Md. Zasim Uddin, Md. Nadim Mahamood, Ausrukona Ray, Md. Ileas Pramanik, Fady Alnajjar, Md Atiqur Rahman Ahad
Rice is one of the most important staple crops globally. Rice plant diseases can severely reduce crop yields and, in extreme cases, lead to total production loss. Early diagnosis enables timely intervention, mitigates disease severity, supports effective treatment strategies, and reduces reliance on excessive pesticide use. Traditional machine learning approaches have been applied for automated rice disease diagnosis; however, these methods depend heavily on manual image preprocessing and handcrafted feature extraction, which are labor-intensive and time-consuming and often require domain expertise. Recently, end-to-end deep learning (DL) models have been introduced for this task, but they often lack robustness and generalizability across diverse datasets. To address these limitations, we propose a novel end-to-end training framework for convolutional neural network (CNN) and attention-based model ensembles (E2ETCA). This framework integrates features from two state-of-the-art (SOTA) CNN models, Inception V3 and DenseNet-201, and an attention-based vision transformer (ViT) model. The fused features are passed through an additional fully connected layer with softmax activation for final classification. The entire process is trained end-to-end, enhancing its suitability for real-world deployment. Furthermore, we extract and analyze the learned features using a support vector machine (SVM), a traditional machine learning classifier, to provide comparative insights. We evaluate the proposed E2ETCA framework on three publicly available datasets, the Mendeley Rice Leaf Disease Image Samples dataset, the Kaggle Rice Diseases Image dataset, the Bangladesh Rice Research Institute dataset, and a combined version of all three. Using standard evaluation metrics (accuracy, precision, recall, and F1-score), our framework demonstrates superior performance compared to existing SOTA methods in rice disease diagnosis, with potential applicability to other agricultural disease detection tasks.
