Customized EfficientNet: A Transfer Learning Approach for Rice Leaf Disease Classification with Advanced Resizing and Augmentation

Accurate and reliable plant disease detection systems have to be implemented for optimizing agricultural profitability and preserving global food security. Deep learning-based image classification has shown significant potential in addressing these challenges. However, its application in resource-limited demands of ecological responsibility solutions that are fast, accurate, and computationally efficient. In order to satisfy these requirements, we have employed a compact transfer learning architecture for rice leaf disease identification that uses nine disease classes. The framework features preprocessing approaches such as illumination correction to enhance image quality. The method we propose integrates a tailored classifier network with a pretrained EfficientNetB0 model for optimal feature extraction and precise classification. With the objective to ensure its efficacy, we evaluate the proposed framework against multiple transfer learning models, including DenseNet121, VGG16, InceptionV3, MobileNetV2, and ResNet50. In accordance with the experimental results, the proposed customized EfficientNetB0 framework exceeds the other frameworks, attaining 98.16% training accuracy and 94.47% testing accuracy. In comparison, DenseNet121, VGG16, InceptionV3, MobileNetV2, and ResNet50 achieved training accuracies of 77.67%, 75.59%, 91.74%, 78.94%, and 93.16%, and testing accuracies of 71.75%, 71.58%, 89.02%, 77.28%, and 89.52%, respectively.