Detection of Bark Beetle Attacks Using Time-Aggregated Satellite Data with Machine Learning
Shokoufa Zeinali, Per-Ola Olsson, Ted Kronvall, Magnus Wiktorsson, Johan Lindström
In this study, we explored how early bark beetle attacks can be detected using weekly aggregated Sentinel-2 data in combination with static data, such as geo- and forestry data. We used an XGBoost classifier, known for its strength and reliability in classification, and compared three sets of data: static data only, satellite data only, and using both together. Having trained the models on cumulative weekly data, we were able to track changes in model performance and feature importance over time, identifying key weeks for the detection of bark beetle attacks. A systematic overview of feature importance identified the Red-edge 3 and blue Sentinel-2 bands as the most important when combined with static data; it also showed changes in feature importance compared to using satellite-only data, e.g., adding static features reduced the importance of red and red-edge 2 bands. Among the static features, land cover and landforms were the most important. Evaluating the temporal features for the combined model highlighted certain weeks as containing key information for detection: week 19, which was the main swarming week; week 25, which is 6 weeks after swarming and just before the second generation is completed; and weeks 31 and 33, more than 3 months after the tree was attacked and well after the new generation has swarmed. The study shows that combining static features with cumulative Sentinel-2, accumulated across weeks, are all important for improving the detection of bark beetle attacks, and that such ideas form an important part of early warning systems.