Deep gradient explainable framework with ensemble adaptive multi-scale feature fusion for plant disease detection

Accurate detection of crop disease is essential to improve crop productivity and produce healthy products. Convolution based approach has proven excellent accuracy for plant disease detection. Existing methods depend on backbone architecture and lack interpretability, limiting their robustness and practical applicability. Capturing both fine-grained texture patterns and multi-scale contextual information is a challenging task for visually similar disease classes. To address these challenges, proposed study introduces novel deep ensemble learning framework that integrates residual and multi-scale feature extraction with explainable artificial intelligence. Features extracted with a deep hierarchical and multi-scale learning network are fused using an adaptive attention-based feature-level fusion strategy. Based on fused features, the proposed Deep Gradient eXplainer (DGX) generates proper visual explanations and disease-relevant regions. The proposed framework is evaluated on benchmark PlantVillage and custom plant disease datasets and achieves a classification accuracy of 99.31%. Extensive analysis demonstrates the effectiveness of the proposed feature fusion technique with DGX. This improves classification accuracy with interpretability. The proposed DGX validates that the model focuses on biologically meaningful regions, which helps to identify the correct disease.