S2G-GCN: A Plot Classification Network Integrating Spectrum-to-Graph Modeling and Graph Convolutional Network for Compact HFSWR
Xiaotong Li, Weifeng Sun, Yonggang Ji, Weimin Huang
Abstract
Plot classification refers to the identification of true target plots among initial detections, and it is crucial for target tracking with compact high-frequency surface wave radar (HFSWR) systems. However, due to the limited spatial resolution and low signal to interference plus noise ratio (SINR) inherent in compact HFSWR systems, traditional classification methods often fail to distinguish true targets from false alarms. Targets, clutter, and noise exhibit different morphological and statistical features in the range-Doppler (R-D) spectrum, and their differences in spatial distribution of echo energy can be described by modeling each detected plot and its surrounding cells as a graph. Based on the above consideration, a novel plot classification network integrating spectrum-to-graph modeling and graph convolutional network (S2G-GCN) is proposed. Firstly, the constant false alarm rate detection algorithm is applied to R-D spectra to obtain potential target plots. For each plot, an echo energy diffusion region is built to include several resolution cells around its spectral peak. Then, these cells are modeled as a graph, where each node corresponds to a cell, and edges are defined using the spatial proximity and energy similarity between neighboring nodes. Finally, a graph convolutional network (GCN)-based classifier is employed to learn discriminative features from the constructed graph and classify each detected plot into one of four classes: true target, sea clutter, ground clutter, or noise. Experimental results demonstrate that the proposed S2G-GCN outperforms three baseline methods, achieving a plot classification accuracy of 93.68%.