The Efficiency of Transfer Learning and Data Augmentation in Lemon Leaf Image Classification

Abstract

Leaf diseases in trees and plants are important factors that directly affect the yield of agricultural products. This problem may cause a decrease in the production capacity and profitability of farmers. For this reason, computer-aided detection and classification systems are needed to accurately detect plant diseases. In recent years, learning algorithms and image-processing techniques have been used effectively in the agricultural sector. In this study, the efficiency of transfer learning and data augmentation methods on a dataset consisting of lemon leaf images is examined and the classification of diseased and healthy lemon leaf images is performed. In our study, VGG16, ResNet50, and DenseNet201 transfer learning methods were applied both with and without data increment, and the effect of data augmentation on performance was evaluated. Among the deep transfer learning methods used, DenseNet201 gave the highest accuracy rate with 98.29%. This study shows that transfer learning methods can effectively distinguish between diseased and healthy lemon leaves. It has also been observed that data augmentation does not always provide performance improvement. In future studies, it is predicted that it will be appropriate to evaluate the effect of data augmentation more effectively by applying deep transfer learning methods to plants with different class numbers.