Artificial intelligence identifies trees from the air as accurately as the human eye: potential applications in the near future - Zeme un valsts

Artificial intelligence identifies trees from the air as accurately as the human eye: potential applications in the near future

A new remote-sensing method can identify individual tree species with accuracy of up to 90%. Detailed species data make forestry more efficient and improve biodiversity monitoring.

Thanks to recent research in tree species classification, tree species can now be identified from the air far more precisely, extensively and quickly than before. The new method combines multispectral laser scanning with machine learning and deep learning methods, in practice, with artificial intelligence (AI).

The research was carried out by scientists at the Finnish Geospatial Research Institute (FGI), part of the National Land Survey of Finland, together with researchers from several universities. Over the past 15 years, forest resource data describing forest structure and tree species have mainly been gathered using a method based on laser scanning, aerial photography and field measurements.

“Aerial photographs, which combine visible-light and near-infrared wavelengths, provided species data that ordinary laser scanning could not, since it relies on only a single wavelength. Aerial photography has had an important role until now,” explains Markus Holopainen, professor of geoinformatics at the Department of Forest Sciences, University of Helsinki.

The more advanced multispectral laser scanning method, like aerial photography, measures several wavelengths, meaning more detailed data can be obtained from laser scanning alone. “With multispectral laser scanning, the accuracy of tree species classification comes very close to that of aerial photography. In the future, forest inventories may no longer need both laser scanning and aerial photography. Instead, a single measurement could provide both (sufficient) datasets,” Holopainen stresses.

A large-scale study in tree species classification

“In remote sensing, AI methods can generate data comparable to what the human eye perceives. Like a person, deep learning can identify a tree species even when colour tones vary between individual trees in an image or dataset.”

The new method developed by FGI and the universities combines multispectral laser scanning with machine learning and deep learning, improving accuracy still further. Deep learning can quickly and efficiently combine complex data, such as tree shape and geometry, along with the reflectance of different wavelengths from trees.

Markus Holopainen notes that the study is among the most ambitious internationally in tree species classification, with a particular focus on multispectral laser scanning and the development of laser scanning methods.

Improved recognition of rarer species

The new remote-sensing method achieved its highest species-detection accuracy with particularly dense 3D laser scanning data, that is, data containing a large number of measurement points. With such data, deep learning helped identify nine species: pine, birch, spruce, aspen, rowan, alder, oak, lime and maple. Trees were identified with an average accuracy of 92%. Detection accuracy also improved even when the data was “less dense”.

The most common species in Finland’s forests, such as Scots pine, Norway spruce, and silver and downy birch, were identified with very high accuracy, while detecting less common broadleaved species proved more difficult. Even so, compared with traditional methods, deep learning models significantly improved the recognition of less common species as well.

Professor Markus Holopainen notes that the challenge lies in the fact that accuracy declines as the number of species increases: “Species diversity in the forests of Finland and other Nordic countries is relatively low; moving towards Central Europe or other regions with a wide range of species, identification quickly becomes more difficult.”

Remote sensing becomes harder when trees overlap or a forest has many different storeys (layers). “The denser the stand, the harder it is to classify the forest. The method works best in mature commercial forests that have undergone thinning or selective felling,” the professor notes.

Benefits for forestry and biodiversity protection

Identifying tree species matters for both forestry and biodiversity. Accurate species data improve, for example, timber harvesting planning, logging operations and timber trade, which overall raises the productivity of forestry.

The new AI-based method also improves the recognition of species, such as aspen, that are important for biodiversity, and helps monitor changes in species composition.

“To work successfully, the machine learning methods used in our study require a large amount of training material, whereas deep learning methods can work with smaller volumes. Both methods perform better as the area of the territories they are applied to increases,” Holopainen stresses. Dead trees or areas of forest damage, for example, can already be identified fairly accurately using aerial and satellite imagery, and deep learning can improve accuracy in this area too. “More effective monitoring of damaged areas allows more detailed measurements to be carried out sooner by drone or on the ground, making it possible, for example, to spot damage caused by the eight-toothed spruce bark beetle in good time.”

Fieldwork will still be needed

If species identification and remote-sensing technologies are becoming this accurate and automated, will it still be necessary to go out into the forest?

“Deep learning methods do genuinely reduce the need for forest visits, but they do not replace them entirely. In the forest resource inventory carried out by the Finnish Forest Centre, species information is just one of around a hundred indicators. Determining most of the others still requires measurements taken on the ground,” explains Markus Holopainen.

On the other hand, if the goal is simply to identify rare species important for biodiversity or to gather information on dead wood, the new method is entirely sufficient, the professor believes: “Deep learning models need material to learn from, and preparing that material is fairly slow work, since visual classification of trees still relies on human effort.”

Development of precision forestry

Traditionally, forest scanning used ordinary LiDAR or laser scanning technology, which measures only tree height and density (at a single wavelength) and cannot reliably distinguish tree species. To compensate, the data was combined with ordinary aerial photographs.

The achievement of the Finnish researchers lies in the use of multispectral lasers. These emit several light wavelengths simultaneously, capturing both a tree’s 3D structure and its “spectral signature”, meaning how the tree reflects light. A human cannot process such a vast volume of data manually, which is where deep learning comes in: neural networks that recognise tree species from their shape and the nuances of light reflected by the foliage.

Already in the near future...

In the coming years (up to 2030), this technology will have huge potential in both the commercial and nature conservation sectors, including in Latvia, where forest structure is fairly similar to Finland’s.

Very fast detection of forest damage and pests will become possible. The technology can be integrated with drones to automatically scan stands and detect, for example, an eight-toothed spruce bark beetle infestation at an early stage, when trees start to change colour in a way the human eye still cannot see from below. Automated, cheaper forest inventory will also become possible. Forest owners will no longer have to pay for expensive, lengthy field surveys to find out the exact composition of a stand. A single drone flight will “deliver” a ready-made map showing each tree’s species and volume.

Even more efficient planning of logging and logistics will become possible. Harvesters will head into the forest with a precise digital map, knowing exactly where the most valuable trees stand, optimising machine movement and reducing CO₂ emissions.

It should be stressed that compliance with biodiversity and EU environmental requirements will also improve. It will become easy to survey and protect ecologically valuable habitats nationwide, and to locate old oaks, aspens or dead wood, which many protected species depend on.

In the near future, this technology will deliver maximum benefit both in commercial forestry and in nature conservation, as well as in cooperation between the two fields.

Benefits for environmental protection and nature surveys

Biodiversity monitoring will become cheaper, broader and more objective. In Latvia and Europe, nature surveys require huge human resources and years of work, with experts walking through forests and so on. With AI and aerial scanning, EU-significant habitats, such as old boreal forests or swamp forests, could be identified nationwide within a few weeks, based on tree age, species and structural diversity.

This method can also recognise and map the rarer broadleaved trees that are home to hundreds of species of moss, lichen and insects. Surveying aspen, which makes up 1–1.5% of Nordic forests, is particularly important, since it is a key biodiversity “anchor” in these forests. In Finland and Sweden, aspen is described as a “keystone species”. In Latvia, aspen covers roughly 8% of forest area. Dead wood is another important factor, critical to forest ecosystems yet difficult to survey. The new system will accurately detect and classify both standing dead trees and fallen deadwood, making it possible to assess the natural health of a forest. States and large companies will be able to demonstrate precisely how much carbon their forests are sequestering. An accurate calculation of tree species and volume will provide reliable data for the carbon credit market.

A sustainable future

This technology’s greatest achievement will be reducing conflict between loggers and conservation activists. Because the data will be fully accurate and available digitally, when planning fellings the system will be able to automatically “flag” and direct harvesters to bypass the most ecologically valuable trees, such as aspens to be retained, old oaks, or areas with a high concentration of dead wood. Commercial forestry will run efficiently while guaranteeing that natural values are not accidentally destroyed.

Add a comment