Design of an iterative method for disease prediction in finger millet leaves using graph networks, dyna networks, autoencoders, and recurrent neural networks

• Objective : Develop an advanced iterative method for early disease prediction in finger millet leaves, improving agricultural yield and disease management. • Employ graph neural networks (GNNs) to model the spatial relationships between different leaf sections, capturing complex disease spread patterns. • Leverage dyna networks to simulate dynamic interactions in plant health data over time, enhancing the model's predictive capabilities. • Utilize autoencoders to reduce dimensionality and extract critical features from leaf image data, providing a more compact and informative representation. • Implement RNNs to model the temporal evolution of disease symptoms, facilitating more accurate prediction of disease progression. Plant diseases are increasingly becoming a serious threat to food security as well as sustainable agriculture sets. Traditional methods for detecting crop diseases, especially in Finger Millet, are cumbersome with chances of error. Therefore, automated solutions are necessary. This work proposes a comprehensive framework for detection and prediction of the disease in Finger Millet leaves using a combination of Graph Networks, Dyna Networks, Autoencoders, and Recurrent Neural Networks. Here, the proposed model brings about both spatial and temporal dynamics in the progression of the disease for earlier and more accurate detection. The core contributions revolve around basic use of Graph Networks in encoding spatial relationships within the leaf structures and application of Dyna Networks in incorporating time-series data for predicting disease progression. Autoencoders allow efficient feature extraction, which is compressing high-dimensional data to meaningful representations, and RNNs allow for plant health to be monitored in real time. Also, the proposed model surpasses the current models in terms of accuracy reaching 95.6 % precision, recall, and F1-score of over 94 %, within those diseases such as Powdery Mildew, Blast, and Leaf Spot. The computational efficiency was also demonstrated to be superior. As shown, training time was reduced to 12 h and prediction time to 0.035 s per image samples. The result would promote the model in actual time application of precision agriculture through more precise, efficient, and scalable solution of plant disease management.