Diagnosis of Paddy Diseases Using Pre-Trained Architectures and a Proposed Enhanced EfficientNetB3 Model

Rice is an important crop in India and is often affected by pests and diseases, which can lead to a significant drop in production. This research investigates advanced deep learning approaches for accurate paddy disease diagnosis, focusing on comparing several transfer learning models. The study specifically targets diseases such as Tungro, Dead Heart, Hispa, Blast, Downy Mildew, Brown Spot, Bacterial Leaf Blight, Bacterial Panicle Blight, and Bacterial Leaf Streak. The base EfficientNetB3 model attains approximately 95.55 % accuracy during training and 95.12% during evaluation on unseen data. However, it encounters challenges when applied to domain-specific tasks such as diagnosing paddy diseases, frequently experiencing issues such as overfitting and inadequate convergence. To overcome these issues, an Enhanced EfficientNetB3 model was developed, incorporating batch normalization, dropout, and data regularization techniques. The training was conducted using the 'Paddy Doctor' dataset, featuring 10,407 high- resolution images of paddy leaves. It reached an accuracy of 98.92 % during training with a loss rate of 0.1385. For validation, the model reached an accuracy of 98.20 % and a loss rate of 0.1450. On an independent test set, the accuracy 98.50 % obtained with a test loss of 0.1505. With remarkable accuracy and a training time of just 68 minutes, the model demonstrates its significant potential for precise paddy disease diagnosis. Its impressive performance plays a crucial role in advancing disease management and boosting crop yields.