Semantic Segmentation of Satellite Imagery using Positive and Unlabeled Learning.
Loading...
Date
2022-10-04
Authors
Eshghi, Mohammad
Journal Title
Journal ISSN
Volume Title
Publisher
University of Oregon
Abstract
The recent advances of deep learning in computer vision field have revolutionized digital image processing. The adoption of vision-based deep learning models in remote sensing has been promising. However, despite their success in remote sensing image processing, deep learning models suffer from labeled data scarcity, which is defined as the lack of large scale labeled datasets. This drawback is important to pay attention to since manually labeling data is labor-intensive and time-consuming. In addition, in many applications, the only information of interest is the presence of the application-specific landcover or object within an image, and thus it is not reasonable to spend extra time and cost to fully label the rest of an image. Therefore, remote sensing image processing benefits greatly from positive and unlabeled learning, which is a more general setting of semi-supervised learning and addresses the availability of only a few labeled examples of the presence of the application-specific event in a dataset.This dissertation investigates the possibility of leveraging transfer learning and ensemble learning frameworks in a positive and unlabeled learning setting for semantic segmentation of satellite imagery. First, I create positive and unlabeled satellite imagery datasets from an available binary positive and negative dataset to be used for model development. Next, I develop a deep homogenous transfer positive and unlabeled learning which utilizes two distinct positive and negative as well as positive and unlabeled satellite imagery datasets acquired by a same satellite sensor (i.e. similar domain images). Building upon this, I extend the homogenous aspect of the developed model to the heterogeneous case. In doing so, the developed model will be able to not only learn from similar domain satellite images and non-satellite images but also to leverage satellite images from dissimilar domains. In the next stage, I develop a deep ensemble positive and unlabeled learning model in order to incorporate the advantages of multiple different models for a same task. Then, I investigate the possibility of a mixture of the proposed models in transfer learning and ensemble learning frameworks for PU learning. Finally, I conclude this dissertation by discussing the possible next steps for future works.
Description
Keywords
Computer Vision, Deep Learning, Positive and Unlabeled Learning, Remote Sensing, Semantic Segmentation