SSDC-DenseNet: A Cost-Effective End-to-End Spectral-Spatial Dual-Channel Dense Network for Hyperspectral Image Classification

Abstract

In recent years, various deep learning-based methods have been applied in hyperspectral image (HSI) classification. Among them, spectral-spatial approaches have demonstrated their power to yield high accuracies. However, these methods tend to be computationally expensive. Specifically, two classic ways to develop spectral-spatial approaches both suffer from significant limitations in cost reduction: multi-channel networks need a large parameter scale, and 3-D filters are inherent of computational complexity. To establish a cost-effective architecture for both training cost and parameter scale, while maintaining the high accuracy of spectral-spatial techniques, an end-to-end spectral-spatial dual-channel dense network (SSDC-DenseNet) is proposed. To explore high-level features, the densely connected structure is introduced to enable deeper network. Furthermore, a 2-D deep dual channel network is applied to replace the expensive 3-D filters to reduce the model scale. The experiments were conducted on three popular datasets: the Indian Pines dataset, University of Pavia dataset, and Salinas dataset. The results demonstrate the competitive performance of the proposed SSDC-DenseNet with respect to classification performance and computational cost compared with other state-of-the-art DL-based methods while obtaining a remarkable reduction of computational cost.

Keywords

• Deep learning
• densely connected convolutional neural network
• feature extraction
• hyperspectral image classification
• multi-scale filter bank