Abstract
Accurate and comprehensive data remain critical for modeling and understanding Earth's complex systems, directly influencing weather forecasting, climate change predictions, and disaster management strategies. However, the scarcity of data, particularly for rare or extreme events, and the inherent imbalance in datasets pose significant challenges to developing robust predictive models. These issues highlight the need for effective data augmentation techniques, a domain where existing methodologies remain underexplored for geophysical data. Addressing this gap, this study introduces a data augmentation framework for geophysical fields, employing a Generative Adversarial Network (GAN) architecture. Our GAN's generator utilizes a UNet architecture combined with depthwise separable convolutions to capture multi-scale spatial hierarchies while also reducing computational cost. The discriminator is enhanced with residual attention mechanisms to distinguish simulations from observations. Beyond the standard GAN loss, a Mean Absolute Error (MAE) regularization term is incorporated to ensure the generated data fields are distinguishable from the original dataset, promoting diversity and enhancing model training. Our approach has been validated through its application to downstream tasks including downscaling, extrapolation, and imputation. It achieves outstanding performance improvements, reducing the Mean Absolute Percentage Error (MAPE) by 25.1%, 19.6%, and 27.4% across these tasks, respectively.