Superpixel Based Active Learning and Online Feature Importance Estimation for Hyperspectral Image Analysis
The rapid development of multi-channel optical imaging sensors has led to increased uti- lization of hyperspectral data for remote sensing. For classification of hyperspectral data, an informative training set is necessary for ensuring robust performance. However, in remote sensing and other image analysis applications, labeled samples are often difficult, expensive and time-consuming to obtain. This makes active learning (AL) an important part of an im- age analysis framework — AL aims to efficiently build a representative and efficient library of training samples that are most informative for the underling classification task. This thesis proposes an AL framework that leverages from superpixels. First, a semi-supervised AL method is proposed that leverages the label homogeneity of pixels in a superpixel, lead- ing to a faster convergence using few training samples. Secondly, a spatial-spectral AL method is proposed that integrates spatial and spectral features extracted from superpixels in an AL framework. The experiments with an urban land cover classification and a wet- land vegetation mapping task show that the proposed methods have faster convergence and superior performance compared to baselines. Importantly, our proposed framework has a key additional benefit in that it is able to identify and quantify feature importance — the resulting insights can be highly valuable to various remote sensing image analysis tasks.