From Unsupervised to Semi-supervised Anomaly Detection Methods for HRRP Targets

TL;DR

This paper benchmarks unsupervised and semi-supervised anomaly detection methods on HRRP radar data, demonstrating how few labeled anomalies can improve detection performance in maritime surveillance contexts.

Contribution

It introduces a comprehensive benchmark of AD methods including novel deep and shallow approaches, and demonstrates the benefits of semi-supervised learning with real radar data.

Findings

Semi-supervised AD improves detection accuracy.

Deep SVDD outperforms traditional methods in this context.

AD methods enhance maritime situational awareness.

Abstract

Responding to the challenge of detecting unusual radar targets in a well identified environment, innovative anomaly and novelty detection methods keep emerging in the literature. This work aims at presenting a benchmark gathering common and recently introduced unsupervised anomaly detection (AD) methods, the results being generated using high-resolution range profiles. A semi-supervised AD (SAD) is considered to demonstrate the added value of having a few labeled anomalies to improve performances. Experiments were conducted with and without pollution of the training set with anomalous samples in order to be as close as possible to real operational contexts. The common AD methods composing our baseline will be One-Class Support Vector Machines (OC-SVM), Isolation Forest (IF), Local Outlier Factor (LOF) and a Convolutional Autoencoder (CAE). The more innovative AD methods put forward by this work are Deep Support Vector Data Description (Deep SVDD) and Random Projection Depth (RPD), belonging respectively to deep and shallow AD. The semi-supervised adaptation of Deep SVDD constitutes our SAD method. HRRP data was generated by a coastal surveillance radar, our results thus suggest that AD can contribute to enhance maritime and coastal situation awareness.