Green Attack or Overfitting? Comparing Machine-Learning- and Vegetation-Index-Based Methods to Early Detect European Spruce Bark Beetle Attacks Using Multispectral Drone Images

Detecting forest insect damage before the visible discoloration (green attacks) using remote sensing data is challenging, but important for damage control. In recent years, the European spruce bark beetle (Ips typographus, L.) has damaged large amounts of forest in Europe, and some studies have been conducted on the early detection of infestations and forest vulnerabilities before attacks. This study assessed the detectability of the green attacks using multispectral drone images and examined the possibility of detecting vulnerable trees before attacks. The study used multispectral drone images covering 24 plots from 6 forest stands in southern Sweden, acquired in May (before attacks), June (green attack), August (green and yellow attack), and October 2021 (red attack). Drone images of individual-tree crowns were segmented and vegetation indices (VIs) were calculated for every single tree. Trees with the same duration of infestation were grouped for the analysis. Random Forest Classification (RF) and linear discriminant analysis (LDA) were used to build and compare models using all bands, sensitive bands, all VIs, and single VIs, respectively. Results were also compared between different ways of dividing training and testing data. When randomly dividing 90% and 10% trees for training and testing, the models could classify vulnerable trees before attacks with low accuracy. However, when training on trees in five stands, no model could predict infestations in the remaining test stand. Similarly, the models could not identify trees infested for fewer than five weeks. We conclude that the detectability of vulnerable trees before attacks and attacked trees with fewer than five weeks of infestation is very low. We noticed a considerable overfitting when using RF with more variables compared to using LDA with single VIs.