Sentinel-2 and Machine Learning Approaches for Early Detection of Forest Pest Outbreaks in Croatia

This study presents an ESA-funded project developing machine learning methods for early detection of forest pest infestations in Croatia using Sentinel-2 satellite images. Targeting Ips typographus (European spruce bark beetle, ESBB) and Corythucha arcuata (oak lace bug, OLB), a multi-temporal TempCNN achieved 96.3% green phase and 94.4% red phase detection accuracy for ESBB, outperforming Random Forest and XGBoost. For OLB, a single-date XGBoost exceeded 99% accuracy. Classification maps at 20 m resolution reached 92.2% (ESBB) and 76.1% (OLB) overall accuracy, validated through area-adjusted error matrices. Key features included red-edge, SWIR, and NIR bands with stress-sensitive vegetation indices.