<p>Early detection of tree decline is critical for effective forest health management, and the integration of remote sensing with artificial intelligence provides an efficient framework for timely monitoring of forest ecosystems. This study was conducted in two oak forest sites in the Zagros Mountains of Iran, and analyses were conducted separately for each site and jointly using combined data. Model performance was evaluated for both individual tree crown detection and crown health classification into three categories, i.e., healthy, declining, and dead, allowing assessment of accuracy, robustness, and generalizability across heterogeneous forest conditions. Mask R-CNN achieved higher segmentation performance than YOLO in both output types, with F1-scores of 0.85 for bounding boxes and 0.83 for masks, indicating more precise crown delineation. Both models could distinguish between healthy, declining, and dead tree crowns in the canopy health classification stage. YOLO outperformed Mask R-CNN in this task, especially when trained on combined data from both forest types. The highest F1-score, 0.72, was achieved for the healthy class using YOLO with bounding box output, while the lowest was recorded for the dead class using Mask R-CNN, with an F1-score of 0.36. YOLO also showed stronger performance for the declining class with an F1-score of 0.61 compared to 0.52 for Mask R-CNN. Overall, the results demonstrated that Mask R-CNN was more effective in tree crown detection, while YOLO provided better results for canopy health classification. Moreover, aggregating data from both forest sites improved the generalizability of both models, enhancing their robustness and applicability to larger areas.</p>