Метод виявлення та розпізнавання хвороб пшениці на основі аналізу зображень

• Datum: 2025
• 1. Verfasser: Smovzh, Serhii
• Format: Artikel
• Sprache: Ukrainian
• Veröffentlicht: V.M. Glushkov Institute of Cybernetics of NAS of Ukraine 2025
• Schlagworte: класифікація; дисбаланс класів; XAI; EfficientNetV2B0; фітопатологія
• Online Zugang: https://jais.net.ua/index.php/files/article/view/529
• Назва журналу: Problems of Control and Informatics

Wheat diseases such as rust, septoria, and powdery mildew cause significant crop losses and pose a serious threat to global food security. As wheat is one of the most important staple crops, providing over 20 % of the daily caloric intake for the average person, timely and accurate disease diagnosis is of critical importance. Traditional methods for identifying phytopathologies are labor-intensive, prone to subjectivity, and often lack sufficient accuracy. At the same time, advancements in deep learning technologies have enabled the development of automated systems capable of performing rapid and scalable image-based diagnostics in real-field conditions. In this study, a custom dataset was created, which revealed the challenge of class imbalance—a key factor that significantly reduced the classification accuracy for diseases with limited image samples. To address this issue, a two-stage classification method was developed. During the preparation phase, classes were grouped into large, medium, and small categories based on the number of training images. In the first stage, a model determines which group the input image belongs to. In the second stage, a corresponding specialized model is invoked to identify the specific disease within the identified group. Additionally, explainable artificial intelligence (XAI) mechanisms were implemented, specifically Grad-CAM, which allows for visualizing and interpreting the model’s diagnostic decisions. The conducted experiments confirmed the high effectiveness of the proposed method: the average classification accuracy reached 88 %, with over 90% accuracy achieved for the numerically dominant classes. These results form a strong foundation for the practical deployment of the method in agricultural enterprises for real-time crop monito­ring. Future research directions include adapting models to varying imaging conditions, developing lightweight architectures suitable for mobile devices, and further enhancing techniques for handling data imbalance.