European canker, caused by Neonectria ditissima, is a major fungal disease that affects apple orchards and leads to significant losses in yields. Traditional detection methods rely on manual visual inspection, which is labor intensive and error prone. This thesis explores an automated, deep learning-based approach to detect canker using RGB and near-infrared (NIR) images across different growth stages of apple trees.

A custom image dataset was collected, annotated, and used to train YOLOv8 models. Experiments were conducted using low-resolution RGB images of young trees, high-resolution RGB images of mature trees, and NIR images, with the latter converted from grayscale to 3-channel format to ensure compatibility with the model. The best performing model achieved an F1 score of 0.95 and mAP using the NIR dataset, demonstrating that NIR imaging significantly improves lesion detection accuracy. Real-time inference was also validated on unseen test images.

This work shows that multimodal imaging combined with lightweight deep learning models like YOLOv8 can enable reliable, real-time detection of canker in field conditions, reducing manual workload and supporting precision agriculture practices. The study laid the groundwork for the deployment of vision-based disease detection systems on edge devices in commercial orchards