Abstract
In this work, we compare deep convolutional and generative models to improve the reconstruction of sea surface temperature (SST) gradients and effective resolution from cloud-contaminated satellite infrared data. Standard gap-filling techniques produce spatially complete datasets, but they inherently smooth out fine-scale features and gradients that may be critical for a better understanding of the ocean dynamics. Here, we investigate how the use of deep learning models as Autoencoders (AEs) and generative models as Conditional Generative Adversarial Networks (C-GANs) differently contribute to recovering the small-scale structures lost during interpolation. Our supervised, observation-based training relies on satellite-derived SST data from the Mediterranean Sea, with a focus on learning the statistical relationship between high-resolution fields and their low-resolution counterparts. We apply a tiling and merging strategy to deal with limited observational coverage and to ensure spatial continuity. Quantitative validation based on mean squared error metrics, spectral analysis, and gradient statistics shows that the AE reduces the reconstruction error, but it fails to recover high-frequency variability. In contrast, the C-GAN effectively restores the statistical properties of the true SST field at the cost of increasing the pointwise discrepancy with the ground truth observations. Our results highlight the potential of deep models to enhance the physical and statistical realism of gap-filled satellite data in oceanographic applications.