A Novel Method for Sea-Land Clutter Separation Using Regularized Randomized and Kernel Ridge Neural Networks

Le Zhang, Jeyan Thiyagalingam, Anke Xue, Shu-Wen Xu

TL;DR

This paper introduces two novel machine learning approaches, Regularized Randomized Neural Networks (RRNN) and Kernel Ridge Regression (KRR), for accurately distinguishing between sea and land clutter in radar data. By developing a new feature set called ECAV, the proposed methods achieve state-of-the-art classification accuracy, outperforming conventional techniques like SVM and ELM.

What Problem Does This Paper Address?

Radar systems are crucial for surveillance and tracking, but "clutter" – unwanted radar returns from the environment – can obscure targets. Distinguishing between different types of clutter, particularly sea and land clutter, is a significant challenge, especially at interfaces where both are present. Traditional methods often rely on pre-existing maps or complex statistical models that can be inaccurate or difficult to adapt. This paper addresses the need for a robust, data-driven method that can effectively and accurately separate sea and land clutter, improving the reliability of radar systems in complex environments.

Methods

The core of this paper lies in proposing two novel machine learning classifiers:

  • Regularized Randomized Neural Network (RRNN): This is a single-hidden-layer feedforward neural network. Its key innovation is that the hidden layer weights are randomly initialized and fixed, while only the output weights are trained using a linear learning rule. This significantly simplifies the hyperparameter tuning process compared to traditional neural networks.
  • Kernel Ridge Regression (KRR): This method extends ridge regression by incorporating kernel functions, allowing it to model non-linear relationships. It transforms the input data into a higher-dimensional feature space where linear regression can be applied.

A crucial aspect of their approach is the development of a novel feature set called ECAV (Energy, Zero-crossing rate, Autocorrelation, and Shape parameter V). These features are extracted from short time-domain segments of the radar signal and are designed to capture the distinct statistical characteristics of sea and land clutter.

Datasets

The study utilizes a mixed dataset:

  • Real Sea Clutter Data: Acquired from the IPIX radar (9.3 GHz, 1 kHz PRF, 30m resolution) with multiple polarization states. This dataset contained a large number of pure sea clutter samples, some with embedded targets.
  • Simulated Land Clutter Data: Since the IPIX dataset lacked land clutter, the authors simulated it. Initially, they modeled land clutter using a Weibull amplitude compound-Gaussian distribution. Later experiments also explored a log-normal distribution for land clutter.

The dataset was split into an 80:20 ratio for training and testing. They also introduced multiple dataset configurations (DS-0, DS-1, DS-2, DS-3) to thoroughly evaluate the classifiers under varying conditions and data volumes.

Key Insights and Ideas

  • ECAV Features are Powerful: The proposed ECAV feature set, combining energy, zero-crossing rate, autocorrelation, and shape parameter V, proved highly effective in distinguishing sea and land clutter. This highlights the importance of carefully engineered features for this classification task.
  • RRNN and KRR Offer Simplicity and Performance: The RRNN's randomized hidden layer and KRR's kernel-based approach provide efficient and effective classification, outperforming traditional methods. The RRNN, in particular, simplifies the complex hyperparameter tuning often associated with neural networks.
  • Parameter Sensitivity Varies: While both RRNN and KRR achieved high accuracy, KRR was found to be more sensitive to hyperparameter tuning than RRNN, requiring more careful selection of parameters for optimal performance.
  • Frame Length Matters: The choice of frame length for signal segmentation significantly impacts classification accuracy and training time. A frame length of 512 was found to be optimal in their experiments.
  • Data Adaptability is Key: The success of machine learning approaches like RRNN and KRR, especially with their data-adaptive nature, underscores the limitations of static statistical models for highly variable clutter statistics.

Key Results

  • High Classification Accuracy: The proposed RRNN and KRR methods achieved impressive classification accuracies of 98.50% and 98.75%, respectively, on their mixed dataset.
  • Superior Performance: These accuracies significantly outperformed conventional methods like Support Vector Machine (SVM) and Extreme Learning Machine (ELM), which achieved lower accuracies (e.g., around 97.4% for ELM and SVM in one comparison).
  • Efficient Training: The RRNN classifier demonstrated a good balance between accuracy and training time, often outperforming ELM and SVM in both aspects.
  • Feature Impact: The ECAV feature set consistently yielded the best classification performance compared to other feature sets like AVC, Wavelet (db3), and Wavelet (fk6), and even when using raw data ("None").

Limitations and Future Work

  • Land Clutter Simulation: While the simulated land clutter aimed to mimic real-world scenarios, the authors acknowledge that modeling land clutter is inherently complex due to its spatial non-uniformity and variability with weather conditions. Future work could explore more sophisticated land clutter modeling techniques.
  • Dataset Specificity: The study used a specific radar (IPIX) and a particular set of simulated land clutter. While the results are promising, further validation on diverse radar systems and real-world land clutter datasets would be beneficial.
  • Unbalanced Datasets: The paper notes that the impact of training on unbalanced datasets is not fully clear, suggesting this as an area for future investigation.
  • Generalization: While the paper demonstrates good performance across different dataset configurations, further research into the generalization capabilities of the proposed methods across a wider range of radar parameters and environmental conditions would be valuable.

Summary updated

BibTeX
@inproceedings{zhang2020novel,
  author = {Le Zhang and Jeyan Thiyagalingam and Anke Xue and Shu-Wen Xu},
  title = {{A Novel Method for Sea-Land Clutter Separation Using Regularized Randomized and Kernel Ridge Neural Networks}},
  booktitle = {Sensors},
  year = {2020},
  url = {https://openreview.net/forum?id=0afGoDZBek},
}

Loading related content...

Figure