A new method for prediction of Vigna mungo millet disease based on deep learning

Various viral illnesses impact plant development, causing farmers to lose a lot of revenue. Early diagnosis and prediction of these viral infections can help farmers take preventive measures and mitigate the impacts on crop productivity and quality. As a result, there is a need to develop automated tools for identifying viral infections in crops that analyze symptoms at various parts of the plant. The prediction of Vigna Mungo millet disease is critical for food security and agricultural sustainability. In this article, a practical and reproducible pipeline is proposed for the automatic detection of leaf diseases in Vigna mungo, which combines ImageNet-pretrained CNN backbones (GoogleNet, MobileNetV2, Xception) with a lightweight recurrent classifier. Our original contribution is to treat convolutional feature maps as ordered spatial sequences and to use a single-layer LSTM to model spatial dependencies across the leaf surface. This design more effectively captures the diffuse and irregular lesion patterns characteristic of viral infections. To address the modest dataset size (660 images, with 220 images per class), we freeze the backbones, apply augmentation on the fly, and utilize dropout, gradient clipping, and early stopping. The models were evaluated with stratified 5-fold cross-validation and statistical tests. It has been revealed that the Xception with LSTM attained the best mean performance (98.34% ± 0.34% across folds; peak 98.48% on the test split). Vigna Mungo/ Black gram plant leaf diseases can significantly reduce crop yields, leading to lower food production and higher food prices. By detecting and identifying these diseases early on, farmers can take appropriate measures to control the spread of the disease and prevent crop losses. • We proposed a hybrid Deep Learning for leaf disease detection of Vigna Mungo plant. • A hybrid Deep Learning model (GoogleNet, MobileNetV2, and Xception) with RNN is designed. • The Xception-RNN network achieved the highest accuracy of 98.48%. • The suggested approach forecasts the health of a plant's leaves and categorizes them into healthy, anthracnose, and yellow mosaic.