LDA-MIG Detectors for Maritime Targets in Nonhomogeneous Sea Clutter

Xiaoqiang Hua, Linyu Peng, Weijian Liu, Yongqiang Cheng, Hongqiang Wang, Huafei Sun, Zhenghua WangIEEE TGRS 2023 · DOI · arXiv

TL;DR

This paper introduces a novel class of radar detectors that combine Linear Discriminant Analysis (LDA) with Matrix Information Geometry (MIG) to find maritime targets in messy, unpredictable sea clutter. By learning a specific mathematical projection that "squashes" high-dimensional data onto a more discriminative manifold, the authors achieve significantly better detection performance than traditional methods, even when clean training data is scarce.

What Problem Does This Paper Address?

Detecting a small boat or a buoy in the middle of a choppy ocean is a nightmare for radar systems. The "background noise" (sea clutter) is nonhomogeneous, meaning its statistical properties change rapidly across space and time.

Traditional radar detectors (like the Adaptive Matched Filter) rely on having plenty of "secondary data"—samples of pure clutter nearby—to estimate the background noise. In a nonhomogeneous environment, you simply don't have enough clean, consistent samples. Standard statistical tools like the Sample Covariance Matrix (SCM) break down, leading to false alarms or missed targets. While Matrix Information Geometry (MIG) has emerged as a robust alternative by treating radar signals as points on a curved mathematical surface (a manifold), choosing the "right" geometric distance to measure these points remains a challenge.

Methods

The authors propose LDA-MIG detectors. Their approach follows a sophisticated pipeline:

  1. Matrix Representation: Radar echoes are modeled as Hermitian Positive-Definite (HPD) matrices. These matrices capture the correlation and power of the signal.
  2. LDA Manifold Projection: This is the core innovation. Inspired by classical Linear Discriminant Analysis, the authors learn a mapping ($f_W(R) = W^H R W$) that projects high-dimensional HPD matrices into a lower-dimensional space.
  3. Optimization on the Stiefel Manifold: The projection matrix $W$ is calculated by solving an optimization problem: maximize the distance between "target" and "clutter" classes while minimizing the spread within each class. This optimization is performed on a Stiefel manifold using Riemannian gradient descent.
  4. Geometric Measures: The authors developed four distinct detectors based on different geometric "rulers":
  5. AIRM (Affine Invariance Riemannian Metric)
    • LEM (Log-Euclidean Metric)
    • JBLD (Jensen-Bregman LogDet divergence)
    • SKLD (Symmetrized Kullback-Leibler divergence)
  6. Robust Estimation: They use geometric means (rather than simple arithmetic averages) to estimate the clutter background, proving via "influence functions" that these means are far more resistant to outliers and interference.

Datasets

The researchers validated their work using two primary sources:

  • Simulated Radar Clutter: Generated using complex Gaussian distributions with added interference and varying signal-to-clutter ratios (SCR) to test robustness.
  • Real-world IPIX Radar Data: Data collected by McMaster University at the shore of Lake Ontario. This is a standard benchmark in the radar community for testing detection in real sea environments.

Key Insights and Ideas

  • Dimensionality as a Feature, Not a Bug: Reducing the dimension of the HPD matrices ($N \times N$ down to $M \times M$) isn't just about saving computation. If the projection is learned via LDA, it actually filters out redundant noise and "sharpens" the contrast between a target and the waves.
  • Geometric Robustness: The paper mathematically demonstrates that the JBLD mean is the most robust against "spiky" outliers in sea clutter, followed by AIRM and LEM. The traditional SCM is shown to be highly sensitive and prone to error.
  • Manifold Awareness: By performing optimization directly on the manifold (the Stiefel manifold), the algorithm respects the intrinsic geometry of the data, leading to more stable and accurate projection matrices than Euclidean methods.

Key Results

  • Superior Detection: In both simulations and real IPIX data, the LDA-MIG detectors consistently outperformed standard MIG detectors and classical benchmarks (like AMF and ACE).
  • Performance at Low SCR: The proposed detectors maintained a high probability of detection ($P_d$) even when the target signal was nearly buried in the clutter (low Signal-to-Clutter Ratio).
  • The "Aha" Moment in Dimension: Interestingly, for several metrics (AIRM, LEM, SLKD), the best performance occurred when the data was projected down to a very low dimension ($M=1$), suggesting that the most discriminative information in sea clutter is concentrated in a very specific subspace.

Limitations and Future Work

  • Supervised Learning Requirement: The LDA projection requires labeled training data (samples known to contain targets vs. samples known to be pure clutter). In a real-time, "blind" tactical environment, obtaining these labels can be difficult.
  • Training Set Selection: The authors note that detection performance can vary based on the quality of the training dataset. Future work is needed to automate the selection of the most "efficient" training samples.
  • Structural Constraints: The current study focuses on diagonal loading for matrix construction. The authors suggest exploring other structures, such as persymmetric or shrinkage estimators, to further refine the detectors.

Summary updated

Abstract

This article deals with the problem of detecting maritime targets embedded in nonhomogeneous sea clutter, where the limited number of secondary data is available due to the heterogeneity of sea clutter. A class of linear discriminant analysis (LDA)-based matrix information geometry (MIG) detectors is proposed in the supervised scenario. As customary, Hermitian positive-definite (HPD) matrices are used to model the observational sample data, and the clutter covariance matrix of the received dataset is estimated as the geometric mean of the secondary HPD matrices. Given a set of training HPD matrices with class labels, which are elements of a higher dimensional HPD matrix manifold, the LDA manifold projection learns a mapping from the higher dimensional HPD matrix manifold to a lower dimensional one subject to maximum discrimination. In this study, the LDA manifold projection, with the cost function maximizing between-class distance while minimizing within-class distance, is formulated as an optimization problem in the Stiefel manifold. Four robust LDA-MIG detectors corresponding to different geometric measures are proposed. Numerical results based on both simulated radar clutter with interferences and real IPIX radar data show the advantage of the proposed LDA-MIG detectors against their counterparts without using LDA and the state-of-the-art maritime target detection methods in nonhomogeneous sea clutter.
BibTeX
@article{hua2023lda,
  author = {Xiaoqiang Hua and Linyu Peng and Weijian Liu and Yongqiang Cheng and Hongqiang Wang and Huafei Sun and Zhenghua Wang},
  title = {{LDA-MIG Detectors for Maritime Targets in Nonhomogeneous Sea Clutter}},
  journal = {IEEE Transactions on Geoscience and Remote Sensing},
  year = {2023},
  doi = {10.1109/tgrs.2023.3250990},
}
Figure