Synthetic data and AI are taking forestry into the future

How can we make better decisions about our forests—faster, more sustainably, and at a tree-specific level? One way is to combine advanced AI with synthetic 3D models of forest environments. This is precisely what has been done in an innovative project.

"The project shows how AI and synthetic data can be used to create sustainable value in an otherwise quite traditional industry, while strengthening our technical ability to use these types of solutions," says Erik Roos, 3D Artist at Arboair.

The project, which received funding from Vinnova, started in autumn 2022 and aimed to develop and test AI models for forest analysis using synthetic data. The work builds on Arboair's previous experience. The goal was to take the technology to a level where it could be tested in practice and used by players in the forestry industry.

A new AI engine for the forest

The project resulted in an AI engine that combines image analysis models with numerical models and is based on an automated pipeline for synthetic data generation. By creating computer-generated images of trees, damage, and various environments in 3D, Arboair has been able to train its models in a way that was previously impossible. The synthetic data has been used for both validation and as a complement to real data.

To further strengthen the development, the project utilized AI Sweden's technical infrastructure, to get access to powerful computing capacity and expertise.

Breakthroughs and insights

A concrete result of the project is the successful implementation of top-breakage analyses in winter environments using synthetically trained models. Top breakage occurs when the top of a tree breaks off, often due to snow pressure or storms. This negatively affects the tree's growth and value, making it important for the forestry industry to detect such damage early. Since the top breakages occurred suddenly during a snowy winter, they were not represented in the training data.

To quickly train a model that could still recognize them, synthetic data was required to replace the missing examples and provide the model with sufficient variation. In this case, top breakage became an example of a use case where the need arose quickly, and synthetic data proved crucial for saving time.

Not all original project goals were met—for example, a finished simulator for forest owners was not created—but the project represents a major step forward in synthetic data generation. It has also led to improvements in Arboair's existing products.

"Currently, the challenge is to achieve sufficient variation in the synthetic images to match the chaotic reality," says Erik Roos. "However, synthetic data can already be used in some cases for testing and R&D to increase understanding of how the AI platform performs under different conditions."

Collaboration and technology boost

The project was carried out in collaboration with Holmen, which provided test areas and requirements specifications, and with the Spillkråkan network, whose members were part of the reference group and contributed important perspectives from smaller landowners. This broad collaboration gave the project both practical relevance and innovative scope.

Technically, development has taken great strides forward. Advances in graphics cards and software have made it possible to render images with thousands of trees and millions of leaves—something that was unthinkable just a few years ago.

The next step for Arboair involves follow-up projects aimed at including more environments and tree types to build an even more robust model. In parallel, a tree generation engine is being developed—a technical platform that could eventually become central for Arboair as well as other players in forestry.

"We have taken important steps forward, but we see this as a beginning rather than an end. Our ambition is to make the technology available to more people and contribute to more sustainable and data-driven forestry—both in Sweden and internationally," says Erik Roos.