“Plenty of opportunities for AI in forestry”

In a few short years, AI has transformed the way we work and make decisions. For the forestry sector this development brings both significant opportunities, and new risks. Kristina Knaving, senior AI expert at RISE, shares her views on the matter.

What opportunities do you see for AI in forestry?

– The opportunities are numerous. Right now, image recognition is a particularly interesting area where AI automatically interprets images from satellites, drones and cameras in the field. This opens the door to large-scale forest inventory, where damage and diseases can be identified at an early stage. It then provides a solid basis for decision-making, in order to balance production targets with environmental and nature conservation goals. Other exciting areas include analysing camera and sensor data to detect machine wear and tear, and the fact that generative AI can program prototypes incredibly quickly. It becomes easier to test ideas, compare scenarios and make informed decisions at an early stage.

What challenges do you see in relation to data accessibility?

– Generally speaking, the more training data we have, the better the models become. This requires various forms of collaboration; companies need to become better at sharing data, but in practice this is difficult, and is often sensitive. GDPR, trade secrets and uncertainties surrounding the division of responsibility all mean that a great deal of information is kept under lock and key. It can also be difficult to coordinate data that has been collected in different ways, and in different formats.

– Information security, in particular, is a major issue going forward. For example, how should forest owners respond to the fact that open datasets, such as satellite images, can provide other industry players with valuable information about their land? And how do we handle privacy issues, such as the fact that harvester data contains information both about the forest, and also about the machine operator?

You often say that the best data is the right data, followed by no data, and last of all, bad data. Please elaborate!

– That the data is relevant to the task and that we understand its limitations is even more important than the sheer volume of data. For example, data on trees selected for felling should not be used to answer questions about the entire stand. Given that many organisations purchase off-the-shelf AI solutions, more people need to be trained in AI and to gain the expertise needed to ask critical questions about the data that has been used. Knowledge of norms and bias is also required here. A planning tool trained on an excessive amount of data on spruce risks suggesting spruce over pine more frequently, and risks being less effective at assessing pine. These norms are built into the goals that the systems are optimised for. If AI is trained to maximise efficiency, values such as biodiversity and social considerations risk being sidelined.

Large language models have become a valued tool. How should organisations approach this?

– Every organisation needs to establish clear guidelines on how AI should be used, and on what company data may be fed into language model services. It’s worth remembering that AI doesn’t reason like we do, even though it may seem that way when reading text generated by AI. Think about what you yourself would have answered, before the language model responds to a question – and always review the result. Verification remains absolutely crucial. AI can save time, but as the user, you are the one making the judgement. Question the answers, refine the questions, and remember that you are the expert on your forestry operations – not AI.