RESUMO
Quercus is one of the important elements of forests worldwide. But the diagnosis of the species in this genus in particular using leaves is pretty challenging due to the presence of natural hybrids and phenotypically plastic trait expression. In this sense, this study aims to classify the leaves of Q. vulcanica and Q. frainetto using convolutional neural networks, VGG16 and VGG19, and Xception deep learning architectures to determine which method has the best performance in species assignment. For this purpose, leaf samples were collected from a total of 300 trees of 6 natural populations using a total of 1459 leaf images, 491 from Q. frainetto and 968 from Q. vulcanica. Before exporting images to the deep learning model, RGB/gray filters are applied and images are optimized with contrast limited adaptive histogram equalization to achieve maximum performance in the deep learning model. Accuracy rates of deep learning architectures varied from 79% (Xception) to 95% (VGG16). The VGG16 deep learning model provided superior performance compared to the others. Developing a mobile device using images from natural populations of many oak species will be beneficial not only for practitioners but also for scientists and local people.