Analyzing spatio-temporal patterns of crown dieback and dead forest trees detected by AI and aerial RGB imagery

Monitoring the quantity and spatial distribution of dead and dieback trees is essential for tracking forest health and understanding drivers of decline. However, individual-tree level assessments remain temporally and spatially discontinuous due to financial constraints. Regularly collected aerial and drone imagery, combined with deep learning algorithms, now offer new opportunities for spatially continuous monitoring of tree health, enabling the identification of spatiotemporal decline patterns valuable for forest management. Here, we used UltraCam-Xp aerial images (2014, 2016) and UAV images (2025) to detect dieback and dead trees in the Alang Darreh Forest, Iran, using a hybrid YOLO11s-SAM model across three time points. Spatial patterns of detected trees were analyzed using Moran's I and Getis-Ord Gi statistics to identify hot and cold spots of tree decline. The identified health decline patterns were then related to topographic variables and the distance from roads using Gi Z -score values, Spearman coefficients and a relative effect analysis to understand the reasons for the observed tree decline. To further understand the drivers of these patterns, we employed multiple regression with two-way interaction terms to also assess the combined effects of slope, elevation, aspect, and road distance on forest decline. Accuracy assessment results showed that the hybrid deep learning model detected dieback and dead trees with overall accuracies of 95.55–98.51% (Kappa coefficients of 0.93–0.97) and 84.79–88.76% (Kappa of 0.74–0.82), at point and polygon levels, respectively. Spatial pattern analyses revealed clustered patterns for dead trees in all years, while dieback trees showed a random distribution in 2014 and a clustered distribution in 2016 and 2025. Spearman correlations revealed significant negative relationships of tree health decline with elevation and slope, and varying positive correlations with aspect across years. Distance from park roads showed mostly negative correlations with tree decline, except for dead trees in 2016. Multiple regression further revealed that road distance was the strongest predictor of forest decline, while significant two-way interactions among road distance, elevation, and aspect confirmed that environmental factors act in combination rather than independently. Relative effect analyses highlighted higher densities of dieback and dead trees at elevations of 400–450 m, slopes below 50%, northeastern to southeastern aspects, and within 150 m of forest roads. In summary, this research demonstrates how high-resolution RGB imagery in combination with the YOLO11s-SAM hybrid deep learning approach provides reliable information for spatio-temporal monitoring of forest decline at individual tree level that among others can serve as a valuable step to guide the sampling procedures of subsequent studies seeking to identify mechanistic reasons for forest health decline. Our approach can be easily adapted for other forest types and regions where high resolution RGB imagery is available.