MGTF-Net: a Multi-Granularity Transformer Fusion Network for cross-fruit disease detection in mango and custard apple

Introduction Automated plant disease detection is essential for securing the food supply and reducing agricultural losses in developing countries. Mango (Mangifera indica) and custard apple (Annona squamosa) suffer heavy yield losses from fungal, bacterial, and environmental diseases, while conventional single-backbone classifiers capture disease morphology at only one semantic granularity. Methods We propose MGTF-Net (Multi-Granularity Transformer Fusion Network), which extracts multi-scale features using three heterogeneous backbones (EfficientNet-B4, ResNet-50, and DenseNet-121) with architecture-matched attention (CBAM and SE) and fuses the tokens using a four-layer, eight-head Transformer encoder trained jointly on both species. Results On MCFD-9K (9,500 images across 10 classes), MGTF-Net achieves 98.73% accuracy, 98.74% macro precision, 98.73% macro recall, 98.73% macro F1-score, Cohen's κ of 0.9859, and a mean AUC of 0.9987 on a held-out test set of 1,500 images. McNemar's test shows statistically significant improvement over a conventional transfer-learning CNN ( p < 0.001), while differences from four recent transformer-based baselines are not statistically significant. Discussion Ablation, cross-fruit, input-resolution, and Grad-CAM++ analyses show that multi-backbone fusion and cross-species joint training improve feature representation and localization. The dataset and code are released publicly.