New data will make AI models smarter

A research group at Umeå University is using AI to improve forest management in a variety of ways. They have now collected detailed data for training AI models to estimate accessibility in the forest, amongst other things. The data will be used in several Mistra Digital Forest projects and will be made available to anyone who wants to work with it more extensively.

The AI models being developed within Mistra Digital Forest can facilitate many aspects of forestry, and in order for them to become really good at their task, they need to train with good data. Using airborne laser scanning and a Komatsu forwarder, researchers in the Digital Physics Group at Umeå University have collected new, high-quality data from two of SCA's forest areas. It will now be used to train AI models in simulations.

 – The idea is that the data will be used as training data in several Mistra Digital Forest projects. When we train AI models in simulations, we save time and we avoid damage to machinery and terrain. The data will also be made available to researchers, machine manufacturers and others who want to work with it in the future, says Erik Wallin, researcher in digital physics at Umeå University.

Training AI to estimate accessibility in the forest 

In the near future, the data will be used to develop an AI model that can estimate trafficability in the forest. By integrating this function into decision support, it is possible to estimate the energy consumption during felling in advance, as well as which forestry machine is best suited to a task, based on the specific conditions of the terrain, for example.

To develop the AI model, the researchers made a virtual copy of the Komatsu forwarder used for data collection. This virtual copy then runs in a simulation that mimics the forest, with great variation in terrain, and is used to train an AI model that can assess energy consumption and trafficability. By comparing the AI model's predictions with the detailed data collected by the physical forwarder, the researchers can validate how well the AI model performs.

 – This lays the foundation for optimisation models that help us become more resource efficient, as well as enabling us to make decisions that are beneficial for both finances and climate. For example, when we can estimate the energy and time required for a particular harvest, we can devise more accurate plans and our compensation to contractors will be more precise. This project is in line with SCA's efforts to make greater use of high-resolution data in order to enable precision forestry, says André Wästlund, Business Developer at SCA.

Aiming for more high performing AI models 

A common challenge for almost all projects where AI models are trained in a simulated environment, is the transition of the model from simulation to reality. It is essential that the virtual copy has been adapted to mimic reality. This currently requires researchers carrying out the majority of the adaptation - the calibration - themselves. However, with the help of the detailed data, the implementation can be streamlined. For example, the data makes it possible to systematically vary the virtual forwarder's settings and to use an algorithm to help to find the optimal choices. In this way, calibration can most probably be simplified and partly automated. The aerial laser data can also be translated into a terrain model that is even closer to reality. This is where the AI model can be trained to recognise real obstacles, for example rocks and low vegetation such as bushes, something that would also facilitate calibration in a physical forestry machine.

 – In order to achieve a well-functioning AI model, we need exactly this kind of data, collected using a disciplined process, and describing the variety of scenarios that forest machines encounter in the field, says Erik Wallin.

The data has been collected on two occasions (2023 and 2024) on SCA's land in Ånge and Björsjö. On each occasion, a Komatsu forwarder, model 895, collected detailed data on position, energy consumption, crane movements, etc. At Björsjö, image data was also collected using a 360 camera on the forwarder. The areas were also scanned with Lidar to produce detailed 3D models of the terrain.