DIRS Laboratory 76-3215
April 25, 2018 at 12:00pm
Ph.D. Thesis Defense



Deep-learning frameworks have made remarkable progress thanks to the creation of large annotated datasets such as ImageNet, which has over one million training images.  Although this works well for color (RGB) imagery, labeled datasets for other sensor modalities (e.g., multispectral and hyperspectral) are minuscule in comparison.  This is because annotated datasets are expensive and man-power intensive to complete; and since this would be impractical to accomplish for each type of sensor, current state-of-the-art approaches in computer vision are not ideal for remote sensing problems.  The shortage of annotated remote sensing imagery beyond the visual spectrum has forced researchers to embrace unsupervised feature extracting frameworks.  These features are learned on a per-image basis, so they tend to not generalize well across other datasets.  In this dissertation, we propose three new strategies for learning feature extracting frameworks with only a small quantity of annotated image data; including 1) self-taught feature learning, 2) domain adaptation with synthetic imagery, and 3) semi-supervised classification.  ``Self-taught''  feature learning frameworks are trained with large quantities of unlabeled imagery, and then these networks extract spatial-spectral features from annotated data for supervised classification.  Synthetic remote sensing imagery can be used to boot-strap a deep convolutional neural network, and then we can fine-tune the network with real imagery.  Semi-supervised classifiers prevent overfitting by jointly optimizing the supervised classification task along side one or more unsupervised learning tasks (i.e., reconstruction).  Although obtaining large quantities of annotated image data would be ideal, our work shows that we can make due with less cost-prohibitive methods which are more practical to the end-user.