Semi Supervised Machine Learning and Deep Learning Based Analysis for Hyperspectral Remote Sensing Images
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Hyperspectral Image Analysis has been an active area of research, especially in scenarios where discriminative features from classes having similar spectral characteristics have to be learned. We propose and implement novel machine learning techniques to address research problems in the field of Hyperspectral Image Analysis using remote sensing images. Each chapter in this dissertation presents a novel method from the field of machine learning with the end goal of robust classification of Hyperspectral Remote Sensing Images. We describe common problems faced in the field of Hyperspectral Image Analysis, and address those problems by proposing novel techniques. One common problem is the lack of large quantities of labeled data, which leads to the problem of models over fitting to the limited number of labeled training samples. We propose a spatial-spectral unsupervised feature extraction / reduction approach in Chapter 2 of this dissertation. Another approach to address the specific problem of the lack of large quantities of labeled data samples is to use the large number of available unlabeled data samples to perform Semi-Supervised learning. Towards this goal, we propose a Semi-Supervised feature extraction / dimension reduction approach in Chapter 3 of this dissertation. Following the same idea, and inspired by the recent advancements in the field of Deep Learning, we also propose a Semi-Supervised Deep Learning approach in Chapter 4 of this dissertation. Another recent development in the field of Deep Learning for color image analysis involves new variants of neural network architectures called Capsule Neural Networks, which can capture the spatial information along with the underlying context from the original images in a much more robust manner. We propose Semi-Supervised Capsule Neural Networks tailored towards hyperspectral image analysis in Chapter 5 of this dissertation. In the penultimate Chapter of this dissertation, Chapter 6, we propose an algorithm to perform label expansion for Semi-Supervised Deep Learning tasks, applied to the domain of large scale Road Segmentation of big cities (we show our results for Road Segmentation in the city of Las Vegas and Caracas, the capital of Venezuela).