GANterpolate: Improving Climate Modeling through Synthetic Data Generation

Data sparsity remains a significant challenge in climate modeling, particularly in remote and under-observed regions that play a crucial role in global climate dynamics. This study proposes GANterpolate, a novel hybrid approach that integrates Generative Adversarial Networks (GANs) with interpolation techniques to reconstruct missing surface temperature data. Utilizing the ERA5 reanalysis dataset, artificial data gaps are systematically introduced in climatologically significant regions such as the Arctic and Equatorial Pacific. The GAN-based generator produces an initial reconstruction, while interpolation methods refine the generated data to enhance accuracy and spatial coherence. The discriminator, a deep convolutional neural network (CNN), evaluates the authenticity of the reconstructed data, optimizing adversarial loss to improve realism. Comparative analysis against traditional interpolation techniques demonstrates that GANterpolate outperforms standalone methods in capturing complex spatial-temporal dependencies. Quantitative results indicate that GANterpolate achieves a Root Mean Square Error (RMSE) of 0.320 and a Mean Absolute Error (MAE) of 0.126 with a perfect Pearson correlation of 1.0 in randomly masked scenarios. In clustered masked regions, it achieves an RMSE of 5.368, MAE of 4.161, and Pearson correlation of 0.958, significantly surpassing both GAN-only and interpolation-based baselines. These findings demonstrate the transformative potential of synthetic data generation in enhancing the quality of climate datasets, paving the way for more robust and reliable climate simulations in data-scarce regions.