Evaluation and Optimization of an AI Model for European Canker Detection in Apple Trees

Abstract

This paper presents a study on the use of Convolutional Neural Networks (CNNs) for the detection of European Canker (Neonectria ditissima) in apple tree leaves. Several CNN architectures were experimentally evaluated using Data Augmentation, ensemble strategies, and threshold-based decision methods, each tested over ten independent replications to ensure robustness and reproducibility. The experiments demonstrated the effectiveness of these approaches for reliable and accurate disease detection, highlighting their potential to support early diagnosis and management decisions in apple orchards. The best-performing models achieved test-set precision values ranging from 0.801138 to 0.871632 and accuracy values from 0.76542 to 0.842679, which are comparable to those obtained by two agronomists who evaluated the same image set, with precision scores of 0.792207 and 0.885496 and accuracy values ranging from 0.838006 to 0.872274.