Deep Learning for Hyperspectral Image Classification: An Overview

Hyperspectral image (HSI) classification has become a hot topic in the field\nof remote sensing. In general, the complex characteristics of hyperspectral\ndata make the accurate classification of such data challenging for traditional\nmachine learning methods. In addition, hyperspectral imaging often deals with\nan inherently nonlinear relation between the captured spectral information and\nthe corresponding materials. In recent years, deep learning has been recognized\nas a powerful feature-extraction tool to effectively address nonlinear problems\nand widely used in a number of image processing tasks. Motivated by those\nsuccessful applications, deep learning has also been introduced to classify\nHSIs and demonstrated good performance. This survey paper presents a systematic\nreview of deep learning-based HSI classification literatures and compares\nseveral strategies for this topic. Specifically, we first summarize the main\nchallenges of HSI classification which cannot be effectively overcome by\ntraditional machine learning methods, and also introduce the advantages of deep\nlearning to handle these problems. Then, we build a framework which divides the\ncorresponding works into spectral-feature networks, spatial-feature networks,\nand spectral-spatial-feature networks to systematically review the recent\nachievements in deep learning-based HSI classification. In addition,\nconsidering the fact that available training samples in the remote sensing\nfield are usually very limited and training deep networks require a large\nnumber of samples, we include some strategies to improve classification\nperformance, which can provide some guidelines for future studies on this\ntopic. Finally, several representative deep learning-based classification\nmethods are conducted on real HSIs in our experiments.\n