Adaptive non-Zero Mean Gaussian Detection and Application to Hyperspectral Imaging

Joana Frontera–Pons, Frédéric Pascal, Jean‐Philippe OvarlezarXiv 2014 · DOI · arXiv

Abstract

Classical target detection schemes are usually obtained deriving the likelihood ratio under Gaussian hypothesis and replacing the unknown background parameters by their estimates. In most applications, interference signals are assumed to be Gaussian with zero mean or with a known mean vector that can be removed and with unknown covariance matrix. When mean vector is unknown, it has to be jointly estimated with the covariance matrix, as it is the case for instance in hyperspectral imaging. In this paper, the adaptive versions of the classical Matched Filter and the Normalized Matched Filter, as well as two versions of the Kelly detector are first derived and then are analyzed for the case when the mean vector of the background is unknown. More precisely, theoretical closed-form expressions for false-alarm regulation are derived and the Constant False Alarm Rate property is pursued to allow the detector to be independent of nuisance parameters. Finally, the theoretical contribution is validated through simulations and on real hyperspectral scenes.

Introduction

In signal processing, the fundamental challenge is often identifying a specific target signal (the "needle") within a complex background (the "haystack"). Most classical detection algorithms operate under the assumption that the background noise is either zero-mean or has a known average value that can be subtracted. However, in Hyperspectral Imaging (HSI), sensors capture hundreds of narrow spectral bands, and the background is rarely zero-mean. Each material possesses a unique spectral signature that shifts the baseline.

If this mean vector is unknown and not properly accounted for, detectors become biased, leading to excessive false alarms. While previous research has utilized "plug-in" estimates for the mean, this paper provides the rigorous theoretical proof and closed-form expressions required to establish precise detection thresholds that remain robust regardless of shifts in background average or power.

Methodology

The study employs Statistical Detection Theory to derive adaptive versions of three prominent detectors under the assumption of a non-zero mean Gaussian background:

  1. Adaptive Matched Filter (AMF): Designed to maximize the signal-to-noise ratio.
  2. Kelly Detector: Derived from the Generalized Likelihood Ratio Test (GLRT). The authors analyze both a "plug-in" version (estimating the mean from separate data) and a "generalized" version (jointly estimating the mean with the pixel under test).
  3. Adaptive Normalized Matched Filter (ANMF): A robust variation that handles changes in background power (scaling).

A primary technical contribution is the derivation of relationships between the Probability of False Alarm (PFA) and the detection threshold. Utilizing complex Wishart and Beta distributions, the authors demonstrate that these detectors possess the CFAR-matrix property. This means the PFA depends solely on data dimensions ($m$) and the number of samples ($N$), remaining independent of the actual background mean or covariance values.

PFA vs. Threshold for the NMF detector, showing perfect agreement between theoretical formulas and Monte-Carlo simulations for different sample sizes $N$.

Experimental Validation

The theoretical derivations were validated using both simulated and real-world data:

  • Simulated Data: $10^6$ Monte-Carlo trials were conducted using 5-dimensional Gaussian vectors to verify that experimental results align with theoretical PFA curves.
  • Real-World HSI Data: Data from the NASA Hyperion sensor was analyzed, specifically a $798 \times 253$ pixel scene with 116 spectral bands. A homogeneous $60 \times 20$ pixel region of water was used to approximate Gaussian conditions for testing.

Results and Discussion

The Monte-Carlo simulations demonstrated a "perfect agreement" with the derived theoretical formulas, with experimental data points falling exactly on the theoretical curves.

A key insight is the "cost" of estimating the mean vector. For AMF and ANMF, estimating the mean reduces the statistical degrees of freedom by one (effectively $N$ becomes $N-1$). In contrast, the "plug-in" Kelly detector exhibits a more complex PFA expression that does not simplify into standard hypergeometric functions, indicating that different detectors respond uniquely to mean estimation.

Comparison of detection probability ($P_d$) vs. SNR for AMF, ANMF, and Kelly detectors, illustrating the superior performance of the Kelly variants.

When comparing detection performance, the Kelly detectors (both "plug-in" and generalized) consistently provided the highest probability of detection ($P_d$) for a given false alarm rate. Real-world validation on the Hyperion water scene further confirmed that the detectors follow theoretical Gaussian predictions closely.

Conclusion

This work establishes that even with an unknown background mean, it is possible to construct detectors that are immune to background variations. This allows for the setting of stable thresholds across varying image regions or lighting conditions. While the Gaussian assumption holds well in controlled environments, future research will focus on extending these derivations to non-Gaussian distributions (such as SIRV or Elliptical distributions) to better accommodate the heavy-tailed, non-homogeneous backgrounds typical of terrestrial hyperspectral imaging.

Summary updated

Abstract

Classical target detection schemes are usually obtained deriving the likelihood ratio under Gaussian hypothesis and replacing the unknown background parameters by their estimates. In most applications, interference signals are assumed to be Gaussian with zero mean or with a known mean vector that can be removed and with unknown covariance matrix. When mean vector is unknown, it has to be jointly estimated with the covariance matrix, as it is the case for instance in hyperspectral imaging. In this paper, the adaptive versions of the classical Matched Filter and the Normalized Matched Filter, as well as two versions of the Kelly detector are first derived and then are analyzed for the case when the mean vector of the background is unknown. More precisely, theoretical closed-form expressions for false-alarm regulation are derived and the Constant False Alarm Rate property is pursued to allow the detector to be independent of nuisance parameters. Finally, the theoretical contribution is validated through simulations and on real hyperspectral scenes.
BibTeX
@misc{fronterapons2014adaptive,
  author = {Joana Frontera–Pons and Frédéric Pascal and Jean‐Philippe Ovarlez},
  title = {{Adaptive non-Zero Mean Gaussian Detection and Application to Hyperspectral Imaging}},
  year = {2014},
  eprint = {1404.2977},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1404.2977},
}
Figure