Date of Award

2024

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Electrical Engineering

First Advisor

Tian Tia

Abstract

Ground Penetrating Radar (GPR) is a non-invasive geophysical method that uses radar pulses to image the subsurface. This technology is widely used to detect and map subsurface structures, utilities, and features without the need for physical excavation. Traditional GPR systems, which rely on fixed radio frequency electronics like Application-Specific Integrated Circuits (ASICs), have significant limitations in their flexibility and adaptability. Adjusting operational parameters such as waveform, frequency, and modulation schemes is challenging, which is crucial for tailoring performance to specific tasks or conditions. The considerable size and weight of these systems restrict their applicability in harsh or confined spaces where mobility and portability are required. Moreover, the process of developing custom GPR systems is both expensive and time-consuming, requiring individual design, testing, and modification.

This dissertation explores the potential of Software Defined Radio (SDR) and Deep Learning (DL) to advance GPR capabilities. The first part of the dissertation addresses limitations of traditional GPR systems by proposing a more flexible solution using SDR. By modifying the software configuration, the SDR-based GPR is able to work in different modes, including a Doppler mode for measuring target vibrations and an imaging mode for locating subsurface objects. Through theoretical analysis and practical experimentation, the first part aims to develop a next-generation GPR platform that offers improved adaptability, ease of use, and efficiency in subsurface imaging for various applications including archaeological exploration, civil engineering, and scientific discovery applications.

The second part of the dissertation investigates deep learning methods for improving GPR image processing, with a focus on the reduction of rough surface clutter in GPR B-scan images. Traditional surface clutter reduction methods have limitations when the patterns of surface and target responses deviate from the assumptions made by these methods. Moreover, current deep learning methods, relying on the supervised learning framework, require that users collect and label a vast collection of B-scans. This requirement is often impractical and resource-intensive in many scenarios. To address these challenges, the second part introduces two deep learning (DL)-based algorithms designed for data efficiency. This significantly reduces the need for users to collect and label a large dataset of B-scan images. The first method employs an auto-encoder, and the second utilizes PixelCNN. Experimental results are provided to demonstrate the efficacy of these methods.

Language

en

Number of Pages

190 p.

Download

Share

COinS