Electronic Thesis and Dissertation Repository

Thesis Format

Monograph

Degree

Doctor of Philosophy

Program

Electrical and Computer Engineering

Supervisor

McIsaac, Kenneth

Abstract

This thesis is focused on deep learning-based, pixel-wise classification of hyperspectral images (HSI) in remote sensing. Although presence of many spectral bands in an HSI provides a valuable source of features, dimensionality reduction is often performed in the pre-processing step to reduce the correlation between bands. Most of the deep learning-based classification algorithms use unsupervised dimensionality reduction methods such as principal component analysis (PCA).

However, in this thesis in order to take advantage of class discriminatory information in the dimensionality reduction step as well as power of deep neural network we propose a new method that combines a supervised dimensionality reduction technique, principal component discriminant analysis (PCDA) and deep learning.

Furthermore, in this thesis in order to overcome the common problem of inadequacy of labeled samples in remote sensing HSI classification, we propose a spectral perturbation method to augment the number of training samples and improve the classification results.

Since combining spatial and spectral information can dramatically improve HSI classification results, in this thesis we propose a new spectral-spatial feature vector. In our feature vector, based on their proximity to the dominant edges, neighbors of a target pixel have different contributions in forming the spatial information. To obtain such a proximity measure, we propose a method to compute the distance transform image of the input HSI. We then improved the spatial feature vector by adding extended multi attribute profile (EMAP) features to it. Classification accuracies demonstrate the effectiveness of our proposed method in generating a powerful, expressive spectral-spatial feature vector.

Summary for Lay Audience

In this thesis, we propose a few approaches to perform hyperspectral image (HSI) classification in the field of remote sensing. As opposed to the regular RGB images which consist of three channels of red, green, and blue, an HSI is composed of a series of images each taken at a specific wavelength. In the field of remote sensing, hyperspectral images are collected by the imaging sensors on board of an airplane or a satellite and due to the valuable information that they can provide about the objects and phenomena on our planet, they are employed in many applications. One common practice in HSI processing is the classification of each individual pixel in the image. In other words, in many applications we are interested in assigning each pixel to a specific category. This kind of classification has been an active research area for years for which different approaches have been proposed. Recently, deep learning-based methods have attracted a lot of attention from the research community because of their superior performance compared to the conventional methods. Therefore, in this thesis, we used one of the deep learning frameworks, stacked autoencoder (SAE) to perform the pixel-wise HSI classification task. As our first contribution, we combined SAE with a supervised dimensionality reduction (DR) technique where labels of the samples are used during the DR step. Second, as one of the common issues in processing the remote sensing hyperspectral datasets is the lack of enough labeled data, we proposed a method to generate virtual samples using the available ground truth data. Since combining spectral and spatial information in an HSI, can dramatically improve classification accuracies as our third contribution, in this thesis project we proposed a new method including a novel spectral-spatial feature vector. In our spatial feature vector, effective pixels have different contributions based on an edge proximity measure obtained from the distance transform image of the input HSI. We applied our methods on several remote sensing hyperspectral datasets and evaluated their performance using various accuracy metrics. Classification results show the superiority of our methods compared to several conventional and deep learning-based approaches

Creative Commons License

Creative Commons Attribution-Noncommercial 4.0 License
This work is licensed under a Creative Commons Attribution-Noncommercial-Share Alike 4.0 License.

Share

COinS