Classification of leaf images using machine learning methods

Leaf classification studies are highly beneficial in fields such as agriculture, determination of biodiversity in a region, accurate identification of plant species, pharmacology, forestry, and food security. In this thesis study, systems have been developed to identify plant classes from the leaf images of 12 different plant categories. Convolutional Neural Network (CNN) architectures, namely AlexNet, VGGNet16, and GoogleNet, have been used for the creation of these systems. The dataset has been divided into two parts: a training set and a test set. Using the training data, both a standard AlexNet and a pre-trained AlexNet, previously trained on a very large image dataset, were trained. The pre-trained AlexNet network was significantly more successful than the one trained from scratch. For this reason, pre-trained networks were also utilized when creating the VGGNet16 and GoogleNet models. Before training the networks, the number of images in the training set was increased through data augmentation. In other words, during each epoch of the training, new images were generated by applying scaling, translation, and rotation operations. This process has improved the models' success by enhancing their generalization capabilities. Additionally, new systems were created using different parameters for the networks, and their performance was measured and compared on the test set using Accuracy, F1-score, and ROC curve metrics. As a result of this thesis study, it is demonstrated that the pre-trained AlexNet network is more successful than the network trained from scratch. The F1-score success rates of the best-performing AlexNet, VGGNet16, and GoogleNet networks, configured with different parameters, have been calculated as 96.8%, 98.8%, and 99.3%, respectively.