Tree crown anomaly detection in UAV RGB-NIR imagery using an unsupervised cosine-distance-based approach in hemiboreal forests

Automated UAV-based data processing can substantially enhance the temporal resolution and spatial accuracy of forest health monitoring. To facilitate practical applicability in operational forestry, there is a critical need for unsupervised workflows that minimise manual intervention. This study explored the efficacy of distance-based anomaly detection (AD) for identifying tree crown damage in RGB-NIR UAV orthophotos. We developed an unsupervised cosine-distance-based AD workflow and evaluated it against a standard NDVI thresholding baseline using both time-series analysis and a large spatial dataset comprising 31 hemiboreal forest stands. The results demonstrated that while both methods achieved comparable accuracy for mapping established deadwood, the AD workflow exhibited heightened sensitivity to subtle, early-stage canopy damage. However, AD requires uniform summer illumination, whereas NDVI—though simpler to execute—demands constant, scene-specific threshold recalibration. Despite the requirement for stable illumination, the AD workflow offers substantial operational value as an automated forest damage detection system that operates without continuous manual threshold recalibration or prior field data.