Early detection of bark beetle infestations in Central Europe using deep learning–based reconstructions of irregular Sentinel-2 time series

Abstract Norway spruce (Picea abies) is among the most abundant tree species in Central Europe. Due to climate change-induced extreme weather events, spruce trees are increasingly stressed and therefore threatened by European spruce bark beetle (Ips typographus) infestations. Recent mass outbreaks led to severe ecological and economic damage in Central European forests. After an infestation, the filial generation of the beetles swarms out within 6 to 10 weeks to infest new trees. Consequently, an efficient bark beetle management needs to remove infested trees within 10 weeks to prevent further dispersal. While remote sensing allows for large-scale monitoring of forests, the detection of bark beetle infestations remains challenging, as many trees show no visible signs of the infestation within the 10-week detection period. Here, we try to achieve early detections by adjusting a state-of-the-art Deep Learning model to be able to cope with irregular Sentinel-2 satellite time series for reconstruction-based anomaly detection. The model is trained on >300 000 time series of undisturbed coniferous forest and the threshold denoting an anomaly is derived independently, i.e. not from the test dataset. We test the model on a geographically independent dataset with known infestation dates. It achieves moderate performance for detections within 10 weeks after the infestation with a producer’s accuracy (PA) of 11.8% ± 8.4% and user’s accuracy (UA) of 43.5% ± 24.5% across three model runs, but yields very good results when extending the detection period to 13 weeks (UA = 84.5% ± 7.6%, PA = 81.5% ± 1%). Since the model responds immediately to an anomaly, we conclude that area-wide bark beetle detections within 10 weeks after infestation are likely impossible using Sentinel-2 alone. Still, our approach can readily be used as a near real-time monitoring system for coniferous forest, be applied on any forest disturbance detection task, and may complement terrestrial surveys in the future.