Demonstrating Superresolution in Radar Range Estimation Using a Denoising Autoencoder

TL;DR

This paper demonstrates that a denoising autoencoder can achieve superresolution in radar range estimation, surpassing traditional limits by leveraging machine learning and optimized signal design.

Contribution

It introduces a machine learning approach using a denoising autoencoder for superresolution radar range estimation, emphasizing the importance of signal design and robustness.

Findings

Bessel signals outperform sinc and triangle pulses in range resolution.

Autoencoder's bottleneck correlates strongly with true scatterer separation.

Method is robust across training sessions and noise conditions.

Abstract

We apply machine learning methods to demonstrate range superresolution in remote sensing radar detection. Specifically, we implement a denoising autoencoder to estimate the distance between two equal intensity scatterers in the subwavelength regime. The machine learning models are trained on waveforms subject to a bandlimit constraint such that ranges much smaller than the inverse bandlimit are optimized in their precision. The autoencoder achieves effective dimensionality reduction, with the bottleneck layer exhibiting a strong and consistent correlation with the true scatterer separation. We confirm reproducibility across different training sessions and network initializations by analyzing the scaled encoder outputs and their robustness to noise. We investigate the behavior of the bottleneck layer for the following types of pulses: a traditional sinc pulse, a bandlimited triangle-type pulse, and a theoretically near-optimal pulse created from a spherical Bessel function basis. The Bessel signal performs best, followed by the triangle wave, with the sinc signal performing worst, highlighting the crucial role of signal design in the success of machine-learning-based range resolution.