Street-level imagery and AI for tree inventories: what studies report
Four studies in US Forest Service Treesearch test imagery and deep learning for street tree inventories. Volunteers counted trees reliably from street-level photos, but diameter and species agreement fell, and one method placed 38 percent of trees.
Checked against the sources at the bottom of this page on October 21, 2026. Rules, fees and pay change: the source has the last word.
Four papers in the US Forest Service Treesearch library test whether images can replace some field work in a street tree inventory. They differ in method: volunteers reading street-level panoramas, a detector trained on panoramas to locate trees, deep learning on aerial orthophotos, and measurements from aerial imagery and LiDAR. Each one reports its own error rates.
Volunteers reading street-level panoramas
Berland, Roman and Vogt published their test in Forests in 2019. They asked 16 volunteers to inventory street trees in suburban Chicago from street-level panoramas, then compared the results with field data from the same places. Agreement among the volunteers was very good for the number of trees on each street segment. It was markedly lower for diameter class and for identification to genus and species.
Only tree count passed 0.8 on Krippendorff's alpha, the agreement score the authors treat as good. Experts agreed with one another more often than novices on all four variables, and novices scored 0.29 for genus and 0.23 for species. Some intermediate analysts matched the experts. The paper also timed the work.
| Group | Minutes per tree |
|---|---|
| All virtual survey analysts | 3.01 (range 1.07 to 8.75) |
| Experts | 1.45 |
| Intermediate analysts | 3.41 |
| Novices | 4.23 |
| Field crew of two, travel time excluded | 3.14 |
The authors conclude that virtual inventories suit basic information such as tree locations, or updating an existing inventory to see where trees were planted or removed.
Locating trees from addresses
Laumer and co-authors, from ETH Zurich, the Forest Service and other institutions, published in the ISPRS Journal of Photogrammetry and Remote Sensing in 2020. Older inventories recorded trees by street address and newer ones by GPS. The method detects trees in street-view panoramas with deep learning, merges the detections from several views into one tree, and matches them to the trees listed at each address by global optimization. The detector was trained on 6,783 labelled trees in 718 panoramas.
In tests on more than 50,000 trees in five California cities, the method gave correct coordinates to 38 percent of street trees (38.6 percent in the paper's closing section). The authors name its main shortcoming: 26 percent of trees were assigned incorrectly. Run on all 48 California municipalities with inventory data, it assigned locations to 619,229 of 1,100,952 trees, or 56 percent.
Deep learning on aerial photos
Velasquez-Camacho and co-authors published in PLOS One in June 2025. They applied deep learning to National Agriculture Imagery Program orthophotos, at 1 m and 0.6 m resolution, to find large street trees, with canopy over seven metres across, in six San Francisco Bay Area counties. The imagery ran from 2005 to 2022. The model was retrained on 100 images.
The paper prints two accuracy figures. The validation section reports a detection accuracy of 75 percent, with precision of 90 percent and recall of 81 percent, tested against San Francisco's official street tree inventory. The discussion reports an accuracy of 81 percent. It compares an earlier study of large trees on private land that had 84 percent precision and 37 percent recall.
Census tracts with higher household incomes, a larger share of residents who self-identify as white, and more families were linked with growth in large tree canopy over the period.
Crown size and height from imagery
Lee, Ko and McPherson tested how far aerial imagery and LiDAR can stand in for field measurement in Sacramento, California, in Urban Forestry & Urban Greening in 2016. They measured 802 street trees of 20 species in the field and compared crown diameter from the aerial photo with the field value.
- Trunk diameter at breast height, derived from crown diameter by curve fitting, came with r2 = 0.883 and RMSE = 10.32 cm.
- Tree height from the LiDAR surface model had an RMSE of 1.64 m.
- Diameter and height from the remote data ran lower than field values without adjustment, and the gap was larger for smaller-stature trees.
- Aboveground biomass from the remote measurements matched field-based results with r2 = 0.881 and RMSE = 799.2 kg.
Sources
- Berland, Roman and Vogt, Can Field Crews Telecommute? Varied Data Quality from Citizen Science Tree Inventories Conducted Using Street-Level Imagery, Forests 10(4): 349 (2019), USDA Forest Service Treesearch record 58791, read October 21, 2026
- Laumer, Lang, van Doorn, Mac Aodha, Perona and Wegner, Geocoding of trees from street addresses and street-level images, ISPRS Journal of Photogrammetry and Remote Sensing 162: 125-136 (2020), USDA Forest Service Treesearch record 60426, read October 21, 2026
- Velasquez-Camacho et al., Monitoring temporal changes in large urban street trees using remote sensing and deep learning, PLOS One 20(6): e0326562, June 26, 2025, USDA Forest Service Treesearch record 69423, read October 21, 2026
- Lee, Ko and McPherson, The feasibility of remotely sensed data to estimate urban tree dimensions and biomass, Urban Forestry & Urban Greening 16: 208-220 (2016), USDA Forest Service Treesearch record 50950, read October 21, 2026