Three-dimensional Feature Representation via Cross Attention for Space Target Recognition

Currently, Radar Cross Section (RCS), High Resolution Range Profile (HRRP), and Joint Time-Frequency (JTF) image are three kinds of important data domains widely used for space target recognition. The last two are derived from RCS through transformation. Targets with completely different appearances have highly differentiated RCS values. However, with the continuous development of modern ballistic missiles, decoys used to mislead recognition systems have become identical in appearance to warheads. The target characteristics reflected by RCS tend to be similar. At the same time, a large amount of irregular space debris, as abnormal targets, can significantly affect the recognition system. Traditional single-domain classifiers are difficult to achieve high accuracy in this scenario. In this article, we propose a 3D feature representation called action-attribute feature space generated from RCS and united HRRP-JTF features using Cross Attention mechanism (CA). Both feature spaces are mined from the original data through time, frequency, and spatial domain. The dot-product operation is implemented to fuse RCS and united HRRPJTF feature spaces. In addition, a ST-3D-CNN is also combined to further extract spatiotemporal information. As a result, the cross fusion of 3D features brought by CA highly improves the recognition accuracy of a simulated dynamic ballistic target dataset with 35 categories including warhead, decoy, and debris. Experimental results also demonstrate the algorithm's robustness under various noise conditions and different signal-to-noise ratios (SNR).