A U-Net network-based downward continuation method for airborne gravity

Downward continuation of gravity data is a critical challenge in the practical applications of airborne gravity, such as global geopotential modeling and geophysical interpretation. This study introduces a machine learning algorithm to address this classical problem in physical geodesy. We first constructed a U-Net-based neural network, termed DWC-Unet, specifically designed for gravity anomaly downward continuation. The model was trained using gravity anomalies simulated by the prior global gravity model. Subsequently, the trained DWC-Unet was used on airborne gravity anomalies to calculate the downward continued gravity anomalies at an arbitrary altitude surface. We applied the proposed method to the airborne gravity data in Colorado, Iowa, and regions in the southeast coast of the United States. The experiment results confirm that the DWC-Unet method shows better performance than the analytic continuation algorithm. The standard deviation of downward continuation errors reached 4.39 mGal in Colorado and approximately 3 mGal in other regions when compared to terrestrial gravity data. This study demonstrates the good performance of applying machine learning to the calculation of gravity data continuation, serving as a reference for the application of artificial intelligence in the field of gravity data processing.