Beech leaf disease symptom detection using deep learning and computer vision tools.

Area of Science:

• Forest pathology
• Plant disease diagnostics
• Artificial intelligence in ecology

Background:

• Beech leaf disease (BLD) poses a significant threat to forests in eastern North America.
• Current visual surveillance methods for BLD can miss early-stage infections, hindering timely intervention.

Purpose of the Study:

• To develop and validate deep learning models for accurate and early detection of beech leaf disease.
• To compare the performance of different deep learning architectures in identifying BLD-affected leaves.

Main Methods:

• Collected datasets of symptomatic and asymptomatic beech leaves from multiple regions.
• Trained and tested various deep learning models, including EfficientNetV2-Small, ResNet50, MobileNetV3-Large, and InceptionV3.
• Utilized Grad-CAM visualizations to interpret model focus and disease characteristics.

Main Results:

• EfficientNetV2-Small achieved 100% accuracy on the primary dataset and 96.55% on an independent validation dataset.
• All tested deep learning models demonstrated high accuracy in distinguishing between healthy and BLD-affected leaves.
• Grad-CAM analysis confirmed model attention on characteristic BLD leaf banding patterns.

Conclusions:

• Deep learning and computer vision offer a promising, efficient approach for monitoring beech leaf disease.
• These AI-driven tools can enhance early detection and support forest management efforts against BLD.
• The study highlights the potential of AI for scalable and accurate forest health surveillance.