Identifying species of trees through bark images by convolutional neural networks with transfer learning method

Abstract

Identifying trees by images of barks via Deep Learning method has a potentially useful contribution to many areas, such as revision of forests, preparation of sustainable management plans for forest resources, operations and processing of trees for paper and furniture industries, preservation of trees having vital importance to environments, definition of species and sub-species of fruits for orcharding, for amateur purposes, and entirely for handling tree sources efficiently. Even though the current progress in Deep Learning has proven to be impressive, the lack or insufficiency of datasets has limited the use of Deep Learning on identification of tree species from barks images. In order make contribution to the researches on this field, and to prove that tree identification via images of barks with high accuracy is possible, 24686 bark images of 59 tree species from different parts of Turkey has been collected within a span of a year, and the data set is used for this work. With the use of seven pre-trained convolutional neural networks, AlexNet, DenseNet201, ResNet18, ResNet50, ResNet101, VGG16, VGG19. It has been demonstrated that identification of tree species by images of barks is possible through transfer learning method. Additionally, it has been inferred that transfer learning method provides fast and accurate solutions to classification problems. Furthermore, the impact of the depth, layer, number of parameters and batch size of the networks has been analyzed. While the average accuracy of all the networks, regarding the ratio of number of images and training data, is between 93.21% and 95.89%, the average of accuracy of the two most successful networks is 99.46%.