Spatio-Temporal Dynamic Graph Attention Network-Based Detector for Sea-Surface Small Targets

This article presents a spatio-temporal dynamic graph attention network (STGAT)-based detector for the detection of sea-surface small targets. It utilizes graphical modeling of radar returns to capture the inherent relationships. With two segmentation operations applied to the received sequence, it is first modeled as a static graph and then arranged chronologically to form a dynamic graph. The topology of a structured graph reflects inherent coherence, while the vertex attributes, which comprise features, facilitate the process of multifeature fusion. The relative maximum degree (RMD) feature are proposed as one of the vertex attributes and shows superiority. Subsequently, the STGAT, a dynamic graph neural network with double convolutions, integrates graph and standard convolution to process dynamic graphs. The network will be trained on simulated datasets and then tested on measured datasets. The outputs of the proposed detector are 2-D vectors belonging to vertices. They are used as classification probabilities for the purpose of distinguishing the target from sea clutter. The decision threshold is obtained through a Monte Carlo experiment at a given false alarm probability. A comparison experiment between the proposed detector and existing detectors is carried out on 80 datasets of IPIX radar database, which indicates the proposed method works effectively in 70% situations.