CBAM-LeafGAN: A selective attention-guided StyleGAN framework for mango leaf disease image synthesis and explainable recognition

The development of robust deep learning systems for plant disease diagnosis is constrained by limited expert-annotated datasets, class imbalance, and insufficient visual diversity, particularly in real-world field mango leaf disease classification. To address these challenges, this paper proposes CBAM-LeafGAN, a Selective Residual-Gated CBAM-enhanced StyleGAN3-t framework for disease-specific image synthesis and augmentation. Lightweight Residual-Gated Convolutional Block Attention Modules are selectively integrated into the intermediate synthesis layers to enhance disease-relevant feature generation while preserving alias-free synthesis. Evaluated on the MangoLeafDS2025 dataset, CBAM-LeafGAN was compared with StyleGAN3-t, StyleGAN3-r, Vanilla Latent Diffusion Model, DiffusionPix2Pix, Variational Autoencoder, and LeafGAN. It achieved the lowest Fréchet Inception Distance (FID) of 12.44 among the evaluated models. Ablation, t-SNE, nearest-neighbour diversity, Turing tests, and statistical analyses further validated the realism and diversity of the images. Synthetic augmentation improved disease classification across 23 deep learning architectures and mitigated class imbalance. Evaluation of the proposed approach on the comparable dataset MangoLeafBD further demonstrated improved performance over state-of-the-art methods. Grad-CAM visualisation and quantitative explainability analysis confirmed the consistent localisation of disease-relevant regions, supporting the reliability and interpretability of the proposed framework for mango leaf disease diagnosis under data-constrained agricultural conditions and indicating its potential for practical field deployment.