Detection of the New Class of Hypersonic Targets under Emerging Hyperspectral Sample Streams: An Unsupervised Isolation Forest Solution

Abstract

Rapid detection of the new class of hypersonic targets (HTs) presenting unknown military threats in space-based surveillance will guarantee aerospace security. This paper proposes an unsupervised subclass definition and an efficient isolation forest based on an anomalous hyperspectral feature selection (USD-EiForest) algorithm to detect the new class of never-before-seen HTs under emerging hyperspectral sample streams. First, we reveal that the hyperspectral features (HFs) of the new class of HTs have no anomaly characteristics when compared to the globally observed samples while having prominent anomaly characteristics when compared to the subclasses of observed samples. Second, an unsupervised subclass definition method adapted to HTs is utilized to classify the observed samples into several subclasses. Then, an efficient isolation forest is designed to determine whether the data stream sample in each subclass indicates anomaly features that mark the detection of the new class of hypersonic targets (DNHT). Finally, we experiment on the simulated hyperspectral HTs data sets considering the RAM-C II HT as the observed samples and the HTV-2 HT as the unknown samples. The results suggest that the performance of our proposal has competitive advantages in terms of accuracy and detection efficiency.