Hyperspectral image classification using deep neural networks

Abstract

Hyperspectral imaging technology has become of high importance nowadays because of the valuable information it provides. Its ability to acquire images with hundreds of adjacent narrow bands in addition to the spatial information made it possible for experts to analyze the target objects more efficiently, and led to an enormous development especially in the remote sensing field which is the focus of this work. Because of the enormous demand on the hyperspectral imaging in various fields, it became a necessity to provide useful tools in order to completely exploit this technology. The aim of this work is to propose some methods that would help reaching this goal by extracting the features from the hyperspectral data needed to identify and classify all the pixels contained in the image. The proposed methods are based on the spectral and spatial information contained in the hyperspectral image. The first method is concerned with the spectral information alone. The second exploits the spatial information alone. Whereas, the last combines both the spectral and spatial information. The technique used to implement the proposed methods is deep learning, which is a powerful system of artificial neural networks that mimics the human brain. Deep neural networks can pass the input data through a bunch of complicated operations in order to extract the most important features and classify the data accordingly.