École Nationale Supérieure de
Techniques Avancées

Abstract

The classification of hyperspectral images is an active field of research, because this type of image has a crucial role in the medical field and the field of remote sensing. In addition the obstacles to overcome in order to provide a good classification algorithm are multiple which justifies the high activity of this field of research. It is these reasons that led to the choice of this subject of study during this internship. First, it was necessary to study the field of hyperspectral imaging and the different methods of classifying hyperspectral images, using small labeled dataset in order to overcome an obstacle specific to this type of image. Once this initiation was done, I was able to compare several research works dealing with the classification of hyperspectral images. Thanks to this study, I was able to choose a class of algorithms for the classification of hyperspectral images. I was thus able to focus on a category of classification method based on the use of graphs. In order to learn more about my choice of method, a comparison and selection of research papers using graphs was necessary in order to optimize the performance of the selected method. My document review went to the research paper published in 2020 by Danfeng Hong and named Graph Convolutional Networks for Hyperspectral Image Classification. I thus tried to deepen and improve the method of the latter in order to obtain a better accuracy during the classification. Two ways for modification of the initial algorithm then emerged by their ability to improve the classification results. Also, this result made it possible to provide a concrete tool for remote sensing thanks to the algorithm for classifying hyperspectral images developed.