Date of Award

Winter 12-2022

Level of Access Assigned by Author

Open-Access Thesis

Degree Name

Master of Science (MS)

Department

Computer Science

Advisor

Salimeh Yasaei Sekeh

Second Committee Member

Roy Turner

Third Committee Member

Chaofan Chen

Abstract

Hyperspectral image (HSI) semantic segmentation is a growing field within computer vision, machine learning, and forestry. Due to the separate nature of these communities, research applying deep learning techniques to ground-type semantic segmentation needs improvement, along with working to bring the research and expectations of these three communities together. Semantic segmentation consists of classifying individual pixels within the image based on the features present. Many issues need to be resolved in HSI semantic segmentation including data preprocessing, feature reduction, semantic segmentation techniques, and adversarial training. In this thesis, we tackle these challenges by employing ensemble methods for HSI semantic segmentation.

Deep neural networks (DNNs) for classification tasks have been employed in HSI semantic segmentation with great success. The ensemble method in traditional classification is often used to increase performance, but research into applying it to semantic segmentation in HSIs is relatively new. Instead of using a single network approach to classification, the ensemble method employs multiple networks to improve performance. Research into ensemble methods in HSI has seen increased accuracy, but often has higher computational complexity and relies on expensive preprocessing techniques. To showcase the performance increase the ensemble method has on semantic segmentation, we propose the novel flagship model Clustering Ensemble U-Net (CEU-Net). In CEU-Net we (1) use a bagging ensemble technique to reduce the computational complexity, (2) utilize clustering on class labels as an intelligent method of delineating which data goes to each network, thereby making each sub-network an expert on a particular cluster, and (3) implement with or without patching for better data flexibility. It is shown that CEU-Net outperforms existing hyperspectral semantic segmentation methods, achieving better performance with and without patching compared to baseline models.

Semantic segmentation models are vulnerable to adversarial attacks and need adversarial training to counteract them. Adversarial attacks are often intelligent attacks that use the knowledge of a trained classifier to create imperceptible perturbations to hurt classification accuracy. Traditional approaches to adversarial robustness focus on training or retraining a single network on attacked data, however, in the presence of multiple attacks these approaches decrease the performance compared to networks trained individually on each attack. To combat adversarial attacks in HSI semantic segmentation, we propose the Adversarial Discriminator Ensemble Network (ADE-Net) which focuses on attack type detection and adversarial robustness under a unified model to preserve per data-type weight optimally while making the overall network robust. In the proposed method, a discriminator network is used to separate data by attack type into their specific attack-expert ensemble sub-network. The ensemble and discriminator networks are trained together using a unified novel loss function to share information between each network. Our approach allows for the presence of multiple attacks mixed together while also labeling attack types during testing. In this thesis, we experimentally show that ADE-Net outperforms the baseline, which is a single network adversarially trained under a mix of multiple attacks, for popular HSI datasets.

Download

Files over 10MB may be slow to open. For best results, right-click and select "save as..."

Share