Inside SmartForest: Norway’s push for AI-driven forestry

This article is based on our podcast episode released on 5 February 2025. [Listen to the full episode here]

Forestry is evolving fast, with AI, LiDAR, and remote sensing changing how forests are monitored and managed. SmartForest, a Norwegian research initiative, is taking digital forestry even further by using machine learning and advanced modelling to enhance forest operations.

One of its key projects is FOR-species20K, a dataset of 20,000 trees from 33 species, designed to train AI models that can identify species using either 3D LiDAR scans or 2D images. But how accurate is AI compared to traditional forestry expertise? And does this level of species identification add real value?

Images vs LiDAR: Which works best?

One of the biggest surprises in SmartForest’s research is that image-based AI models outperform LiDAR-based AI models in species identification.

"The image-based models actually perform better," says Rasmus. "It’s not by much, but they generally have a higher accuracy."

According to their findings, image-based AI models achieved 79.5% accuracy, while LiDAR-based models reached 75.6%.

While LiDAR provides detailed 3D structural data, it struggles with distinguishing between certain species. Image-based models, on the other hand, capture colour and texture, making it easier for AI to tell species apart.

"I would assume that point clouds are better for other use cases than detecting species," says Rasmus. "It’s great for measuring structure, but maybe not for identification."

AI performs better on large trees than small ones

A major limitation in AI-based species identification is tree size. SmartForest’s findings reveal that AI models perform significantly better on large trees than small ones.

"Once trees get over 8 metres, the accuracy is much higher," says Rasmus. "But as soon as you drop below 5 metres, it becomes a real challenge."

Younger trees have simpler crown structures, making them harder for the models to differentiate, and their bark and leaf characteristics aren’t as well developed. This suggests that AI-powered species recognition might be more effective in mature forests rather than early-stage forest management or regeneration sites.

Is AI-powered species identification necessary?

While AI can successfully classify tree species, there is still the question of whether it is solving a real-world problem.

"Most foresters can identify species easily," says Jens. "So is this really solving a real problem?"

In practical forestry, stand-level data is more important than identifying individual trees. Foresters rely on species composition, growth trends, and overall forest health rather than cataloguing every single tree.

There are, however, cases where AI-powered species recognition could prove useful. It could improve the precision of carbon storage calculations, assist in biodiversity monitoring by detecting rare or invasive species, and reduce the need for manual species surveys in large-scale forest inventories.

Still, Jens remains sceptical. "If you’re running a forestry operation, do you really need to know every single tree? I’m not so sure."

Geographic limitations: AI is trained on European species

One challenge with AI-based species recognition is the data used to train it. The FOR-species20K dataset consists mostly of European tree species, meaning AI models might struggle when applied to forests in North America, Asia, or tropical regions.

"It’s a classic AI problem," says Jens. "If your dataset is biased towards certain conditions, the model won’t generalise well outside those areas."

To truly make AI species recognition scalable, future projects will need larger, more diverse training datasets that incorporate tree species from different climates and ecosystems.

LiDAR struggles in dense forest stands

Another key limitation is how AI handles forests with dense canopies. In unmanaged forests or multi-layered stands, LiDAR struggles to differentiate individual trees, leading to misclassifications or missing species entirely.

"It works well when trees are standing apart," says Rasmus, "but if you’re scanning a really dense forest, LiDAR has a hard time separating individual trees."

This suggests that LiDAR-based species identification might be best suited for plantation forestry, selective logging, or open forests rather than dense, old-growth stands.

Beyond AI: The rise of digital twins

While AI focuses on species recognition, another emerging concept in forestry is digital twins – virtual models that simulate growth, predict climate impact, and assess risks like drought and fire.

"If we had full digital models of forests, we could test things like drought scenarios or fire risks," suggests Rasmus.

By continuously scanning forests and feeding AI models with real-time data, foresters could predict changes and optimise forest management in ways that were never possible before.

But this technology is still in development. Rasmus warns against full automation.

"You don’t want to go full automation. There’s a balance. Machines can do a lot, but we still need foresters to make the big decisions."

How accurate does AI need to be?

A key question in forestry AI development is how much accuracy actually matters. Even though image-based AI models outperform LiDAR-based AI models, the difference is only around 4%.

"If we go from 75% to 79%, is that actually making a difference?" asks Jens. "At the end of the day, foresters need reliable data, but perfect data isn’t always necessary."

Instead of focusing on perfect identification, AI's biggest value might be in automating large-scale surveys, making forest monitoring faster and more efficient without requiring perfect precision.

Modern forestry is constantly evolving, and it can be hard to keep up with all the new technologies – even for us who work with exactly that every day.

For more details, visit SmartForest’s official website or read the full report on FOR-species20K here.