Machine learning forest map suggests North America has fewer large trees than previous estimates

New research published in How many trees are there in North American forests? forest ecosystem Researchers are getting closer to answering that question by mapping tree density in Canada, the United States, and Mexico using a combination of forest survey data, satellite observations, and machine learning.

Tree density is an important indicator of forest structure and plays an important role in studies of carbon storage, biodiversity, ecosystem function, and forest management. Although national forest inventories provide valuable information, differences in data availability and sampling methods can make it difficult to consistently assess tree density over large geographic areas.

To address this issue, the research team combined forest inventory data from more than 600,000 forest inventory plots with environmental information such as climate, soil, and topography from satellites and other sources. They then compared several machine learning approaches to determine which best predicted tree density across North America at approximately 3 km resolution.

Among the models tested, feedforward neural network (a type of machine learning model) outperformed other approaches in prediction accuracy and was chosen to generate the continental map.

The map reveals clear regional patterns. The highest tree densities are found in boreal and temperate coniferous forests across Canada, Alaska, and the Pacific Northwest. Many eastern forests have moderate densities, but deserts and other arid regions have far fewer trees.

Using this approach, researchers estimate that North America includes: 339 billion and 514 billion trees with diameter >10 cm at breast height. This range is less than the widely cited figure of 603 billion trees across the continent.

This study investigated where predictions are more reliable. Areas with rich inventory data, especially the United States and southern Canada, have lower uncertainty. In contrast, regions with more complex environments or less comprehensive study areas, such as northern boreal forests and mountainous regions, exhibit higher uncertainties.

One of the key findings of this study is that Tree abundance estimates are sensitive to forest definitions and tree size thresholds.. Results varied depending on the forest map used and the size of the smallest tree included in the analysis. These technical choices are often overlooked but have major implications for policy and carbon accounting.

This framework supports forest monitoring, biodiversity assessment, and carbon accounting applications. Its modular design allows for updates as new inventory data and satellite observations become available, providing a consistent approach to tracking forest conditions over time.

This study establishes a reproducible foundation for large-scale forest assessments across North America by integrating harmonized data, remote sensing observations, and explicit uncertainty analysis.

DOI link:

https://doi.org/10.1016/j.fecs.2026.100466