ArbNation Join free

Computer vision and tree species recognition: what studies report on accuracy

Three open-access papers report how well computer models name tree species from satellite images and laser point clouds. Accuracy runs from 85.65% to 96.94%, and look-alike conifers are the usual mistake.

Checked against the sources at the bottom of this page on October 16, 2026. Rules, fees and pay change: the source has the last word.

Tree species recognition by computer vision means a model that labels each tree, or each patch of forest, with a species from remote sensing data. Three peer-reviewed papers in open-access journals give the figures below. They test satellite imagery and drone laser scanning on forest stands in South Korea and North China. None of them tests street trees, and none is a field guide.

Satellite imagery and a convolutional network

Cha and colleagues at the Korea National Institute of Forest Science published their study in Forests in 2023. They fed a U-net convolutional neural network with RapidEye and Sentinel-2 images taken on several dates in 2019, plus texture data from the near-infrared band and vegetation indices. Reference data came from 20 m by 20 m field plots, over 800 of them.

The Gwangneung study area has nine species. They cover 28,200 hectares, led by Quercus aliena, dentata and serrata together at 29.46% and Pinus koraiensis at 23.70%. The Chuncheon area has eight species on 590 hectares. The authors report 90.5% accuracy in Gwangneung against 84.5% for a random forest model, a 6 point gain. The same model reached 92.1% in Chuncheon, and the authors say further validation is needed.

Patch size changed the result a great deal. In Gwangneung, a patch size of 304 gave recall of 14.0% to 42.3% in one scenario. A patch size of 784 gave 59.0% to 90.5%. The authors saw the most misclassification among the needle-leaved trees, Pinus rigida and Larix kaempferi, and tie it to similar spectral signatures among conifers. They add that a small study area limits how well a trained model carries over to new places.

Drone laser scanning and point-cloud models

Zhang and colleagues at Shanxi Agriculture University published in Frontiers in Forests and Global Change on June 4, 2025. They scanned 12 plots in North China with a drone-mounted laser. The data covered 2,622 trees of four species: Populus alba, Populus simonii, Pinus sylvestris and Pinus tabuliformis. They segmented individual trees, then compared two machine learning methods with two deep learning methods.

MethodOverall accuracyKappa
PointMLP (deep learning)96.94%0.9589
Random forest95.62%0.9329
Support vector machine94.89%0.9221
PointNet++ (deep learning)85.65%0.8083

The authors found that the feature for the 99th percentile of return height had an important influence on the result. PointNet++ confused species with similar crown structures: it labelled 33 Pinus sylvestris as Pinus tabuliformis and misread eight Populus alba as Pinus sylvestris. For the deep learning models they recommend 1,024 to 2,048 points per tree. PointMLP stayed between 95% and 97% across the sampling counts they tried.

A review of 300 publications

Zhong and colleagues at Southwest Forestry University reviewed 300 publications on mapping tree species with remote sensing. Their review appeared in Forests on May 13, 2024. It reports that the single data types used most are hyperspectral images, laser scanning, ordinary colour images and very high resolution imagery. The most used pairing is hyperspectral with laser data. The classifiers used most are convolutional neural networks, random forests and support vector machines.

The review also reports a split by scale. Traditional machine learning serves large study areas, while deep learning has mostly served small ones, and is starting to move to larger ones. In one survey it cites, 347 forestry professionals ranked tree species first among 63 parameters they wanted from remote sensing.

The review lists limits on accuracy: spatial, temporal and spectral resolution, shadows, cloud cover, small canopies, overlapping crowns and similar canopy spectra. It names labelled field data as a burden, since the field surveys behind labels are slow and laborious. Patch size, it adds, has not been studied enough.

Study areas behind the figures

  • Cha and colleagues: 90.5% and 92.1% apply to forest stands in two Korean regions, not to individual street trees.
  • Zhang and colleagues: 96.94% applies to four species in North China, scanned from a drone at a flying height of 50 m.
  • Zhong and colleagues: the review reports trends and limits, with no single accuracy figure.

Each paper measures accuracy on its own species list, sensors and plots, so the percentages are not a ranking of methods. Other pages cover LiDAR and remote sensing for tree work and tree inventories.

Sources