Comparative study of near real-time monitoring algorithms for early detection of bark beetle infestations in Germany with Sentinel-2

Recent mass outbreaks of bark beetle infestations have severely affected Germany's spruce forests. Forest management requires detection within 10 weeks after infestation to prevent further dispersal. Several near real-time (NRT) satellite time-series change-detection methods have been developed. Existing methods, however, have been evaluated in diverse contexts and data types, leaving their comparative performance unclear. Here, we compare four NRT change detection methods including FORDEAD, S-CCD, BFAST Monitor (harmonic model-based) and TimesNet RSAD (deep learning-based), with respect to their bark beetle detection capability on a dataset of spruce monocultures in Germany with precisely dated infestations using Sentinel-2 time series. While no method achieves satisfactory Producer's Accuracy (PA) in the critical detection window of 10 weeks, BFAST Monitor remains best for early detection (PA ~50%). When given a longer detection window (14 weeks), TimesNet RSAD achieved the highest PA (>80%), followed by FORDEAD (~77%), while BFAST Monitor showed the lowest User's Accuracy (UA) (~11%) compared to TimesNet RSAD (>98%). We conclude that TimesNet RSAD is likely best suited for early bark beetle detection using Sentinel-2, mainly by avoiding false positives. Nevertheless, reliable early detection in Germany based on currently freely available multispectral satellite imagery in large-scale NRT monitoring still seems unachievable.