Integrated deep learning framework for automating tree health assessment using ground-based images

Urban trees are increasingly vulnerable to pests, pathogens, and climate induced stressors, threatening their health and resilience. Effective monitoring is essential for proactive management, yet traditional inspection and remote sensing methods are limited in scalability and timeliness. Ground-based RGB imagery from DSLR cameras offers a practical, affordable alternative with a horizontal perspective suitable for assessing structural health indicators. However, extracting reliable information from such imagery requires advanced image analysis. This study presents an integrated automated deep learning framework for assessing urban tree health using convolutional neural network (CNN)-based multiclass semantic segmentation. As a proof-of-concept case, the framework is developed and validated using ash trees images. A UNet model was trained both standalone and with pre-trained backbones (ResNet34, ResNet50) to classify images into foliage, wood, ivy, and background. The UNet–ResNet50 model outperformed others, achieving an overall accuracy of 84.3% and an F1 score of 0.81 for foliage. Segmented outputs were used to derive tree health indicators (THIs): defoliation, tree height, crown length-to-height ratio (ClThR), ivy index, tree tilt, and crown symmetry. A weighted sum of THIs was used to develop a Tree Risk Index (TRI). Defoliation predictions showed strong agreement with Observers’ scores ( R 2 =0 . 83), while other indicators reflected varying degrees of accuracy, influenced by image quality and visibility constraints. The framework demonstrates a scalable, non-invasive method for systematic tree health assessment. By combining deep learning with ground-based imagery, it enables consistent extraction of structural indicators and provides a robust foundation for urban tree risk monitoring and management.