Space Target Classification Improvement by Generating Micro-Doppler Signatures Considering Incident Angle

Abstract

Classifying space targets from debris is critical for radar resource management as well as rapid response during the mid-course phase of space target flight. Due to advances in deep learning techniques, various approaches have been studied to classify space targets by using micro-Doppler signatures. Previous studies have only used micro-Doppler signatures such as spectrogram and cadence velocity diagram (CVD), but in this paper, we propose a method to generate micro-Doppler signatures taking into account the relative incident angle that a radar can obtain during the target tracking process. The AlexNet and ResNet-18 networks, which are representative convolutional neural network architectures, are transfer-learned using two types of datasets constructed using the proposed and conventional signatures to classify six classes of space targets and a debris–cone, rounded cone, cone with empennages, cylinder, curved plate, and square plate. Among the proposed signatures, the spectrogram had lower classification accuracy than the conventional spectrogram, but the classification accuracy increased from 88.97% to 92.11% for CVD. Furthermore, when recalculated not with six classes but simply with only two classes of precessing space targets and tumbling debris, the proposed spectrogram and CVD show the classification accuracy of over 99.82% for both AlexNet and ResNet-18. Specially, for two classes, CVD provided results with higher accuracy than the spectrogram.

Recognition of Ballistic Targets by Fusing Micro-Motion Features with Networks

“…Compared with manual feature extraction, feature extraction methods based on neural networks can automatically extract significant features of the targets and avoid the impact of human errors on the results, which has become the main manner of feature extraction in micro-motion target recognition methods. Current micro-motion feature extraction methods based on neural networks are mostly performed by CNN [19][20][21][22][23][24][25][26]30,31]. Wang et al performed feature extraction on the TR map based on the designed CNN [20], and Kim et al used Googlenet to perform feature extraction on the CVD map of the targets [19].…”

“…However, this kind of fusion method loses a large amount of detailed information. Compared with decision-level fusion, feature-level fusion retains the details of different forms of data and has a stronger expressive ability [24][25][26]. The feature maps obtained by feeding the CVD map and TF spectrogram are spliced along the same dimension to achieve the classification of spatial targets [24,25].…”

“…Compared with decision-level fusion, feature-level fusion retains the details of different forms of data and has a stronger expressive ability [24][25][26]. The feature maps obtained by feeding the CVD map and TF spectrogram are spliced along the same dimension to achieve the classification of spatial targets [24,25]. By splicing the targets' macro-motion and micro-motion features after weighting, Choi et al effectively identified targets and decoys [26].…”