Beech leaf disease (BLD) has rapidly emerged as a significant threat to forests across the eastern United States and Canada, and early detection is a major challenge. Current surveillance relies on visual identification of characteristic leaf banding, a method that can miss early infections. To address this limitation, we developed deep learning models capable of distinguishing between BLD symptomatic leaves and asymptomatic leaves. A primary dataset of symptomatic and asymptomatic leaves collected in Maryland (Dataset I) was used for model development, and an independent set of images collected in North Carolina, Ohio, and New England (Dataset II) provided real-world validation. In Dataset I model testing, EfficientNetV2-Small was the most accurate model (100%), followed by ResNet50 (99.32%), MobileNetV3-Large (97.95%), and InceptionV3 (94.88%). Independent testing on Dataset II also identified EfficientNetV2-Small as the most accurate model (96.55%). Grad-CAM visualizations confirmed that EfficientNetV2-Small focused on banded regions of BLD leaves that are a main characteristic of the disease. These findings demonstrate the potential of deep learning and computer vision approaches to support more efficient monitoring of BLD in forested regions.
Waldo et al. (Thu,) studied this question.