Dual-branch dense residual network for hyperspectral imagery classification. Issue 7 (2nd April 2020)

• Record Type: Journal Article
• Authors: Wang, Yuhao; Liang, Binxiu; Ding, Meng; Li, Jiangyun
• Is Part Of: International journal of remote sensing. Volume 41:Issue 7(2020)
• Page Start: 2581
• Page End: 2602
• Publication Date: 2020-04-02
• Subjects: Remote sensing -- Periodicals; Télédétection -- Périodiques; 621.3678
• DOI: 10.1080/01431161.2019.1693078
• Languages: English
• ISSNs: 0143-1161

Abstract

ABSTRACT: Hyperspectral imagery (HSI) classification is an essential task in remote sensing. In recent years, convolutional neural network (CNN) applied to HSI has achieved significant improvement compared with the traditional machine learning approaches. However, three main issues still exist in applying CNN. First, excessively increasing the number of layers in CNN will hamper the convergence of network. Second, most existing methods ignore the correlated information among different layers. Third, there still exist challenges when jointly extracting spectral and spatial features. To address these problems and improve the performance, we propose a novel dual-branch dense residual network (DRN), where one branch based on 1D convolution is used to extract the spectral features, the other branch based on 2D convolution is used to extract the spatial features, which effectively make use of the complementary benefits of the spectral and spatial information. Moreover, DRN utilized in each branch is designed to ease the training procedure of network by residual units, and the dense structure is introduced to fuse the information of different convolution layers. We conducted experiments on four well-known hyperspectral imagery datasets, i.e. Indian Pines, Pavia University, Salinas and Pavia Centre. The promising experimental results demonstrate that our dual-branch dense residual network achieved superior classification performance compared with other state-of-the-art methods. Abbreviations: The following abbreviations are used in this manuscript:CNN: Convolutional Neural Network; HIS: Hyperspectral Imagery; RU: Residual Unit; DRN: Dense Residual Network; PCA: Principal Component Analysis; ICA: Independent Component Analysis; LADA: Locality Adaptive Discriminant Analysis; SIFT: Scale-invariant Feature Transform; LBP: Local Binary Pattern; HOG: Histogram of Oriented Gradient; KNN: K Nearest Neighbours; RF: Random Forest; SVM: Support Vector Machine; AEs: Automatic Encoders; GSR: Group Sparse Representation; DD: Diverse Density; ISBDD: Instance Space-based Diverse Density; MCNN: Multiscale Convolutional Neural Network; DCPN: Deep Cube-Pair Network; FDSSC: Fast Dense Spectral-spatial Convolution; SSUN: Spectral-spatial Unified Network; AVIRIS: Airborne Visible Infrared Imaging Spectrometer; OA: Overall Accuracy.