Title:
AN ENHANCED DEEP LEARNING FRAMEWORK INTEGRATING VISION TRANSFORMERS-CNN FOR ACCURATE SEED-BORNE DISEASE PREDICTION IN CROPS

Authors:
Deepa Bhadana and Purandhar. N, India

Abstract:
Seed-borne diseases remain a serious threat to world agriculture and, frequently, they cause extensive yield losses and compromised food safety. Conventional diagnosis mainly involves manual visual inspection and laboratory methods like culturing and PCR that are labour-intensive, time-consuming, and susceptible to human error. These approaches are also not very scalable and lack adequacy in diagnosing diseases early, particularly if visual signs are not striking. Current machine learning-based image-based models mostly use only phenotypic data and hence do not capture essential genetic signals of susceptibility or resistance to diseases. For overcoming these deficiencies, a deep learning model involving Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) is introduced here for seed-borne disease prediction accurately. CNNs are used to extract local visual features, whereas ViTs extract global contextual information using selfattention operations. At the same time, genomic features are embedded into dense feature spaces and passed through a neural network to extract disease-related genotypic features. These visual and genomic streams are combined through a cross-attention operation to produce joint representations, allowing the model to effectively relate phenotypic and genotypic clues. The model is tested on a dataset of high-resolution seed images and genomic sequences. Experimental outcomes reveal an improved predictive accuracy over the standard single-modality models. The method presented in this paper records an accuracy boost of about 12.5% over standard CNNbased models and also provides improved precision, recall, F1-score, and AUC scores. In addition, attention-based interpretation mechanisms yield visual and genomic saliency maps to support domain experts in interpreting the decision-making process. This combined strategy provides a scalable and generalizable solution for early disease detection in agricultural systems, with potential applications in smart farming, crop monitoring, and plant breeding.