Estimation and Target CFAR Detection in Sea Clutter with Compound Gaussian Distribution

Abstract

Radar signal processing plays a critical role in modern surveillance, remote sensing, and defense applications, where accurate target detection based on clutter parameter estimation are crucial in complex sea or land environments. This thesis investigates advanced signal processing techniques for radar clutter modeling, parameter estimation, and target detection, with a particular focus on integrating deep learning methodologies. The study begins with an in-depth review of sea clutter modeling and Constant False Alarm Rate (CFAR) detection strategies, followed by the introduction of a model selection framework using cross-validation techniques for compound clutter distributions, including K-distributed, Pareto Type II, and Compound Gaussian-Inverse Gaussian (CG-IG) models. This framework allows the radar processor to dynamically select the most appropriate clutter model from real-time data, improving detection robustness. Building on this foundation, the dissertation presents four main contributions addressing modeling, parameter estimation, and detection in challenging radar clutter environments. The first contribution investigates statistical model selection for sea clutter, proposing a systematic framework based on information-theoretic criteria and cross-validation strategies to identify the most suitable distribution under varying sea states and operating conditions. The second contribution introduces a combined Convolutional Neural Network (CNN) with Long Short-Term Memory (LSTM) architecture for estimating the shape parameter of K-distributed clutter in the presence of additive Gaussian noise. This data driven estimator demonstrates improved robustness compared with conventional moment based approaches, particularly in low Signal to Noise Ration (SNR) and nonstationary scenarios. The third contribution develops a multi-headed deep learning framework for correlated Pareto Type II clutter, enabling joint estimation of shape and scale parameters under correlated returns. By integrating convolutional feature extraction with recurrent modeling and autoencoding structures, the proposed method achieves enhanced estimation accuracy and computational efficiency. The final contribution addresses distributed radar detection by formulating the tuning of decentralized Greatest-Of (GO)CFAR and Smallest-Of (SO)-CFAR factors as a Neyman-Pearson constrained optimization problem and solving it using the Moth-Flame Optimization (MFO) algorithm. The resulting framework ensures a prescribed global false alarm rate while improving global detection probability in both identical and non-identical sensor configurations. Extensive numerical evaluations validate the efficiency of the proposed methods using both synthetic and real radar data from the Intelligent Pixel processing X-band (IPIX) radar experiment data. Results show that the integration of Deep-Learning (DL)-based estimators into radar signal processing chains can lead to highly accurate, real-time adaptive detection systems that are robust to clutter non-stationarity and correlation. This work paves the way for future research into fully cognitive radar systems, where clutter modeling, parameter estimation, and detection thresholds are dynamically controlled through learning-driven paradigms. The results demonstrate the practical potential of combining statistical signal processing and deep learning for next-generation intelligent radar systems.