ERTGAN: Electrical resistivity tomography subsurface model generator with evaluation using TabSynDex

• Main Author: Ducut, Jullian Dominic
• Format: text
• Language: English
• Published: Animo Repository 2023
• Subjects: Tomography; Computerized tomography; Other Engineering; Other Operations Research, Systems Engineering and Industrial Engineering
• Online Access: https://animorepository.dlsu.edu.ph/etdm_mem/9; https://animorepository.dlsu.edu.ph/context/etdm_mem/article/1010/viewcontent/ERTGAN__Electrical_Resistivity_Tomography_Subsurface_Model__Gener.pdf
• Institution: De La Salle University

In the ever-evolving landscape of urban infrastructure, the repercussions of leaking pipes on pavements have emerged as a pressing concern, demanding innovative solutions to ensure both structural integrity and public safety. Recognizing this challenge, the focus of this comprehensive study is on harnessing the potential of subsurface imaging techniques, with a particular emphasis on Electrical Resistivity Tomography (ERT). Traditional ERT methodologies, despite their advanced nature, are being integrated with modern Deep Learning (DL) techniques that brings exhaustive dataset requirements. Such extensive data prerequisites often lead to prolonged processing durations, necessitating significant manpower and computational resources, impeding the overall efficiency and cost-effectiveness of ERT DL techniques. To address these and to pave the way for a new era in subsurface imaging, this study introduces and extensively elaborates on ERTGAN, a type of Generative Adversarial Network (GAN). ERTGAN's primary objective is the generation of synthetic apparent resistivity profiles. These profiles, synthesized based on subsurface models, encapsulating the nature of urban underground structures, including both metallic and plastic pipes. This study illustrates a methodical process and can be partitioned into distinct yet interconnected phases. The initial phase revolves around the creation of a synthetic dataset, which is achieved through subsurface modeling of apparent resistivity profiles. Building on this foundation, the subsequent phase involves the development and iterative refinement of the ERTGAN model. By leveraging the synthetic dataset, the model undergoes a series of fine-tuning processes, incorporating feedback loops and ensuring alignment with real-world subsurface structures. The validation of the model's efficiency and accuracy is of paramount importance, leading to the third phase, which is performance evaluation. The primary evaluation metrics employed is the TabSynDex. Providing an overall score of 0.559 and 0.421 underscores a moderate degree of alignment between the synthetic and real-world apparent resistivity profiles, there are areas demanding attention. Notably, the synthetic data's ability to replicate key statistical attributes is commendable, as evidenced by a score of 0.934, which indicates alignment with real data's intrinsic statistical properties. However, challenges emerge in the form of correlation and machine learning scores, which, at 0.337 and 0.168 respectively, signal the need for further refinement. Despite these hurdles, the overarching narrative of this study remains positive. In the last phase of this study, a GUI was implemented to easily bridge future users on using ERTGAN, that promotes its potential on the field of ERT and DL. With this study’s pioneering approach and robust framework, has the potential to revolutionize subsurface imaging. Its ability in replicating key statistical parameters and its ability to encompass a vast range of values offers a promising pathway for the early and efficient detection of subsurface models.