Forests

Bark beetle prediction

We ranked forest in three areas of southern Sweden by how likely spruce bark beetles were to attack in the 2024 season, using data up to 2023. In forward tests, plots in the highest risk class were attacked about 3.8 times as often as the average plot.

The goal: where to look first

Nobody can inspect every stand of spruce. The map ranks forest so that inspection can start where an attack is most likely. It says where to look first, not how much forest is threatened.

How we built it

1. Field truth

   SLU's inventory recorded freshly attacked spruce on field plots of 20–25 m radius from 2020 to 2024. Plot positions are confidential and do not appear on the map.

2. What the forest looks like from above

   Sentinel-2 imagery, land cover, tree height, soil wetness, terrain, climate, nearby clear-cuts, beetle traps and 9.2 million AI-detected dead spruces from Skogsstyrelsen.

3. More examples to learn from

   We added 4,200 dead-spruce events and 3,216 spruce controls at low weight. This raised the score on unseen Östergötland forest from 0.17 to 0.23. We call this model B5+.

4. Tests on data the model had not seen

   Train outside Östergötland and test there; train on earlier years and test the next; compare with Skogsstyrelsen's published risk index and with later dead-spruce detections.

5. The map

   Forest is ranked within each area and split into five classes, at 100 m resolution, using the October forecast for the 2024 season.

The classes separate risk

Attack rates fall steadily from the very high class to the low class. The percentages shift between years, so the multiples are the steadier guide.

How it compares

B5+ clearly beat Skogsstyrelsen's risk index in the forward 2023 test. In Östergötland the two cannot be told apart.

AUPRC measures how well attacked plots rank at the top; higher is better. Östergötland has 418 plot-years with 38 attacks, and forward 2023 has 994 with 63. With no skill, the riskiest 10% would hold about 10% of attacks.

Simple field records and attack history did about as well as anything on Östergötland, and with only 38 attacks the ranges overlap widely. In forward tests, B5+ pulled ahead of the field model in 2022 and 2023 (2.5× and 3.9× better than chance, against 1.5× and 1.8×) but not in 2024, with its 12 attacks (2.9× against 3.6×).

How far to trust it

• It ranks; it does not count. In backtests, risk levels fitted 2022 and 2023 but came out 3–4 times too high in 2024 (10.6–13.9% predicted against 3.0% observed). Hectares at risk are not yet reliable.
• Few attacks mean wide ranges. 2024 has only 12 attacks in the forward test, and field plots are not a random sample of the forest.
• The map is a look back. B5+ learned from some 2024 aerial photos, which did not exist in spring 2024, so this map could not have been issued then.
• Dead-spruce agreement is flattering. Agreement with later detections is high, but part of it is the model learning what the detector sees, and trees in a 2024 photo mostly died from attacks up to 2023.
• Colours do not compare between areas. Classes are ranks within each area.

What comes next

Later forecast dates help only slightly: the riskiest 10% held 29% of attacks for an October forecast, 30% for April and 34% for June. That supports the main goal, a forecast map for 2027 issued before the season and scored when the 2027 field data arrive.

Data

Boundaries from SCB. Satellite background: contains modified Copernicus Sentinel data 2023, via Microsoft Planetary Computer.