Hybrid augmentation for multi-channel deep learning in guava leaf disease detection

Guava (Psidium guajava) faces significant threats from leaf diseases that compromise its yield and quality. Traditional diagnostic methods that rely on manual inspection are inefficient and subjective, necessitating automated solutions. This study introduces a robust ensemble deep learning framework for guava leaf disease classification by combining hybrid data augmentation with advanced architectures to address the variability in environmental conditions and imaging. A dataset of 2,063 guava leaf images in five categories was expanded using traditional geometric augmentation and synthetic GAN-generated images. Seven state-of-the-art deep learning models, and Vision Transformer B16) were evaluated, with InceptionV3 (92.50% accuracy on GAN data) and ResNet50 (93.12% accuracy on augmented data) selected for their complementary strengths. A multi-channel model fused their features, achieving a 97.50% test accuracy, 0.975 F1-score, and 0.9934 AUC. The results demonstrate the viability of integrating CNNs with transformer architectures under unified augmentation strategies, thereby offering a scalable solution for real-time field deployment.