Dual-boosted capsule networks: Attention-enhanced routing and feedback learning for leaf disease classification on heterogeneous datasets.

Abstract

Identifying the complex leaf disease is a major challenge in crop production. Traditional Capsule Networks face challenges in capturing the complex features associated with various plant diseases, even though they can effectively retain spatial hierarchies. To overcome challenges, this research presents a novel Capsule Networks framework for the identification of leaf disease and its classification. The research proposes a Feedback-Enabled Capsule Routing and a Capsule-Level Attention mechanism. The Feedback-Enabled Capsule Routing method that refines classification outcomes also introduces top-down contextual signals, whereas the Capsule-Level Attention mechanism dynamically weighs capsule outputs based on their relevance. Incorporating this, we also integrated Primary Capsule Squashing and Dynamic Routing. All methods are grounded in formal mathematical definitions. The XAI techniques are integrated to visualize the model's decision-making process for transparency. Empirical evaluations demonstrate the model's effectiveness in achieving classification accuracies of 98.97% and 98.49% in the PlantVillage and the PLD datasets, respectively. This framework addresses the trust gap in AI-driven agricultural diagnostics, allowing for a transparent and reliable method of disease detection in agricultural applications.