A faster, cheaper way to map the world’s forests with AI
New research from the University of Cambridge offers an easy and readily available solution using AI and satellite data. In a paper published in the journal Science of Remote Sensing, researchers used ...
New research from the University of Cambridge offers an easy and readily available solution using AI and satellite data. In a paper published in the journal Science of Remote Sensing, researchers used embeddings by Tessera, a geospatial foundation model, to successfully map tree species across the Trentino region of the Italian Alps.
The results outperformed traditional satellite methods and were achieved using standard computing. Lead author James Ball, a postdoctoral researcher at Cambridge's Conservation Research Institute, said the research opened the door to accurate, low-cost, continent-scale forest mapping.
"Tessera is a democratizing force that brings satellite data to the masses," Ball said. "A team in a region without access to huge computing resources can apply this approach on forest inventories and embeddings that have already been produced by the Tessera team and roll it out for their region without too much expense or effort."
The study is among the first to use geospatial foundation models to create detailed maps of tree species distribution from space. The ability to cost-effectively generate continent-scale forest maps in this way will be of particular interest to government bodies such as the European Union and conservation organizations such as the UN Environment Program's World Conservation Monitoring Center and IUCN. It is also relevant to companies and organizations dealing with carbon credits.
The trouble with tree mapping
Mapping tree species from space is notoriously difficult. That's especially true in mountainous regions such as Trentino, where steep slopes and cloud cover make it virtually impossible to separate one species from another in individual images. The challenge has long been a bottleneck for forest mappers.
Previously, researchers had hoped that satellites carrying hyperspectral sensors, which capture light across hundreds of wavelengths, could disentangle the subtle, species-level signal. That has proved difficult, even with supercomputer-powered models to process the highly complex data produced by these sensors.
Foundation models like Tessera analyze a year's worth of satellite images and distill them into a clean, ready-to-use embedding. The embedding captures seasonal changes in the canopy as trees flower, seed, drop and regrow leaves, information that's unique to each species. A simple model analyzes the embedding to reveal individual species.
In the current study, Ball trained a model to map 18 tree species and species groups in a Tessera embedding of Trentino. The results were tested against traditional methods of mapping forests as well as another foundation model, Google's AlphaEarth.
Both Tessera and AlphaEarth performed well, correctly classifying around 82%. Traditional methods ranged from roughly 73% to 79% accuracy. Tessera was better at picking out rarer species and more accurately predicted the correct species proportions within each map unit.
According to co-author Michele Dalponte, a forest ecologist at Italy's Fondazione Edmund Mach who supplied the Trentino dataset, roughly three-quarters of Trentino's forest is publicly owned and actively managed, while the remaining quarter is private and largely unmanaged. The new mapping is important for forest administration because of its coverage and high level of detail.
"It gives us the species composition of private land that's typically data-scarce, and it adds spatial detail within public forests," Dalponte said. "It serves as a powerful tool for the forest service—not only for locating economically valuable species but also for identifying localized biodiversity hotspots."
A soft approach to AI training
The study is also notable for its use of a machine-learning technique called soft labels, in which the model takes a probabilistic approach to the labels it uses for training, something that most models don't normally do. Instead of discarding areas that contain a mix of species, it uses the proportion of each species within a map unit as the training target.
"We don't have to throw away all the mixed-up data," Ball said. "We can use soft labeling to extract more information from the messy parcels."
Ball, whose usual research is on tropical forests, is now applying the same methods to a forest-plot network spanning more than 11,000 sites across Latin America.
"This sets up the opportunity to integrate more data and see how far we can push this for automated, continental-scale forest mapping that can represent more of its interesting traits," he said.
More information
James G.C. Ball et al, Geospatial foundation models enable data-efficient tree species mapping in temperate mountain forests, Science of Remote Sensing (2026). DOI: 10.1016/j.srs.2026.100466
Who's behind this story?
BA art history, MA material culture. Former museum editor, paramedic, and transplant coordinator. Editing for Science X since 2021. Full profile →
Master's in physics with research experience. Long-time science news enthusiast. Plays key role in Science X's editorial success. Full profile →
Citation: A faster, cheaper way to map the world's forests with AI (2026, July 29) retrieved 29 July 2026 from https://phys.org/news/2026-07-faster-cheaper-world-forests-ai.html
This document is subject to copyright. Apart from any fair dealing for the purpose of private study or research, no part may be reproduced without the written permission. The content is provided for information purposes only.