Electronic Thesis and Dissertation Repository


Doctor of Philosophy


Electrical and Computer Engineering

Collaborative Specialization

Planetary Science and Exploration


McIsaac, Kenneth

2nd Supervisor

Osinski, Gordon R.



Autonomous techniques in the context of planetary exploration can maximize scientific return and reduce the need for human involvement. This thesis work studies two main problems in planetary exploration: rock image classification and hyperspectral image classification. Since rock textural images are usually inhomogeneous and manually hand-crafting features is not always reliable, we propose an unsupervised feature learning method to autonomously learn the feature representation for rock images. The proposed feature method is flexible and can outperform manually selected features. In order to take advantage of the unlabelled rock images, we also propose self-taught learning technique to learn the feature representation from unlabelled rock images and then apply the features for the classification of the subclass of rock images. Since combining spatial information with spectral information for classifying hyperspectral images (HSI) can dramatically improve the performance, we first propose an innovative framework to automatically generate spatial-spectral features for HSI. Two unsupervised learning methods, K-means and PCA, are utilized to learn the spatial feature bases in each decorrelated spectral band. Then spatial-spectral features are generated by concatenating the spatial feature representations in all/principal spectral bands. In the second work for HSI classification, we propose to stack the spectral patches to reduce the spectral dimensionality and generate 2-D spectral quilts. Such quilts retain all the spectral information and can result in less convolutional parameters in neural networks. Two light convolutional neural networks are then designed to classify the spectral quilts. As the third work for HSI classification, we propose a combinational fully convolutional network. The network can not only take advantage of the inherent computational efficiency of convolution at prediction time, but also perform as a collection of many paths and has an ensemble-like behavior which guarantees the robust performance.