Coffee is one of the most consumed beverages worldwide, with Brazil leading exports in 2024. However, production is threatened by biotic stresses such as cercospora, phoma, rust, and leaf miner, which reduce grain quality and cause major losses. Disease identification is usually performed manually by specialists, a process that is slow, limited, and error-prone. This study proposes automating classification through Convolutional Neural Networks (CNNs) combined with various image processing techniques. The JMunBEN and JMunBEN2 datasets were used for this task, together with image preprocessing techniques such as cropping, augmentation, Gaussian, Clahe, Graythresh and Wavelet. Three CNN architectures—ResNet50, MobileNetV2, and AlexNet—were tested for multiclass classification, trained for 100 epochs with Adam optimizer and CrossEntropy loss. The article proposed wavelet filter yielded the best results, with ResNet achieving 99.07% accuracy, confirming CNNs with image pre-processing techniques to be efficient tools for rapid, reliable detection of coffee biotic stresses.
Yamakawa et al. (Tue,) studied this question.