Abstract
The Mellin transform offers a robust framework for parameter estimation in complex statistical models. This paper proposes a novel estimator for the K distribution plus noise (KpN) model, which is essential for accurate radar sea clutter characterization. By deriving second characteristic functions and log-based moments via the Mellin transform, we establish a relationship between log-cumulants and KpN parameters (shape, scale, and noise power). Monte Carlo simulations demonstrate that the proposed Mellin-based estimator significantly outperforms existing methods, such as Higher-Order Moments Estimation (HOME) and zlog(z) approaches, particularly in scenarios with low clutter-to-noise ratios (CNR) and limited sample sizes.
Introduction
In maritime radar surveillance and Synthetic Aperture Radar (SAR) imaging, the ability to distinguish targets from background "sea clutter" is paramount. Sea clutter represents the backscattered signal from the ocean surface, which can obscure small or low-reflectivity targets. While the Compound K distribution has long been the standard for modeling high-resolution radar clutter, it often fails to account for the thermal noise inherent in radar hardware.
The K plus noise (KpN) model provides a more realistic representation by incorporating this additive noise. However, estimating the three parameters of the KpN model—shape ($v$), scale ($b$), and noise power ($p_n$)—presents a significant mathematical challenge. Traditional estimation techniques, including the HOME method and the zlog(z) approach, often struggle with accuracy or require prohibitively large datasets to achieve convergence.
Methodology: The Mellin Transform Approach
This research introduces a parameter estimation method leveraging the Mellin transform. Unlike the more common Fourier or Laplace transforms, the Mellin transform is uniquely suited for distributions involving products of random variables, making it ideal for the multiplicative texture/speckle nature of sea clutter.
The proposed approach follows a three-step process:
- Derivation of Log-Cumulants: Instead of relying on raw moments, which are susceptible to instability in heavy-tailed distributions, the authors use the Mellin transform to derive the "characteristic function of the second kind." This allows for the calculation of log-cumulants—statistics derived from the logarithm of the radar data.
- Analytical Modeling: These log-cumulants are expressed through complex mathematical functions, specifically the Tricomi function and Polygamma functions, which capture the interplay between the K-distribution and additive noise.
- Numerical Optimization: A system of three equations based on the first three log-cumulants is solved numerically to extract the shape, scale, and noise parameters simultaneously.
Experimental Evaluation
The efficacy of the Mellin-based estimator was evaluated using extensive Monte Carlo simulations. Synthetic data following the KpN distribution was generated across various Clutter-to-Noise Ratios (CNR), ranging from -3dB (high noise) to 10dB (low noise). The study also examined the impact of sample size, testing the algorithm on datasets as small as $M=300$ and as large as $M=10,000$. The performance was benchmarked against the established HOME method and the zlog(z)-based closed-form method.
Key Results
The proposed estimator demonstrated superior accuracy and robustness across nearly all tested scenarios. Notably, the Mellin-based approach maintained stability even at a CNR of -3dB, where the noise power exceeds the clutter signal—a condition that typically causes other estimators to fail.
As shown in the figure above, the Mean Square Error (MSE) for noise power estimation is significantly lower for the Mellin-based approach (red squares) and the zlog(z) method (blue stars) compared to the HOME method (green circles). The Mellin method consistently provides the most stable and accurate estimates across multiple trials.
While the Mellin-based method requires a numerical solver, resulting in a slightly higher computational cost (approximately 0.89 seconds per trial compared to 0.31 seconds for zlog(z)), the gains in accuracy and reliability, especially in small-sample regimes, justify the trade-off.
Conclusion
The Mellin transform-based estimator provides a powerful new tool for radar signal processing. By operating in the log-domain, the method achieves high precision in parameter estimation for the KpN model, even under challenging low-CNR conditions. Future work will focus on validating these results using real-world sea clutter data, such as the IPIX database, and optimizing the numerical solvers for real-time applications.
Summary updated