International Journal of Scientific Research and Engineering Development

Detection of plant diseases in the field of agriculture is a high-priority issue in terms of the safety of crops and the availability of food products. The identification of disease in the wild is complicated by a complex background clutter, varying luminance, occlusion, scale differences and subtle inter- class visual differences not to mention that the identification of disease in the wild is not in controlled laboratory settings. Under these free conditions, early and accurate detection is necessary to facilitate an intervention in good time and minimize the loss of yield. Recent literature has investigated convolutional neural networks and vision transformer-based design to identify and recognize plant disease and localization. Although CNN-based models prove to be very effective in the local feature extraction, they usually fail to penetrate long-range contextual dependencies and time variation. Vision Transformer and multitask learning methods are better global feature models but have many dis- advantages, such as high-cost computation and poor resistance in use on fine-grained disease patterns in the field. In addition, current approaches are to a large extent based on single scale representations and do not provide efficient means to combine sequential or hierarchical dependencies between features leading to diminished generalization capability in real-life context. The paper uses a Multi-Scale RNN-Transformer Hybrid Model to solve theseconstraints to provide a fine-grained plant disease recognition in the wild. An attention mechanism that is based on transformer is introduced to improve the learning of global context and discriminative features. This hybrid structure is useful in effectively integrating local detail sensitivity, modeling time dependency and long-range contextual representation and is useful in recognizing diseases under unconstrained field condi- tions. The experimental analyses made using the real-world plant disease data show that the suggested model clearly outperforms the current CNN and transformer-based models.

Bathula Prasanna Kumar, P.Gnana Sai Jayanth, G. Venkateswara Rao, M.Lakshmi Ganesh, Y.Sankar,"Multi-Scale RNN–Transformer Hybrid Model for Fine-Grained Plant Disease Recognition in the Wild" International Journal of Scientific Research and Engineering Development, V9(2): Page(822-827) Mar-Apr 2026. ISSN: 2581-7175. www.ijsred.com. Published by Scientific and Academic Research Publishing.