Active learning for hyperspectral image classification
Obtaining labeled data for supervised classification of remotely sensed imagery is expensive and time consuming. Further, manual selection of the training set is often subjective and tends to induce redundancy into the supervised classifier. Active learning (AL) integrates data acquisition with the classifier design by ranking the unlabeled data to provide advice for the next query that has the highest training utility. Thus, it explores the maximum potential of the learner toward both the labeled and unlabeled data, and the training set can be maintained as small as possible by focusing only on the most informative samples for the learning task. This potentially leads to greater exploitation of the information in the data, while significantly reducing the cost of data collection. Due to its promising advantages, active learning has been studied for various applications, such as document retrieval and natural language processing. However, research in active learning has been very limited in remote sensing. In this dissertation, three major active learning frameworks are proposed for hyperspectral image classification, including multi-view adaptive maximum disagreement based active learning, local proximity data regularization based active learning, and critical class oriented active learning. Experiments were conducted on both the AVIRIS and Hyperion hyperspectral data sets. Classification performance was superior to Random Sampling (RS), and SVM based margin sampling SVM-MS, a state-of-the-art active learning method. ^
Melba Crawford, Purdue University.
Engineering, Civil|Engineering, Electronics and Electrical|Computer Science