RICE-TDL: Leveraging on Deep Transfer Learning Architecture for Improved Multi-Class Classification of Rice Leaf Diseases

The paper presents a deep transfer learning-based approach for detecting and classifying rice leafdiseases, driven by the need to improve the accuracy and reliability of traditional visual inspection methods,which are often prone to errors due to overlapping disease symptoms common in tropical regions like Nigeria.The methodology involved collecting 200 images of four major rice diseases like bacterial leaf blight, rice blast,brown spot, and false smut collected from Kaggle and expanding the dataset to 1,960 images using standardimage augmentation techniques (flipping, rotation, shearing, and zooming ) to enhance the dataset's diversity andgeneralization potential. The dataset was split into 80% training, 10% validation, and 10% testing. Two modelswere developed: a custom CNN model and a pre-trained VGG-16 model. The custom CNN achieved 68%accuracy on the original dataset, with False Smut Disease recording the best metrics (68.00% accuracy, 66.70%precision, 65.20% recall, and 65.90% F1-score). The VGG-16 model outperformed it with 74.00% accuracy anda 71.40% F1-score for False Smut. When tested on the augmented dataset, the optimized VGG-16 modeldemonstrated a significantly improved accuracy of 99.55% for False Smut, highlighting its robustness andeffectiveness. The system is implemented as a user-friendly web application, enabling farmers to upload imagesfor instant disease diagnosis, thereby offering a practical and scalable solution for enhancing rice diseasemanagement in Nigeria.