AI image analysis of forest: how the technology counts every tree
AI image analysis of forest requires millions of parameters and synthetic trees built branch by branch. This is how the technology behind Arboair's tree counting works.
A forest owner with five hundred hectares in northern Sweden can today ask a question that had no answer five years ago: exactly which trees should I thin, and where on the property are they? The answer is generated by an AI system that analyses aerial imagery pixel by pixel and returns individual data for every tree in the stand. Building such a system — and building it correctly — has turned out to be one of the harder technical problems in modern forestry technology. It is not a problem solved with more data and more compute. It requires a different approach.
Two models read the forest in different ways
Behind Arboair's AI image analysis of forest there is not a single model but several parallel systems, each with its own task. The first system reads colour. It analyses differences in pixel values and tries to determine which pixels belong together — whether a given pixel belongs to a spruce, a pine or a deciduous tree, and whether the pixel next to it belongs to the same tree or a different one. The model builds a probability matrix and clusters the pixels into objects based on what it estimates they represent.
The second system reads shape and height. It works with three-dimensional data and analyses differences in height profiles, textures and geometric forms to determine what constitutes a delimited object and which object it is. Both systems contain millions of parameters and nodes, trained over years of work with real forest data from Scandinavia and beyond.
It is the combination of the two ways of reading — colour and three-dimensional form — that allows the system to recognise and classify trees in forests with widely differing characteristics and geographic conditions. Neither the colour model alone nor the height model alone is enough. It is the interplay between them that carries the result.
Why training data is the forest's hardest problem
A natural impulse is to think of AI training as a mathematical problem with a clean solution: gather enough data, run it long enough, and the model finds the patterns. That holds in theory. The reality in a forest is more complicated.
One of the fundamental challenges concerns class balance. A well-trained model needs comparable quantities of every tree type it is meant to recognise. But the forest rarely provides that balance. Many stands are spruce-dominated, pine-dominated or dominated by a single deciduous species. The training data reflects that reality — and a model trained on imbalanced data tends to become better at the common and worse at the uncommon, precisely the opposite of what you want in a forest where the uncommon is often the economically decisive part.
On top of that, trees do not look the same in different places. A spruce from Norrbotten and a spruce from Skåne are genetically the same species but grow under such different conditions that they can look fundamentally different in an aerial image. Crown density varies. The height stratification of a stand varies. Light falls differently depending on latitude and season. The model has to learn to recognise spruce as a species, not the specific appearance of spruce in a given location — and that distinction is hard to teach.
The errors, when they occur, are often understandable. They usually happen when the colours in the image are very similar, making it hard to determine where one tree ends and another begins. Noble broadleaves are among the hardest categories: the crown of an old oak can be so wide and irregular that the system hesitates over whether it is one tree or four.
Trees built branch by branch inside a computer
To handle the complexity of training data, Arboair has taken a step that sets the company apart from most other players in the market: its own system for synthetic training data. It starts with constructing trees inside a computer — not as simplified 3D models but at a level of detail that takes time to fully appreciate.
Branch by branch. Needle by needle. Until every synthetic tree is photorealistic in detail. An algorithm then generates thousands of variants of each tree type — different ages, different crown shapes, different branch structures and densities. The synthetic trees are placed into digital landscapes and the system takes drone-perspective images of them, exactly as a real drone would have done during an assignment.
The decisive difference is what the computer knows when it takes the images. It knows exactly which tree it placed where. It knows the exact species, height, crown size and composition. No human needs to sit and interpret and annotate. That means the training data lacks the bias and misinterpretations built into all manual labelling — the kind of systematic error that is otherwise impossible to eliminate entirely, however careful the annotator. The ground truth of the reference data is one hundred per cent.
It has taken years to build. The result is a knowledge lead that Arboair estimates corresponds to a decade of conventional training-data development — a lead that shows most clearly when the models encounter forests they have never trained on and still perform well.
Millions of trees as a reference base
Arboair's database today contains millions of trees — real trees from real forests, measured and analysed over years. It is that volume that makes the difference. A model that has seen enough spruce in enough contexts begins to distinguish spruce as a species rather than as a specific pattern in a specific place — a generalisation that is fundamental to the system working outside the places it was originally trained on.
That scale would not have existed five years ago. The parallel development of more powerful AI architectures and falling compute costs has made it possible to use more advanced models at a cost that is genuinely commercially defensible. It is the combination of data at the right volume, models with sufficient capacity and compute infrastructure at a reasonable cost that explains why the technology has matured now and not ten years earlier.
From pixels to decisions — what AI image analysis of forest actually enables
What a forest owner sees is not pixels and probability matrices. It is a map of the property with individual data for every tree: species, height, crown size, stem volume, health status. From that data you can ask questions that have never had precise answers before.
Which trees should be thinned, and where on the property are they? Which trees have the greatest potential to become high-quality sawlogs? What is the property's exact timber volume ahead of a generational transfer or a property transaction? Which parts of the stand show signs of poor health that require action?
This is documentation that fundamentally changes the decision process. Arboair's assessment is that forest owners working with this level of precision can extract around five per cent more timber value from their property over a five-year period, compared with those using low-resolution data or manually surveyed management plans. That is an estimate grounded in a known difference in data quality and a known correlation between decision precision and financial outcome — not a measured average from a controlled trial.
The technology has also opened up possibilities for continuous-cover forestry. Selective harvesting and conservation-oriented management require exact knowledge of every individual in the stand — it is hard to plan which trees should remain without knowing exactly where they are, what they are and what they are worth. Individual tree-level data gives that planning a basis it previously lacked.
Five hundred hectares in northern Sweden. With AI image analysis of forest, it is no longer an area to manage in broad strokes — it is five hundred hectares of trees you can actually know.