Enhancing the performance of transferred efficientnet models in leaf image-based plant disease classification

Plant-leaf diseases have become a significant threat to food security due to reducing the quantity and quality of agricultural products. Plant disease detection methods are commonly based on experience through manual observations of leaves. Developing fast, accurate, and automated techniques for identifying of crop diseases using computer vision and artificial intelligence (AI) can help overcome human shortcomings. In the current study, the EfficientNet architectures with pre-trained Noisy-Student weights were implemented using the transfer learning approach to classify leaf image-based healthy and diseased plant groups. The deep learning models were performed on the extended and enhanced PlantVillage datasets, consisting of leaf images of 14 different plant species, with background and augmented images. Early stopping and learning rate scheduler techniques were used to speed up learning and improve the efficiency of the training and testing process. The experimental results obtained on the two test sets showed that the EfficientNet-B3 and EfficientNet-B5 architectures achieved the highest performance metrics on the non-augmented and augmented datasets, respectively. The average testing accuracy of both models was up to 99.997% with good precision and sensitivity. The improved networks also revealed an excellent efficacy compared to several popularly convolutional neural networks in the literature, such as AlexNet, GoogleNet, VGGNet, ResNets, DenseNets, and MobileNets. Growingly, the enhanced artificial intelligence models may provide more powerful and practical solutions to the detection of plant-leaf diseases.