Classification of hyperspectral images by fusion of spectral and spatial features in convolutional neural networks
|
Obeid Sharifi * , Behnam Asghari Beirami , Mehdi Mokhtarzade |
K.N. Toosi University of Technology |
|
Abstract: (2828 Views) |
Hyperspectral images are useful in monitoring the Earth surface phenomena due to the acquisition of large number of spectral bands. Hyperspectral image classification is the most important field of the hyperspectral data processing, and so far, there have been many attempts to increase its accuracy. Convolutional neural networks (CNNs) and spatial features have had a great role in improving the accuracy of the hyperspectral image classification in recent years. In the previous researches not much attention has been paid to the simultaneous use of the capabilities of the low spatial feature deriving methods in convolutional neural networks. For this reason, in the present paper, a new architecture of convolutional neural networks is introduced for the classification of hyperspectral images that uses the different combinations of spectral features and spatial features which are derived from morphological profiles, Gabor filter and local binary pattern (LBP) as input vectors to the proposed CNN. The experiments which are conducted on two real hyperspectral images from agricultural and urban areas, show the superiority of the proposed method (about 2.5%) in comparison to some recent spatial-spectral classification methods. |
|
Keywords: Hyperspectral image classification, convolutional neural networks, morphological profiles, Gabor filter, local binary pattern |
|
Full-Text [PDF 2191 kb]
(963 Downloads)
|
Type of Study: Research |
Subject:
RS Received: 2019/07/29 | Accepted: 2019/12/29 | Published: 2021/10/22
|
|
|
|
|
Send email to the article author |
|