Multi-Scale Vision Transformer-Assisted Banana Disease Diagnosis by Optimized Deep Transfer Learning and XAI-Aided Interpretation

Abstract

Early diagnosis of banana leaf nutrient deficiency and disease identification is essential for protecting crop yield and thus sustaining agricultural productivity. The proposed model multi-feature fused Deep transfer Inception ResNet V2 (MFF-DTIRNetV2) scheme presents a compact, end-to-end automated model which integrates multiscale feature learning, an optimized deep transfer learning classifier, and explainable AI-driven interpretation. The system employs comprehensive preprocessing and Multiscale Vision Transformer (MViT-MFE) for extracting fine-grained spatial-spectral cues from PSFD-Musa banana leaf dataset, followed by refined Inception-ResNetV2 classifier that are optimized with the use of Adaptive Evolutionary Gorilla Troops Optimizer (AEGTO). This scheme supports multi-task analysis, which includes localization of lesions, categorization of disease, and severity grading, with Grad-CAM++ improving interpretability through the meaningful visual explanations. An experimental outcome shows that the suggested scheme attains over 99% accuracy highly enhancing class-wise precision, recall, and MCC while outperforming traditional baseline models compared. These outcomes confirm the robustness of model and reliability for real-time banana crop disease monitoring and broad range of agricultural applications.