A Dual-Layer UAV Workflow for Tree-Level Monitoring of Tree Decline Using RGB Imagery and a Lightweight Deep Learning

Objective: Tree decline is a serious multifaceted problem in Zagros forests of Iran. The prevalence of this complex phenomenon in the recent decades highlights the need for high-resolution geospatial monitoring approaches. While unmanned aerial vehicle (UAV)-based photogrammetry offers a flexible and cost-effective means of capturing forest health, conventional top-of-canopy imaging fails to sufficiently represent critical under canopy features, including stem morphology and lower crown structure that is commonly prone to early symptoms of tree decline.
Method: We presented a dual-layer UAV photogrammetric framework that combines above- and below-canopy imagery to detect phenotypic decline in Quercus brantii using a novel 3d tree reconstruction method followed by a MobileNetV2 deep learning model to detect decline symptoms on stems. Using this detection, we computed the Phenotypic Decline Index (PDI) and Decline Acuteness Index (DAI) to describe decline severity and trends in continuous form.
Results: The MobileNetV2 achieved overall classification accuracy of 96.3% (F1-score = 0.94) in distinguishing healthy and declined stems (n = 299). This performance, derived from a confusion matrix with 166 true positives, 9 false negatives, 11 false positives, and 133 true negatives, highlights the model's high reliability. Furthermore, the UAV-derived DAI correlated strongly with multi-year field-based decline trajectories (Pearson r = 0.718, Spearman ρ = 0.928) , confirming the method’s reliability.
Conclusions: Decline severities during three years of field data collection followed by suggested consistent shift amongst levels of tree decline, while UAV-based phenotypic analysis was shown to enable capturing nuanced changes in tree vitality. By yielding high correlations, we provide a cost-effective and high-resolution workflow for phenotyping oak decline, which enables multi-scale analysis of structural and symptomatic indicators using RGB-only data.