Radar signal sorting with GraphSAGE on weighted directed graph

Radar signal sorting is a key component of electronic reconnaissance missions. In satellite-based electronic reconnaissance systems, intercepted radar pulses are collected through long-duration and wide-area spectrum monitoring, resulting in high pulse density and significant overlap across multiple signal parameters. Such a complex signal environment often leads to substantial performance degradation in conventional radar signal sorting methods. To address these challenges, this paper proposes a radar signal sorting method that integrates weighted directed graph modeling with an improved Graph Sample and Aggregate (GraphSAGE)-based network. Specifically, radar pulses are modeled as nodes in a weighted directed graph constructed using an adaptive sparsity-aware strategy, where edge weights encode inter-pulse similarity and connection strength. Then, the proposed GraphSAGE-based network with adaptive weighted-attention aggregation is used to enhance neighborhood representation learning for pulse-level node classification. Finally, experimental results show that the proposed method consistently outperforms representative baseline algorithms in terms of classification performance under multiple evaluation metrics. Furthermore, the method exhibits strong robustness under various non-ideal conditions, including pulse loss, spurious pulses, and measurement noise.