PaperlayerView exact revision on arXiv

Source-linked AI summary

Target detection in synthetic aperture radar imagery: a state-of-the-art survey

Khalid El-Darymli, Peter McGuire, Desmond Power, Cecilia Moloney

arXiv:1804.04719v1eess.IV

TL;DR

Target detection is the front-end of SAR-ATR, where computational complexity, detection efficacy, false alarms, and detection probability must be balanced. This survey taxonomizes SAR target-detection methods, reviews their assumptions and examples, and analyzes CFAR through signal-processing and pattern-recognition perspectives, showing CFAR’s filtering and classification interpretations.

  • Problem

    SAR-ATR detectors must identify regions of interest while balancing computational simplicity, low probability of false alarm, and high probability of detection.

  • Method

    The paper surveys and taxonomizes target-detection methods, overviews their implementation assumptions and representative examples, and develops signal-processing and pattern-recognition analyses of CFAR.

  • Results

    CFAR is characterized as a finite impulse response band-pass filter and a suboptimal one-class classifier; one-parameter CFAR is a Euclidean distance classifier, while two-parameter CFAR is a quadratic discriminant with a missing term.

  • Takeaways & Limitations

    The survey provides a framework for comparing SAR target detectors and supports objective analysis of CFAR design and implementation.

  • Takeaways & Limitations

    CA-CFAR assumes isolated targets and homogeneous, identically distributed reference clutter, and its performance suffers significant detection loss when these assumptions do not hold.

Abstract

from arXiv · show

Target detection is the front-end stage in any automatic target recognition system for synthetic aperture radar (SAR) imagery (SAR-ATR). The efficacy of the detector directly impacts the succeeding stages in the SAR-ATR processing chain. There are numerous methods reported in the literature for implementing the detector. We offer an umbrella under which the various research activities in the field are broadly probed and taxonomized. First, a taxonomy for the various detection methods is proposed. Second, the underlying assumptions for different implementation strategies are overviewed. Third, a tabular comparison between careful selections of representative examples is introduced. Finally, a novel discussion is presented, wherein the issues covered include suitability of SAR data models, understanding the multiplicative SAR data models, and two unique perspectives on constant false alarm rate (CFAR) detection: signal processing and pattern recognition. From a signal processing perspective, CFAR is shown to be a finite impulse response band-pass filter. From a statistical pattern recognition perspective, CFAR is shown to be a suboptimal one-class classifier: a Euclidian distance classifier and a quadratic discriminant with a missing term for one-parameter and two-parameter CFAR, respectively. We make a contribution toward enabling an objective design and implementation for target detection in SAR imagery.

REVIEW

The paper is a state-of-the-art survey of SAR target detection, motivated by the need to organize diverse research on this topic.

  • The survey addresses target detection methods in synthetic aperture radar imagery.
  • It covers single-feature-based methods, constant false alarm rate detection, multifeature-based methods, and expert-system-oriented methods.

1 Introduction

SAR-ATR divides processing into detector, low-level classifier, and high-level classifier stages to manage the computational burden of high-resolution SAR imagery. The survey focuses on organizing detection research, primarily for single-channel SAR imagery.

  • SAR-ATR commonly splits processing into detector, low-level classifier, and high-level classifier stages.The detector is also called the prescreener, while the low-level classifier is also called the discriminator.
  • The detector identifies regions of interest for later analysis while balancing computational complexity, detection probability, and false-alarm probability.It should operate in real time or near real time while maintaining low PFA and high PD.
  • The survey responds to the difficulty of relating numerous detection strategies and their findings across the literature.Its stated goal is to provide an umbrella for broadly probing and taxonomizing this research.
  • The survey restricts its attention to single-channel, or single-polarization, SAR imagery, although many topics may extend to multichannel data.
  • The paper introduces a taxonomy, surveys implementation strategies, compares representative methods, and discusses additional detection issues.The discussion includes CFAR window size, CFAR loss, one-class classification, and CA-CFAR as a comparison baseline.

2 Taxonomy of the Detection Methods

Detection algorithms are organized into single-feature-based, multifeature-based, and expert-system-oriented taxa. Increasing sophistication introduces a complexity-performance tradeoff that must be balanced when selecting an approach.

  • The survey evaluates detection modules by computational complexity, probability of detection, and false-alarm rate.The module should reduce clutter false alarms and pass regions of interest to later processing stages.
  • Single-feature-based: Single-feature-based methods use one image feature, typically pixel-intensity brightness or radar cross-section.This taxon is described as the most common and as a building block for the other taxa.
  • Multifeature-based: Multifeature-based methods fuse two or more extracted features, such as radar cross-section, multiresolution radar cross-section, and fractal dimension.The survey describes this taxon as expected to provide improved detection performance with fewer false alarms.
  • Expert-system-oriented: Expert-system-oriented methods use multistage artificial intelligence and prior knowledge about the scene, clutter, or targets.Prior knowledge can come from image segmentation, scene maps, or previously gathered data.
  • Greater detection sophistication creates a complexity-performance tradeoff that requires careful approach selection.

3 Taxa, Methods, and Selected Examples

The paper organizes SAR target-detection methods into three taxa and surveys their methods, assumptions, and representative examples. It emphasizes CFAR variants, their operating conditions, and limitations across clutter and target scenarios.

  • Taxonomy: The literature is broadly classified into single-feature-based, multifeature-based, and expert-system-oriented detection taxa.These taxa and their sub-taxa are organized in the paper’s taxonomy.
  • CFAR-based methods: CFAR methods combine parametric or nonparametric estimation with implementation strategies such as CA-CFAR, SOCA-CFAR, GOCA-CFAR, and OS-CFAR.The implementation strategies differ in how they estimate statistics from the boundary ring and apply thresholds.
  • Single-feature-based methods: Single-feature-based methods commonly use RCS alone, with CFAR as the dominant approach and CA-CFAR as the baseline.The Bayesian detector is presented as theoretically optimal, while practical CFAR methods operate on features from a sliding-window stencil.
  • CFAR limitations: CA-CFAR assumes isolated targets and homogeneous, identically distributed reference pixels; violating these assumptions causes significant detection loss.The reference pixels are assumed to have a probability density similar to that of the pixel under test.
  • CFAR limitations: SOCA-CFAR handles strong clutter returns but is susceptible to clutter edges, whereas GOCA-CFAR performs better at clutter edges but degrades with strong boundary-ring returns.Both methods incur additional CFAR loss relative to CA-CFAR because they use only part of the boundary ring.
  • Advanced and adaptive methods: OS-CFAR is reported to outperform CA-CFAR in heterogeneous clutter and for contiguous targets, but its performance degrades during clutter transitions.Expert-system-oriented CFAR mixtures and adaptive reference-window methods address specialized clutter conditions, although SAR-specific work remains limited.

4 Comparison

Table 3 compares selected detection modules across SAR image types, features, clutter and target types, and applicable clutter or target models. The comparison emphasizes methodological differences and scenario applicability rather than reported detection performance.

  • Table 3 compares representative detection modules across image type, features, clutter and target type, and clutter or target models.Examples are selected to cover different methods under each taxon.
  • The comparison is intended to show major methodological differences and applicability to particular scenarios, not to rank algorithms by reported PD or PFA.The authors note that sensor characteristics, operating conditions, and data used differ across examples.
  • Nonrectangle-shaped methods replace the conventional hollow stencil with alternative structures, including gamma kernels.Gamma-CFAR uses two-dimensional gamma kernels in a sliding stencil and estimates thresholds from radial pixel intensities around the PUT.
  • Multifeature-based methods combine features such as RCS, variance, extended fractals, or multiresolution RCS for detection.Reported examples include fusion of three features and sliding-window CFAR within a multistage feature-generation process.
  • Representative applications span synthetic, emulated, and real SAR imagery involving land vehicles, ships, and heterogeneous or homogeneous clutter.The examples include MSTAR, TESAR, Radarsat-1, synthetic X-band SAR, and real SAR imagery.

5 Discussion

The discussion examines the suitability of SAR clutter models and reinterprets CFAR from signal-processing and pattern-recognition perspectives. It characterizes CFAR as a band-pass FIR filter and as a suboptimal one-class classifier with distinct interpretations for one- and two-parameter variants.

  • Suitability of SAR data models: Parametric CFAR detection requires an appropriate probability distribution because the detector’s thresholding operation depends on the clutter distribution.The discussion emphasizes model suitability before applying parametric detection methods.
  • Suitability of SAR data models: Goodness-of-fit validation should precede distribution selection, with CVM measuring distance between design and empirical CDFs and Anderson-Darling giving more weight to distribution tails.The discussion notes that SAR dependence violates the independence assumptions underlying the commonly used Kolmogorov-Smirnov test.
  • Understanding the Multiplicative SAR Data Models: SAR multiplicative models represent the observed image as the product of independent backscatter and speckle random processes.The backscatter variable is often positive real, while speckle may be complex or positive real depending on the image domain.
  • Understanding the Multiplicative SAR Data Models: The square root, reciprocal square root, and constant cases of the gamma-based model yield the K-distribution, Go-distribution, and scaled-speckle homogeneous case, respectively.The Go-distribution can model extremely heterogeneous regions that the K-distribution cannot model with the same two parameters.
  • Signal processing perspective: A CFAR stencil is a band-pass FIR filter implementable using two low-pass filters centered on the guard-region target area and the whole stencil.The spatial-domain filter is referred to as a convolution kernel, or kernel.
  • Pattern recognition perspective: CFAR is a suboptimal one-class classifier; one-parameter CFAR becomes a Euclidean distance classifier, while two-parameter CFAR is a quadratic discriminant with a missing term.The missing term in two-parameter CFAR contributes to performance degradation under the stated optimal-classification assumptions.

6 Conclusions

The survey taxonomizes SAR target-detection methods, reviews their implementation strategies, and compares representative examples. It also develops signal-processing and pattern-recognition interpretations of CFAR, while explicitly limiting its coverage to selected major approaches rather than the entire literature.

  • Survey scope and taxonomy: The detector methods are organized into single-feature-based, multifeature-based, and expert-system-oriented taxa, with further classification of implementation strategies.Representative examples and a comparison table are provided for methods under each taxon.
  • Survey scope and taxonomy: The survey emphasizes Bayesian detection as optimal, CFAR as the most popular approach, and expert-system-oriented methods' advantages.
  • CFAR interpretations and data models: Beta-prime and Go distributions are reported as better background-clutter models than the K-distribution for single-look and multilook SAR imagery, respectively.
  • CFAR interpretations and data models: From a signal-processing perspective, CFAR is interpreted as a finite impulse response band-pass filter implementable with two low-pass filters in spatial or frequency domains.
  • CFAR interpretations and data models: From a statistical pattern-recognition perspective, CFAR is a suboptimal one-class classifier: Euclidean distance for one-parameter CFAR and a quadratic discriminant with a missing term for two-parameter CFAR.
  • Scope boundary: The survey selects representative popular methods rather than exhaustively covering all detector implementations, and a separate survey of intermediate and back-end SAR-ATR stages was still under preparation.
Loading 1804.04719v1…