Abstract
Deep Convolutional Neural Networks (CNNs) are the heart of deep neural network research and have accomplished remarkable masterstrokes in various domains. In this research, a new attention-driven deep neural network approach is proposed to localize and classify late blight crop disease by improving the embedding network performance using visual attention. Early disease detection is critical to plan expeditiously and reduce crop losses. Late blight diagnosis is still mainly performed physically. Additionally, this work also expands the visualized gradient-based method by tiding over the gap and facilitates the exploration of the rich dynamics of the core behavioral trained CNNs to identify the unusual patterns that are hidden behind huge data to localize the target object. Studies on decision support for target localization have increased drastically achieving very significant results using deep learning techniques that still fail to explain the black box. As proof of concept, we applied our approach assessing the ongoing benchmark expertly curated images on healthy and late blight contaminated harvests through the current online stage PlantVillage. Experimental comparative results show that our proposed approach achieves a test accuracy of 98.99%, which offers higher localization with classification and assists with taking care of the issue of yield misfortunes in harvests because of irresistible infectious diseases. The result of this research will upgrade the implementation of a deep neural network for early disease diagnosis and management in the agricultural field.