Quantum Information-Empowered Graph Neural Network for Hyperspectral Change Detection

TL;DR

It is demonstrated that through the unitary feature extraction procedure, QUEEN provides radically new information for deciding whether there is a change or not and the superior HCD performance of the proposed QUEEN-empowered GNN will be experimentally demonstrated on real hyperspectral datasets.

Abstract

Change detection (CD) is a critical remote sensing technique for identifying changes in the Earth's surface over time. The outstanding substance identifiability of hyperspectral images (HSIs) has significantly enhanced the detection accuracy, making hyperspectral change detection (HCD) an essential technology. The detection accuracy can be further upgraded by leveraging the graph structure of HSIs, motivating us to adopt the graph neural networks (GNNs) in solving HCD. For the first time, this work introduces quantum deep network (QUEEN) into HCD. Unlike GNN and CNN, both extracting the affine-computing features, QUEEN provides fundamentally different unitary-computing features. We demonstrate that through the unitary feature extraction procedure, QUEEN provides radically new information for deciding whether there is a change or not. Hierarchically, a graph feature learning (GFL) module exploits the graph structure of the bitemporal HSIs at the superpixel level, while a quantum feature learning (QFL) module learns the quantum features at the pixel level, as a complementary to GFL by preserving pixel-level detailed spatial information not retained in the superpixels. In the final classification stage, a quantum classifier is designed to cooperate with a traditional fully connected classifier. The superior HCD performance of the proposed QUEEN-empowered GNN (i.e., QUEEN-G) will be experimentally demonstrated on real hyperspectral datasets.

Figures (12)

• Figure 1: Overall network framework of the proposed QUEEN-$\mathcal{G}$. The bitemporal HSIs are first dimension-reduced by a $1 \times 1$ convolutional layer. Subsequently, the graph feature learning (GFL) and quantum feature learning (QFL) modules, further detailed in Figs. \ref{['fig:GFL']} and \ref{['fig:QFL']}, are adopted to learn the graph-structured and quantum unitary-computing features at the superpixel level and pixel level, respectively. The extracted quantum feature and graph feature are then fused, and weighted by the spectral similarity matrix ${\bm Z}$ (obtained from ${\bm X}_1$ and ${\bm X}_2$). Then, the extracted feature map, the dimension-reduced input HSIs, and the spectral similarity map are further fused by a feature fusion module. Finally, the quantum-enhanced classifier (QEC), detailed in Fig. \ref{['fig:QEC']}, is employed to detect the changed areas. With the additional quantum branch, new feature information yields upgraded decision-making in the final classification/detection.
• Figure 2: Detailed architecture of the GFL module, which is outlined in Fig. \ref{['fig:architecture']}. The GFL module begins by converting the pixel-level feature map into a graph using a graph encoder. It then extracts graph-structured features at the superpixel level through a multi-head GAT layer followed by a single-head GAT layer. Finally, the feature graph is transformed back to the pixel level via a graph decoder. The block at the left describes the generation of the adjacency matrix $\textbf{A}$ and the association matrix ${\textbf{O}}$. The block in the middle details the process to compute the attention coefficient $\rho_{i j}$ in the GAT layers. The block at the right illustrates the information propagation of the nodes in the multi-head GAT layer.
• Figure 3: Detailed architectures of the QFL module, and of the QUEEN (together with its input dimension). The block at the top left is the QFL module, which is composed of a $1 \times 1$ convolutional layer, a QUEEN with FE to extract quantum unitary-computing features at the pixel level, and a $1 \times 1$ convolutional layer. The block at the bottom is the QUEEN with mathematically provable FE that can generate any possible quantum state or implement any valid quantum function. The block at the top right illustrates how to input the feature map into the QUEEN for learning the quantum information.
• Figure 4: Detailed architecture of the QEC module with the input dimension of the QUEEN. QEC first learns the information separately through a quantum network and a traditional fully connected layer. Then, a trainable weight ${\bm W}$ highlights the important elements, followed by another fully connected layer to fuse the quantum and the traditional features. Finally, the output feature map ${\bm M}$ is computed by the Softmax function. The block at the right illustrates how to input the feature map into the QUEEN to learn the quantum information.
• Figure 5: Pseudo-color images of the benchmark Yancheng dataset. (a) HSI acquired on May 3rd, 2006. (b) HSI acquired on April 23rd, 2007. (c) Ground-truth change map, where white pixels denote the changes.

Theorems & Definitions (1)

• Theorem 1