Loading...
Citations
Altmetric:
Date
2025-12
Abstract
In this thesis, we explore Machine Learning (ML) methods for solving inverse problems, focusing on both static imaging with Electrical Impedance Tomography (EIT) and dynamic imaging with Synthetic Aperture Radar (SAR). Inverse problems, which are at the basis of any imaging modality, are inherently challenging due to their ill-posed nature, where small errors in measurements can lead to large discrepancies in the reconstructed image. ML techniques offer a powerful framework to address these challenges by learning complex mappings from measurement data to the desired image or parameter distribution of interest. In the first part of this thesis, we propose an ML approach to EIT. For 2D EIT, we employ Support Vector Machines (SVM) to detect the presence of one or more inclusions using the Dirichlet-to-Neumann (DN) matrix Lσ, which encodes the relationship between applied voltages and resulting currents on the boundary of the domain under investigation. We further investigate the issue of detecting the size of one or more inclusions and the identification of different types of anisotropy, including diagonal anisotropy and non-constant anisotropy within such inclusions, using Artificial Neural Networks (ANNs). Our results in the 2D setting in EIT demonstrate the efficacy of our ML-based approach and lay the groundwork for our extension to 3D EIT. In 3D anisotropic EIT, we extend our 2D analysis to assess the generalisability of our methods, observing that in 3D there is a decrease in accuracy due to the increased complexity of the problem and the higher dimensionality of the data.
In the second part of this thesis, we address object shape detection in SAR using Convolutional Neural Networks (CNNs), which are well suited for extracting spatial features from image data. We achieve high accuracy (≥ 90%) with both raw SAR data and reconstructed SAR images, but we highlight that the use of raw SAR data leads to better classification with CNNs. This suggests that the computationally expensive image generation step can be skipped in machine learning workflows for SAR data, which has the potential to save significant computational time. We also apply our approach to real-world satellite SAR data for classifying different ice types, a crucial task for climate monitoring and navigation, again demonstrating high accuracy and the robustness of our ML models. A comparative analysis of the inclusion determination problem reveals that dynamic imaging with SAR offers advantages over static imaging with EIT in certain scenarios, particularly in the case of detecting the number of inclusions. On the other hand, EIT offers better inclusion size detection than SAR.
Supervisor
Description
Publisher
University of Limerick
Citation
Collections
Files
Loading...
Naidu_2025_Machine.pdf
Adobe PDF, 2.72 MB
ULRR Identifiers
Funding code
Funding Information
Sustainable Development Goals
External Link
License
Attribution-NonCommercial-ShareAlike 4.0 International