Attention guided convolutional neural network with explainable AI for papaya leaf disease detection in edge and drone...

Existing disease discovery in papaya leaves is most significant in achieving yield and profitability stability in the tropics but has proven difficult in the presence of deficiencies in manual exploration and tailored crop models in crop-AI systems. Therefore, this study introduces PapayaNet, a lightweight attention-guided convolutional network specifically structured for the automated classification of six papaya leaf states, including major diseases and healthy leaves. For real-world deployment in scarce-resource farming contexts, PapayaNet adopts batch norm and hierarchical attention steps in five convolution stages and accelerates both computational celerity and discriminability. Trained on 6618 manually annotated orchard images sourced from orchards in Bangladesh at a very high resolution, it has a 98.79% classification accuracy, all of which was realized using 483,926 parameters and an average infer time of 0.01 s, which is significantly better when evaluated using EfficientNetB6, DenseNet121, and VGG16. XAI methods, including Grad-CAM and LIME, showed model decisions towards the biologically informative parts of the leaf, thus boosting interpretability and user confidence. Systematic ablation analysis also confirmed the importance of distributed attention in ensuring robust generalization towards visually similar disease classes. An in-browser diagnostic portal deployed using Gradio provides intra-browser predictive deployment and interpretability overlay in real time, thus inviting field practicability. Given its low-latency inference and minimal computational footprint, PapayaNet is well-suited for integration into edge devices and drone platforms, offering a scalable solution for real-time in-situ crop health monitoring. This study advances the field of precision agriculture by delivering a crop-specialized, explainable, and deployable AI system for sustainable management of papaya diseases.