Fact-checked by Grok 2 weeks ago

K-distribution

The K-distribution is a continuous that models the statistics of returns from textured surfaces, such as clutter, by combining a gamma-distributed texture component representing local reflectivity variations with an speckle component arising from coherent . Introduced in by E. Jakeman and P. N. Pusey to explain non-Rayleigh in microwave , it captures the spiky, heavy-tailed nature of such signals, which deviate from Gaussian assumptions in traditional models. The (PDF) for the intensity z in a single-look K-distributed clutter is given by the compound form
P(z) = \int_0^\infty P(z|x) P_c(x) \, dx,
where P(z|x) = \frac{1}{x} \exp\left(-\frac{z}{x}\right) is the conditional exponential PDF for speckle given local power x, and P_c(x) = \frac{b^\nu}{\Gamma(\nu)} x^{\nu-1} \exp(-b x) is the gamma PDF for x, with \nu > 0 controlling texture roughness (lower \nu yields spikier distributions) and b = \nu / p_c related to clutter power p_c = \mathbb{E}. This integral lacks a simple closed form but evaluates to
P(z) = \frac{2 b^{\nu/2} z^{(\nu-1)/2}}{\Gamma(\nu) } K_{\nu-1}(2 \sqrt{b z}),
involving the modified of the second kind K_{\nu-1}(\cdot). Key properties include infinite higher-order moments for \nu \leq k (where k is the moment order), a shape parameter \nu that quantifies deviation from (as \nu \to \infty, it approaches ), and versatility in extensions like the multivariate or polarimetric K-distribution for correlated channels in () imagery. Applications span radar performance analysis, including () detection in sea clutter, where (e.g., Gauss-Laguerre ) computes detection probabilities, and simulation of nonhomogeneous environments for target discrimination. The model has been validated extensively against X-band and other measurements, influencing , , and even adaptations in for heavy-tailed intensity data.

Definition and Parameters

Probability Density Function

The probability density function of the K-distribution for a positive random variable x is given by f(x; \nu, b) = \frac{2 b^{\nu/2} x^{(\nu-1)/2}}{\Gamma(\nu)} K_{\nu-1}\left(2\sqrt{b x}\right), \quad x > 0, where \nu > 0 is the shape parameter, b > 0 is the scale parameter, \Gamma(\cdot) is the , and K_{\nu-1}(\cdot) is the modified Bessel function of the second kind of order \nu - 1. This form depends on \nu, which controls the degrees of freedom in the underlying compound model, and b, which scales the distribution along the positive real line. The support is the , and the normalization ensures \int_0^\infty f(x; \nu, b) \, dx = 1. This PDF arises from a compound model in which the x (representing intensity) conditionally follows an with \tau, modulated by a gamma-distributed texture \tau with shape \nu and rate b. Specifically, the conditional density is f(x \mid \tau) = \frac{1}{\tau} \exp\left(-\frac{x}{\tau}\right) for x > 0, and the marginal texture density is f(\tau) = \frac{b^\nu \tau^{\nu-1} \exp(-b \tau)}{\Gamma(\nu)} for \tau > 0. The unconditional PDF is then f(x) = \int_0^\infty f(x \mid \tau) f(\tau) \, d\tau, which evaluates to the involving the modified via its known integral representation. For large x, the asymptotic behavior of the PDF is governed by the large-argument approximation of the modified , K_{\nu-1}(z) \sim \sqrt{\frac{\pi}{2z}} \exp(-z) as z \to \infty, leading to modulated by a power-law prefactor: f(x; \nu, b) \sim \frac{2 b^{\nu/2} x^{(\nu-1)/2}}{\Gamma(\nu)} \sqrt{\frac{\pi}{4 \sqrt{b x}}} \exp\left(-2 \sqrt{b x}\right). This sub-exponential tail reflects the heavy-tailed nature suitable for modeling spiky phenomena. The \nu influences tail heaviness, with smaller \nu yielding heavier tails for applications in modeling heavy-tailed data such as clutter.

Shape and Scale Parameters

The K-distribution is characterized by two primary parameters: the \nu > 0 and the b > 0. The \nu governs the and the heaviness of the tails in the distribution; smaller values of \nu produce heavier tails and greater spikiness, indicative of increased variability in radar cross-section or , while larger values reduce this , with the distribution approaching an form as \nu \to \infty. The b scales the overall intensity and variance of the , directly relating to the mean power p_c through the relation b = \nu / p_c, thereby controlling the average level of the observed signal. Parameter estimation for the K-distribution commonly employs the method of moments, which leverages the sample and variance to derive estimates of \nu and b, often via ratios such as the second normalized by the squared first moment; this approach is straightforward but can suffer from high variability in higher moments, particularly for small sample sizes. Alternatively, (MLE) maximizes the likelihood function with respect to \nu and b, requiring a two-dimensional numerical search; however, it presents computational challenges due to the involvement of modified , and approximations may yield biased or negative estimates in low-signal-to-noise scenarios. To address , especially for large \nu, reparameterization is often used, such as substituting t = 1/\nu to model the inverse variance of the texture component or expressing the in terms of the \mu instead of b. Sensitivity to the shape parameter \nu is pronounced in higher-order statistics: reductions in \nu elevate both kurtosis and skewness, amplifying the presence of extreme values and asymmetry due to heavier tails, whereas increases in \nu diminish these measures, yielding a distribution closer to Gaussian characteristics with lighter tails. This parameter's influence underscores its role in capturing non-Gaussian behaviors, such as those observed in radar clutter texture modeling.

Statistical Properties

Moments

The K-distribution, commonly used to model clutter intensity, exhibits moments that reflect its nature, leading to heavier tails and higher variability compared to Gaussian distributions. The raw moments are derived from its representation as a gamma-distributed modulating an speckle component. The first raw moment, or , is given by \mathbb{E}[X] = \mu, where \mu > 0 is the average intensity level. The second raw moment is \mathbb{E}[X^2] = \mu^2 \frac{\Gamma(\nu + 2)}{\Gamma(\nu)} \left( \frac{\mu}{\nu} \right)^2 \cdot 2! / \mu^2, simplifying to \mathbb{E}[X^2] = 2\mu^2 \frac{\nu + 1}{\nu}, with \nu > 0 the controlling clustering or fluctuation. The central second moment, or variance, follows as \mathrm{Var}(X) = \mathbb{E}[X^2] - (\mathbb{E}[X])^2 = \mu^2 \left(1 + \frac{2}{\nu}\right). This variance exceeds that of an (\mu^2) by the amount \mu^2 \frac{2}{\nu}, highlighting the additional variability from the . Higher-order raw moments for integer k \geq 1 are \mathbb{E}[X^k] = k! \left( \frac{\mu}{\nu} \right)^k \frac{\Gamma(\nu + k)}{\Gamma(\nu)}. These moments grow factorially with k, modulated by the ratio, which for large \nu approaches unity but for small \nu amplifies tail heaviness. The is \sqrt{1 + \frac{2}{\nu}}, decreasing to 1 as \nu \to \infty, indicating reduced relative . The excess , \kappa = \frac{\mathbb{E}[(X - \mu)^4]}{\mathrm{Var}(X)^2} - 3, is positive and greater than 6 for finite \nu, confirming leptokurtic behavior with peakedness and heavy tails relative to the Gaussian (\kappa = 0). Computations use the fourth raw moment \mathbb{E}[X^4] = 24 \left( \frac{\mu}{\nu} \right)^4 \frac{\Gamma(\nu + 4)}{\Gamma(\nu)}. In the limiting case as \nu \to \infty, the texture variance vanishes, and the moments converge to those of a scaled : \mathbb{E}[X^k] \to k! \mu^k, with variance \mu^2, 1, and excess 6. This transition underscores the K-distribution's in bridging Gaussian-like speckle to more spiky clutter scenarios at low \nu.

Mode and Median

The of the K-distribution, which represents the value at which the (PDF) attains its maximum, is determined by solving the equation f'(x) = 0, where f(x) is the PDF. This leads to a involving the derivative of the modified of the second kind, requiring numerical methods such as Newton-Raphson for solution. For large shape parameter \nu, the approaches 0, reflecting the distribution's convergence to an . The of the K-distribution lacks a and is typically computed numerically via the inverse (CDF) or by solving \int_0^m f(x) \, dx = 0.5 using methods like Gauss-Laguerre . For large \nu, the approaches \mu \ln 2 \approx 0.693 \mu. In the K-distribution, the ordering of measures highlights its positive , particularly for \nu < 2, where the exceeds the , which in turn exceeds the ( > > ). This skew arises from the heavy-tailed nature due to the gamma component. For varying \nu, the shifts rightward with increasing \nu; low \nu (e.g., \nu \approx 0.1) yields a near zero with a long tail, while higher \nu (e.g., \nu > 10) centers the closer to the , maintaining throughout despite apparent risks in highly heterogeneous clutter scenarios.

Related Distributions and Derivations

Compound Gamma Representation

The K-distribution arises as a compound through a representation that models the intensity X as the product X = Y \cdot S, where Y and S are independent random variables, with Y \sim \mathrm{Gamma}(\nu, b) representing the intensity texture (a slowly varying random controlling power fluctuations) and S following an with mean 1 (equivalently, a normalized with 2 , representing the speckle from coherent interference of a Gaussian signal). This formulation captures the compounding of a gamma-distributed mean level with normalized exponential noise, where the exponential arises from the modulus squared of a zero-mean Gaussian with unit variance. To derive the marginal probability density function (PDF) of X, one integrates the conditional PDF of X given Y = y (which is the scaled exponential PDF) over the gamma prior on Y. Specifically, the conditional PDF is f_{X|Y}(x|y) = \frac{1}{y} \exp\left(-\frac{x}{y}\right), and the marginal PDF is then f_X(x) = \int_0^\infty f_{X|Y}(x|y) f_Y(y) \, dy, with f_Y(y) = \frac{b^\nu}{\Gamma(\nu)} y^{\nu-1} \exp(-b y). This integral evaluates to a form involving the modified Bessel function of the second kind, yielding the characteristic K-distribution PDF. This compound representation interprets the K-distribution as modeling systems with a fluctuating intensity, such as in coherent detection where the texture Y accounts for underlying variations in target reflectivity or , while the speckle S arises from the random and of scattered . The uniqueness of this gamma-compounding mechanism lies in its precise production of the K-form via the Bessel integral, which distinguishes it from other gamma mixture models (e.g., those with inverse gamma textures leading to Student's t-distributions) by generating heavier tails suited to spiky clutter statistics.

Relation to Bessel Functions

The (PDF) of the K-distribution incorporates the modified of the second kind, K_{\nu-1}(z), where \nu > 0 is the and z is a scaled or variable. This function arises naturally from the compound gamma representation of the distribution, reflecting the modulation of an exponential speckle component by a gamma-distributed . A key representation that connects this to the underlying compounding process is K_{\alpha}(z) = \frac{1}{2} \left( \frac{z}{2} \right)^{\alpha} \int_0^\infty t^{-\alpha-1} \exp\left(-t - \frac{z^2}{4t}\right) \, dt, valid for \operatorname{Re}(\alpha) > -1/2 and \operatorname{Re}(z) > 0, which parallels the integral form of the gamma mixture in the K-distribution's derivation. Asymptotic expansions of K_{\nu}(z) provide insight into the behavior of the K-distribution's PDF tails. For small arguments (z \to 0^+, \nu > 0), K_{\nu}(z) \sim \frac{1}{2} \Gamma(\nu) \left( \frac{z}{2} \right)^{-\nu}, leading to a finite non-zero value for the PDF as z \to 0^+, consistent with the behavior of exponentially distributed speckle modulated by texture. For large arguments (z \to \infty), K_{\nu}(z) \sim \sqrt{\frac{\pi}{2z}} \, e^{-z} \left(1 + O\left(\frac{1}{z}\right)\right), which induces sub-exponential tails in the PDF, modeling the heavy-tailed nature of radar returns from distributed targets. These expansions influence parameter estimation and tail probability computations in applications. The order of the in the K-distribution PDF is \nu - 1, where \nu governs the in the gamma texture. For integer orders (corresponding to integer \nu), K_{\nu-1}(z) admits recursive relations and finite series expansions, simplifying analytical approximations and moment calculations. Non-integer orders, common for fractional \nu to fit empirical data, require more general hypergeometric representations, increasing computational demands but allowing flexible modeling of clutter variability. Numerical evaluation of K_{\nu-1}(z) in software implementations of the K-distribution relies on series expansions for small z (using the integral or form) and asymptotic or methods for large z, with libraries such as those in or providing stable routines to avoid overflow. For non-integer orders, these methods ensure accurate computation across the parameter range, essential for simulation and fitting.

Applications

Radar Clutter Modeling

The K-distribution provides a physical basis for modeling clutter in environments exhibiting non-Rayleigh statistics, such as sea surfaces or vegetated , where returns arise from the coherent summation of signals from multiple scatterers modulated by underlying surface variations. This model represents clutter as the product of a speckle component—modeled by a Rayleigh (or intensity) distribution, capturing rapid fluctuations from small-scale scatterers like capillary —and a texture component following a , which accounts for slower-varying reflectivity due to larger-scale features like breaking or vegetation density. This structure arises from the inherent heterogeneity in natural clutter scenes, where the number of effective scatterers fluctuates, leading to spikier amplitude distributions than predicted by Gaussian assumptions. In fitting the K-distribution to radar data, the shape parameter \nu serves as an estimate of the effective number of independent scatterers within the resolution cell; higher \nu values indicate more Gaussian-like behavior from abundant scatterers, while low \nu (e.g., \nu < 1) characterizes spiky clutter dominated by few dominant reflectors, such as sea spikes in high sea states. Parameter estimation typically employs moment-matching or maximum likelihood methods on or data, enabling adaptation to specific clutter types like or forested areas. This interpretability facilitates across varying geometries and environmental conditions. For target detection in K-distributed clutter, (CFAR) algorithms are adapted to account for the heavy-tailed nature of the distribution, outperforming traditional cell-averaging CFAR (CA-CFAR) designed for by incorporating texture estimates or order-statistic thresholding to maintain stable false alarm rates. These adaptations, such as the K-CFAR or VI-CFAR variants, leverage the compound structure to normalize against local variations, yielding improved detection probabilities in heterogeneous sea clutter scenarios compared to Rayleigh-based methods. Numerical methods, including simulations, are often required for threshold computation due to the lack of closed-form expressions for detection probabilities under the K-model. Empirical studies from the 1980s and 1990s, using real () data from platforms like and ERS-1, validated the K-distribution's superior fit to heterogeneous clutter over simpler exponential or Weibull models, particularly in capturing the extended tails observed in sea and land returns. For instance, analyses of high-resolution X-band radar data demonstrated that the K-distribution more accurately modeled amplitude histograms in vegetated and maritime scenes, with goodness-of-fit metrics like Kolmogorov-Smirnov tests showing reduced error compared to Weibull approximations. These validations underscored the model's utility for non-Gaussian environments, though it requires multi-look processing for stable parameter estimates.

Synthetic Aperture Radar Imaging

In (SAR) imaging, the K-distribution provides a robust for the multiplicative speckle noise that degrades images, arising from the coherent of scattered waves within each resolution cell. Unlike Gaussian models, which assume homogeneous , the K-distribution captures the compound nature of speckle by combining a Rayleigh-distributed speckle component with a gamma-distributed underlying , making it particularly suitable for heterogeneous scenes such as urban areas or vegetated . The shape parameter \nu quantifies the degree of resolution cell heterogeneity: lower values of \nu (e.g., \nu < 10) indicate high texture variability typical of rough surfaces, while higher values approach Gaussian-like behavior in smoother regions. Adaptive filtering techniques leverage the K-distribution to enhance SAR image quality by estimating local parameters for targeted denoising. Variants of the Lee filter, such as the enhanced Lee filter, incorporate K-distribution statistics to adaptively weight contributions based on estimated \nu and scale parameters within a sliding window, preserving edges while suppressing speckle in homogeneous areas. Similarly, Frost filter adaptations use an exponential weighting scheme derived from the K-distribution's moments to iteratively reduce , with local \nu estimates guiding the for better performance in textured regions compared to fixed-parameter approaches. These methods improve visual interpretability by balancing with detail retention, often applied in preprocessing pipelines for SAR data from missions like . The K-distribution's parameters also enable texture analysis for SAR image segmentation, where variations in \nu and texture strength distinguish terrain types based on scattering heterogeneity. Segmentation algorithms, such as those employing Markov random fields or superpixel partitioning, fit K-distribution models to local image patches to classify regions; for instance, forests exhibit low \nu values (indicating spiky, highly variable backscatter) contrasting with calmer sea surfaces showing higher \nu (more uniform ). This approach facilitates automated land cover mapping, with \nu maps highlighting transitions between textured and non-textured areas for applications in environmental monitoring. In terms of performance, K-distribution-based despeckling outperforms Gaussian-assuming methods in metrics like (PSNR), particularly on heterogeneous data. Recent post-2010 advances integrate for efficient K-distribution parameter inference, using convolutional neural networks to estimate \nu and scale from raw patches, enabling denoising in convolutional architectures that surpass traditional maximum likelihood estimators in accuracy and speed. As of 2024, studies have compared K-distribution-based CFAR detectors with approaches for ship detection in spaceborne imagery, showing improved robustness in heavy-tailed sea clutter scenarios.

History and Extensions

Origins in Physics

The K-distribution emerged from foundational studies in wave propagation through turbulent media during the , where models described the cumulative effects of phase perturbations on propagating fields. These models, rooted in the of fluctuating refractive indices, linked intensity fluctuation statistics to modified through Hankel transforms, providing a mathematical framework for non-Gaussian behaviors in scattered waves. A pivotal advancement came in 1976 with E. Jakeman and P. N. Pusey's introduction of the K-distribution as a specific form for modeling statistics in scenarios, particularly non-Rayleigh echoes from turbulent surfaces like the sea. This formulation arose from physical interpretations of coherent imaging in random , where the received results from exponential-distributed speckle—due to random —compounded by a gamma-distributed reflectivity representing large-scale fluctuations in the medium. Jakeman and Pusey further emphasized in 1978 the broader significance of K-distributions across experiments, highlighting their derivation from random sums in fluctuating environments, which naturally yield the modified of the second kind in the probability density. Initially applied to optical and acoustic problems, the K-distribution captured fluctuations in turbulent atmospheres and before its adoption in clutter modeling during the . In these early contexts, it provided a versatile tool for describing how wave degrades in inhomogeneous media, bridging theoretical predictions with empirical observations of non-exponential tails in fluctuation spectra. The "K" derives directly from the modified of the second kind, K_{\nu}(z), which appears in the distribution's normalizing factor and reflects the asymptotic behavior of scattered intensities.

Modern Developments

Introduced in the late , the multivariate K-distribution has become a key extension for modeling correlated, vector-valued data in polarimetric systems, particularly for (SAR) imaging where multiple polarization channels (e.g., , , , ) exhibit spatial and temporal dependencies. This model compounds a gamma-distributed with a complex Wishart-distributed speckle component to account for non-Gaussian clutter statistics, enabling better characterization of heterogeneous scenes like urban or vegetated terrain. Seminal work by Gini and Greco demonstrated estimation techniques for correlated K-distributed clutter, improving (CFAR) detection performance in partially homogeneous environments. Further advancements, such as generalized forms of the multivariate K-distribution, have enhanced estimation for multilook polarimetric SAR data, addressing limitations in traditional Wishart models by incorporating variability. Approximate inference methods have addressed challenges in parameter for high-dimensional K-distributed , moving beyond classical moment-based approaches to handle in clutter analysis. Markov chain Monte Carlo (MCMC) techniques, including , provide robust Bayesian of shape and scale parameters by sampling from posterior distributions, particularly effective for sea clutter modeling where exact likelihoods are intractable. Similarly, variational Bayes approximations optimize evidence lower bounds to infer the in K-distributions, offering faster convergence for real-time applications like while maintaining accuracy comparable to MCMC in low-sample regimes. These methods have proven superior in high-dimensional settings, reducing estimation bias in textured clutter scenarios. Software implementations have facilitated practical adoption of the K-distribution across disciplines. In R, the VGAM package supports fitting of gamma and generalized gamma distributions, enabling simulation and parameter estimation for K-distributed processes through compound modeling, though direct K-family functions require custom compounding. Python's SciPy library provides essential tools via the special module's modified Bessel functions (e.g., kv for order ν), allowing computation of the K-distribution's probability density function involving the second-kind Bessel kernel. Simulation algorithms typically employ gamma-gamma mixing: generate a gamma texture variable τ ~ Gamma(α, β), then a chi-squared speckle variable, and compute the amplitude as √(τ · χ²_{2L}/L) for L looks, yielding efficient Monte Carlo samples for validation in radar simulations. Theoretical extensions include the generalized K (GK) distribution, which introduces asymmetry to capture heavier tails on one side, extending the symmetric compound gamma form for applications beyond radar. Post-2010 developments have applied GK variants, such as the κ-generalized distribution, to model heavy-tailed income distributions, where skewness and kurtosis better fit empirical data compared to normal or Student's t distributions. Emerging open areas involve integrating machine learning for real-time radar processing, such as gradient boosting decision trees for self-learning parameter estimation in K-distributed sea clutter, achieving low-latency shape and scale inference with minimal training data. Neural networks have also enhanced shape parameter estimation accuracy in heterogeneous clutter, paving the way for adaptive detection in dynamic environments. As of 2025, recent advancements include refined CFAR detection algorithms tailored for heterogeneous K-distributed sea clutter backgrounds, improving target discrimination in maritime radar systems.

References

  1. [1]
    A model for non-Rayleigh sea echo
    **Summary of Content from https://ieeexplore.ieee.org/document/1141451:**
  2. [2]
    [PDF] Calculation of Radar Probability of Detection in K Distributed Sea ...
    The detection performance of maritime radars is usually limited by sea clutter. The K distribution is a well established statistical model of sea clutter ...
  3. [3]
    [PDF] Estimating the Parameters of the K Distribution in the Intensity Domain
    This paper reviews a number of different moment-based methods for esti- mating the parameters of the K distribution. The K distribution is a model.
  4. [4]
    K-distribution and polarimetric terrain radar clutter
    A multivariate K-distribution is proposed to model the statistics of fully polarimetric radar data from earth terrain with polarizations HH, HV, VH, and VV.
  5. [5]
    Modeling and simulation of radar sea clutter using K-distribution
    This paper discusses the modeling and simulation of K-distributed sea clutter to help in understanding the clutter characteristics from a statistical viewpoint.
  6. [6]
    Sea Clutter: Scattering, the K Distribution and Radar Performance
    This book attempts to bring together those aspects of maritime radar relating to scattering from the sea surface, and their exploitation in radar systems.
  7. [7]
    Probability density function formalism for optical coherence ...
    Probability density function formalism for optical coherence tomography ... K distribution for both OCT amplitude and intensity. The PDF formalism is ...
  8. [8]
    [PDF] Modelling the Statistics of Microwave Radar Sea Clutter - ARPI
    The most popular compound model is the K distribution which is characterised by two parameters. (shape and scale) which can then be related to variations in the ...
  9. [9]
    [PDF] On the Approximation of the Generalized-K Distribution by a Gamma ...
    Feb 5, 2010 · and third moments of the generalized-𝐾 distribution and the ... Gamma function as defined in [20, eq. 8.350.2]. Similar to the BER measure ...
  10. [10]
    Generalized K distribution: a statistical model for weak scattering
    **Summary of Generalized K-Distribution (J. Opt. Soc. Am. A, Vol. 4, Issue 9, pp. 1764)**
  11. [11]
    https://ieeexplore.ieee.org/document/1458425
  12. [12]
    https://www.mathworks.com/help/matlab/ref/besselk.html
  13. [13]
    besselk - Modified Bessel function of second kind - MATLAB
    K = besselk(nu,Z,scale) specifies whether to exponentially scale the modified Bessel function of the second kind to avoid underflow or loss of accuracy.
  14. [14]
    https://apps.dtic.mil/sti/tr/pdf/ADA424271.pdf
  15. [15]
    [PDF] Clutter Spatial Distribution and New Approaches of Parameter ...
    Statistical properties of the Weibvtll and K- distributions in the log domain are derived and then used in new approaches, named as imbiased estimation schemes, ...
  16. [16]
    Statistical Modeling of SAR Images: A Survey - PMC - PubMed Central
    Several empirical models have been used to characterize the statistics of SAR amplitude or intensity data, such as Weibull, log-normal, and Fisher PDFs. The log ...
  17. [17]
    Synthetic Aperture Radar Image Background Clutter Fitting Using ...
    Sep 30, 2015 · Among the SAR image clutter statistic models, the modeling ability of K distribution is good for the heterogeneous clutters, but it cannot ...
  18. [18]
    https://arxiv.org/abs/2210.05833
  19. [19]
    Superpixel-Based Classification Using K Distribution and Spatial ...
    where K v ( · ) is the modified Bessel function of the second kind with order v . The K distribution is parameterized by the shape parameter α , the number of ...<|control11|><|separator|>
  20. [20]
    Parameter estimation of the homodyned K distribution based ... - arXiv
    Oct 11, 2022 · In this work, we propose a machine learning based approach to the estimation of the HK distribution parameters. We develop neural networks ...Missing: SAR post- 2010
  21. [21]
    Significance of $K$ Distributions in Scattering Experiments
    Feb 27, 1978 · Jakeman and P. N. Pusey, IEEE Trans. Antennas Propag. 24, 806 (1976); H. R. Pitt, Integration Measure and Probability (Oliver and Boyd ...
  22. [22]
    [PDF] Bayesian Estimation of Sea Clutter Parameters for Radar - kth .diva
    May 23, 2024 · For a considerable duration, the K-distribution model has served as the predominant statistical model for characterizing the backscattered ...
  23. [23]
    https://cran.r-project.org/web/packages/VGAM/VGAM.pdf
  24. [24]
    [PDF] VGAM: Vector Generalized Linear and Additive Models - R Project
    Feb 12, 2025 · This package fits many models and distributions by maximum likelihood estimation (MLE) or penalized MLE, under this statistical framework.
  25. [25]
    scipy.special.kv — SciPy v1.16.2 Manual
    Returns the modified Bessel function of the second kind for real order v at complex z. These are also sometimes called functions of the third kind, Basset ...<|separator|>
  26. [26]
    A new model of income distribution: the κ-generalized distribution
    Jun 24, 2011 · This paper proposes a three-parameter statistical model of income distribution by exploiting recent developments on the use of deformed ...
  27. [27]
    [PDF] Self-learning parameter estimation of K-distributed clutter using ...
    Aug 21, 2022 · K distribution is a widely-used and effective amplitude probability model of sea clutter at low and moderate range resolution [4]. Currently, ...<|control11|><|separator|>