Semi-Supervised Virtual Support Vector Machines with Self-Learning Constraint for Remote Sensing Image Classification

In real-world applications, it is difficult to collect labeled samples, and supervised learning methods rely on the quality of this labeled training data. Therefore, in this research, a semi-supervised learning approach is developed in order to benefit from the unlabeled samples that can be produced effortlessly. These semi-supervised methods are built on a popular machine learning technique called support vector machine, which is used to classify remote-sensing imagery in this thesis. Moreover, this work aims to enhance the accuracy of the methods in settings with very few labeled samples and deploy a constrained set of unlabeled samples with a self-learning strategy. Additionally, the aim includes model evaluation for existing support vectors and virtual samples. Moreover, the methodology is further extended with an active learning method. This extension involves uncertainty visualizations in order to increase the model accuracy by relabelling the uncertain samples in a prioritized way. To evaluate these models, experimental results were obtained over the city of Cologne, Germany, and the Hagadera Refugee Camp, Kenya from a very high spatial resolution multispectral data set. Results from newly proposed methods showed favorable performance properties, especially on the few labeled samples. Furthermore, the uncertainty of the models was compared with the active learning extension, and this extension also increased the accuracy