Spectral-spatial classification of hyperspectral imagery using a dual-channel convolutional neural network

Haokui Zhang, Ying Li, Yuzhu Zhang, Qiang ShenRemote 2017 · DOI

Authors: Haokui Zhang, Ying Li, Yuzhu Zhang, and Qiang Shen
Affiliations: Northwestern Polytechnical University, Xi'an, China; Aberystwyth University, Aberystwyth, UK

Abstract

Hyperspectral imagery (HSI) classification faces significant challenges due to high dimensionality and the scarcity of labeled training samples. This paper introduces a dual-channel convolutional neural network (DC-CNN) designed to extract spectral and spatial features simultaneously. By combining a 1D-CNN for spectral signatures with a 2D-CNN for spatial textures, and employing a robust data augmentation strategy, the proposed model achieves state-of-the-art performance on benchmark datasets.

Introduction

Hyperspectral imagery (HSI) provides a detailed "spectral fingerprint" of the Earth's surface by capturing hundreds of narrow, continuous spectral bands. However, classifying these images is difficult due to the curse of dimensionality and the spatial-spectral disconnect. Early methods often ignored spatial context, while later deep learning models frequently required flattening data into 1D vectors, destroying spatial structures. Furthermore, the high cost of labeling HSI data often leads to overfitting in complex models.

Methodology: The DC-CNN Framework

The proposed Dual-Channel Convolutional Neural Network (DC-CNN) addresses these issues by processing target pixels through two specialized parallel paths:

1. Spectral Channel (1D-CNN)

This channel focuses on the chemical and physical signatures of materials. It extracts a $3 \times 3 \times L$ cube (where $L$ is the number of spectral bands) around a target pixel. These are rearranged into pixel vectors, and 1D convolutional kernels slide along the spectral dimension to capture deep spectral features.

Architecture of the 1D-CNN spectral channel

2. Spatial Channel (2D-CNN)

To capture spatial textures and shapes without excessive computational cost, the model first uses Principal Component Analysis (PCA) to compress the data into the three most informative bands. A $41 \times 41 \times 3$ patch is then fed into a 2D-CNN, allowing the model to utilize a larger spatial window to identify complex structures.

Architecture of the 2D-CNN spatial channel

Feature Fusion and Data Augmentation

The high-level features from both channels are concatenated and fed into a Softmax regression layer for final classification. To combat the shortage of training data, the authors implemented a data augmentation strategy—rotating and flipping training patches—which increased the training set size by a factor of six.

Experimental Results

The DC-CNN was evaluated on two widely used HSI benchmarks: the Pavia University scene and the Indian Pines agricultural scene.

  • Pavia University: The model achieved an Overall Accuracy (OA) of 99.68% with data augmentation, compared to 96.57% for a standard CNN.
  • Indian Pines: On this more challenging dataset, the model reached an OA of 98.76%. The results demonstrated that data augmentation is vital for smaller datasets, providing a significant accuracy boost from 96.88%.

Visual analysis of the classification maps showed that DC-CNN produces smoother regions with fewer "salt-and-pepper" noise errors, indicating superior spatial consistency compared to methods like Stacked Sparse Deep Belief Networks (SSDL).

Conclusion and Future Work

The DC-CNN effectively bridges the gap between spectral and spatial feature extraction. While the model achieves high accuracy, future work will focus on reducing computational costs and exploring semi-supervised learning to leverage the abundance of unlabeled HSI data available in real-world applications.

Summary updated

BibTeX
@inproceedings{zhang2017spectral,
  author = {Haokui Zhang and Ying Li and Yuzhu Zhang and Qiang Shen},
  title = {{Spectral-spatial classification of hyperspectral imagery using a dual-channel convolutional neural network}},
  booktitle = {Remote},
  year = {2017},
  doi = {10.1080/2150704x.2017.1280200},
}
Figure