SwinGhost-ClustNet: An explainable deep ensemble model for papaya leaf disease detection and field deployment in Bangladeshi agriculture

Early detection of papaya leaf disease is essential for Bangladeshi farmers who experience 39.9% post-harvest losses and pesticide misuse risks. The proposed research presents SwinGhost-ClustNet, which is an ensemble of Swin Transformer to capture the global context, GhostNet to capture the local textures, cross-attention fusion, and K-means clustering (k=8, silhouette score 0.67) on a 9,342-image dataset across eight classes (anthracnose, bacterial spot, healthy, leaf curl, mealybug, mite, mosaic, ring spot) in eight districts. The model reached 99.25% accuracy, 99.28% precision, 99.25% recall, 99.25% F1-score, and 1.000 ROC-AUC on 1,401 test images, which is 2.10% higher than base models using pseudo-labels and cross-attention (ablation-validated). Explainable AI (Grad-CAM++ IoU 0.72, Layer-CAM 0.68, and Grad-CAM 0.65) visualizes disease-specific features such as lesions and webbing, and enables building farmer trust. An API based on Flask provides real-time diagnostic, confidence ratings (>70% criterion), and recommendations of pesticides in the Bengali language, which makes it possible to deploy it to the field regardless of single-label restrictions. Future efforts: severity scoring + dosage recommendations, multi-label classification, model pruning (<100MB), cloud deployment (AWS/GCP), and multi-crop validation with domain adaptation (e.g., 98.89% on lemon leaves; Section 4.7) to address domain shift and species differences. This promotes precision farming, which minimizes losses and excess of chemicals in Bangladesh.