Dual-Branch Subpixel-Guided Network for Hyperspectral Image Classification

TL;DR

The paper tackles hyperspectral image classification under mixed-pixel conditions by proposing DSNet, a dual-branch network that jointly learns subpixel abundances via a deep autoencoder unmixing network and pixel-level class features via a CNN classifier, unified through a subpixel fusion module. The learning objective combines a reconstruction loss based on spectral angle distance $L_{RE}$ with a cross-entropy loss $L_{CE}$, balanced by $L = \lambda L_{RE} + (1-\lambda) L_{CE}$, enabling unsupervised abundance estimation to guide supervised classification. Evaluations on Indian Pines, Berlin, and Augsburg show DSNet outperforms state-of-the-art DL-based HSI methods, with ablation studies demonstrating the benefits of nonlinear mixing modeling and subpixel fusion for robust class separation. The work suggests practical impact for real-world HSI tasks by better handling mixed pixels and offering a path toward integrating richer spectral-spatial priors into end-to-end learning.

Abstract

Deep learning (DL) has been widely applied into hyperspectral image (HSI) classification owing to its promising feature learning and representation capabilities. However, limited by the spatial resolution of sensors, existing DL-based classification approaches mainly focus on pixel-level spectral and spatial information extraction through complex network architecture design, while ignoring the existence of mixed pixels in actual scenarios. To tackle this difficulty, we propose a novel dual-branch subpixel-guided network for HSI classification, called DSNet, which automatically integrates subpixel information and convolutional class features by introducing a deep autoencoder unmixing architecture to enhance classification performance. DSNet is capable of fully considering physically nonlinear properties within subpixels and adaptively generating diagnostic abundances in an unsupervised manner to achieve more reliable decision boundaries for class label distributions. The subpixel fusion module is designed to ensure high-quality information fusion across pixel and subpixel features, further promoting stable joint classification. Experimental results on three benchmark datasets demonstrate the effectiveness and superiority of DSNet compared with state-of-the-art DL-based HSI classification approaches. The codes will be available at https://github.com/hanzhu97702/DSNet, contributing to the remote sensing community.

Figures (9)

• Figure 1: Illustration to clarify the similarities and differences between the existing HSI classification method and the subpixel-guided HSI classification method using DL. (a) Workflow for the existing DL-based HSI classification method. (b) Workflow for the subpixel-guided HSI classification method.
• Figure 2: The framework of the proposed DSNet, including deep AE unmixing network, CNN-based classifier network and subpixel fusion module. The deep AE unmixing network is designed by considering a general mixing decoder with physically nonlinear constraints, and further extract useful subpixel-level abundance information from the HSI in an unsupervised manner. The CNN-based classifier network extracts the spectral-spatial information within the HSI to obtain pixel-level class features. The subpixel fusion module aims at integrating the abundance information and class features to ensure high-quality information fusion and enhance model representation capability.
• Figure 3: False-color image, GT and classification maps obtained by different methods on the Indian Pines dataset.
• Figure 4: False-color image, GT and classification maps obtained by different methods on the Berlin dataset.
• Figure 5: False-color image, GT and classification maps obtained by different methods on the Augsburg dataset.