PIRESGAN: Radio Echo Sounding Signal Synthesis with Physic-informed Generative Adversarial Networks
Qingchuan Ma, Tong Hao · 2024
Recognizing subglacial lakes (SLs) through radio echo sounding (RES) data has become a prominent research area in glaciology. Despite the continuous collection of extensive heterogeneous RES data, extracting meaningful information remains a significant challenge. Recent advancements in machine learning (ML) techniques have spurred scientists to explore and leverage more advanced algorithms to process geophysical data collected in polar regions. However, such progress has primarily occurred in the supervised learning domain, which demands large amounts of annotated data. Generating synthetic data can serve as an effective and cost-efficient approach to provide extensive labelled datasets to train ML models. In this study, we propose a novel "Physic-informed Radio Echo Sounding Generative Adversarial Networks (PIRESGAN)" framework, which utilizes the GAN model based on the physical properties embedded within RES data to generate a large volume of synthetic SL RES data. The whole procedure can be trained end-to-end in an adversarial training paradigm, which alleviates the challenges associated with manual annotation. To validate the effectiveness of our proposed method, we showcase our experimental results and demonstrate that, by combining physical properties and data-driven approaches, the simulated SL RES signals exhibit a high degree of structural similarity to the real RES signals.