Cosine annealing scheduler added with momentum factor: classification of tomato plant diseases

Plant disease classification is vital for improving agricultural productivity through timely, accurate diagnosis. This study explores deep learning methods for classifying tomato leaf diseases using four Convolutional Neural Network (CNN) architectures: InceptionResNetV2, VGG16, ResNet50, and ResNet152. To enhance training, we employ three optimizers-Adam, RMSprop, and SGD and apply a cosine annealing scheduler to adjust the learning rate cyclically. Momentum is fine-tuned alongside learning rate to improve convergence and reduce overfitting. These strategies allow efficient weight updates and better exploration of the loss landscape. Data augmentation techniques such as blurring, rotation, and sharpening are used to improve data diversity and model robustness. Results show that ResNet50, combined with the Adam optimizer and cosine annealing, achieves the best classification accuracy. Findings highlight that tuning both learning rate and momentum enhances generalization. The proposed approach enables early plant disease detection, supporting smart agriculture and crop protection.