Automated Potato Tomato Disease Detection Using EfficientNetB3 Transfer Learning

Abstract

Plant disease significantly influences agricultural production performance, particularly in valuable crops such as potatoes and tomatoes. The disease detection should be precise and automated in order to improve crop management and sustainable agriculture. Leaf disease classification for a number of diseases via multi-class approach using EfficientNetB3 architecture is proposed as the optimal structure of deep learning. Weights trained on ImageNet are used to facilitate transfer learning and feature extraction efficiency, while customized classifiers are designed to provide improved generalization, including batch normalization and dropout techniques. Input data diversity and robustness is provided via data augmentation techniques (rotation, translation, and horizontal flips). The adaptive learning approach together with Adamax optimizer are used to train the classifier. The classifier was tested using the PotatoTomato dataset from the PlantVillage repository which consists of six disease and healthy categories. The achieved accuracy score was 99.25%, while achieving precision and recall of various categories at 100%. The good performance in multi-classification proves the soundness of the classifier.