, meaning species or broader taxonomic labels where precise identification was unavailable.

Each tree was linked to detailed National Agriculture Imagery Program aerial photographs and airborne LiDAR measurements, which use laser pulses to measure height. Sentinel-2 satellite observations captured seasonal vegetation changes, Google Satellite Embeddings supplied numerical summaries of Earth observations, and PRISM weather data described temperature, rainfall and atmospheric dryness. The satellite and weather records covered 2021–2023, giving the model information about changes through time as well as physical appearance.

A tree-detection model helped correct location errors by matching inventory entries to visible crown centres within five metres. Researchers built the initial grouping system around 280 better-represented taxa, then assigned another 299 less-common taxa to the closest established groups. Taxa with fewer than 50 eligible records were excluded, leaving the rarest trees outside the study's coverage.

Six groups shaped by what sensors can differentiate

Trying to recognise hundreds of species from publicly available imagery runs into a practical problem: many trees look too similar from above. The researchers developed 11 intermediate groups and combined them into six broader categories after comparing separation and prediction performance. The resulting categories describe patterns the observations can distinguish, rather than a new botanical classification.

The first group contained warm-climate evergreen and coastal trees, including palms, eucalypts and evergreen oaks. The second brought together cool-temperate ornamental broadleaf trees such as cherries, crabapples and many maples, and the third covered temperate deciduous canopy trees including elms, ashes and several oaks. The fourth included warm- and dry-climate ornamentals such as crape myrtle and Chinese pistache, the fifth centred on plane trees and some riverside broadleaf trees, and the sixth was dominated by conifers.

Trees from the same genus did not always land in the same group, with silver maple separated from several other familiar maples. The groupings captured differences in visible structure, seasonal behaviour and surroundings that botanical relationships alone could not explain. Sharing a group does not establish that trees provide the same cooling, pollution removal or pollen exposure, so those uses would need separate validation.

What the accuracy figures reveal

The final classifier was evaluated on 161,508 records held apart from model development and group selection. Its 79.3% accuracy was accompanied by a macro-F1 score of 0.772, a measure that gives each group equal importance and balances correct identification against mistaken assignments. Retaining all 11 intermediate groups produced 68.8% accuracy, showing how much confusion involved categories that were eventually combined.

For records meeting the two-metre vegetation-height threshold, the warm-climate evergreen and coastal group earned the highest group F1 score, at 0.867. Conifers scored 0.705, and the largest two-way confusion occurred between cool-temperate ornamentals and temperate deciduous canopy trees, with roughly 15% in each category assigned to the other. Shorter vegetation and less-common taxa were harder overall, though rare taxa with clear, stable group assignments performed almost as well as better-represented taxa.

Detailed aerial imagery and its combination with canopy-height information received most of the model's internal weighting. Removing the combined Sentinel-2 and weather sequence caused the largest performance drop, reducing accuracy by about 2.6 percentage points and macro-F1 by 2.8 points. That sequence was also the strongest input when used alone, showing that seasonal and weather information added value beyond a detailed aerial view.

Why a new city remains a harder test

The main evaluation included different trees from cities already represented during training, where nearby records could share satellite pixels and environmental conditions. When researchers excluded a target city from training, average accuracy fell to 68.5% and macro-F1 to 0.478. Models trained on just one city performed much worse across target cities, averaging 34.3% accuracy and demonstrating the value of geographically varied training data.

A Sentinel-2 observation covers a ten-metre grid cell, which can contain several crowns, lawns, shrubs and pavement. Performance weakened where vegetation occupied little of that cell, where more satellite observations needed gap-filling, and where inventory locations were farther from detected crowns. These patterns expose uncertainty in linking a tree's label to its surroundings, especially for small trees or imperfect records.

The filtering process favoured cities with stronger inventories and more complete imagery, restricting how widely the reported accuracy can be applied. Differences in acquisition dates, LiDAR coverage and inventory practices complicated comparisons, and the final model's five-checkpoint ensemble came from one training run rather than independent repeated runs. More consistent field records, broader geographic coverage, finer seasonal imagery and tests of regional models could help address those weaknesses.

The work provides a foundation for mapping broad urban tree groupings using publicly available observations. Its outputs could help describe city forests, but decisions requiring exact species or ecological traits still need more specific evidence. The researchers have made their code, training datasets and models available for others to examine and build on.