A Deep Learning Approach for Coffee Leaf Disease Classification with Centroid-Based Loss Optimization

As global demand for coffee increases, coffee trees, particularly their leaves, suffer more from diseases and pests. Leaf health, being the central organ for photosynthesis, directly impacts the yield and quality of coffee. Therefore, developing a computationally efficient and scalable method for disease and pest management in coffee trees has become imperative. This study introduces a centroid-based loss adjustment method to enhance the focus on distant samples from their class centroid. It involves calculating the centroid for each class and adjusting the loss according to the distance between the sample features and their class centroid. This method is essential concerning the spatial distribution of features, which helps improve classification accuracy and model robustness when dealing with complex data. The experiments use the BRACOL and RoCoLe datasets containing images of multiple coffee leaf diseases. Multiple Deep Learning pre-trained models were tested to validate the proposed method's effectiveness. The experimental results show that this method significantly enhances the accuracy of coffee leaf disease classification and surpasses State-of-The-Art methods. This study presents a method to classify and estimate coffee leaf disease severity, improving model performance without greatly increasing the computational burden.