Conditional Generative Models for OceanBench Sea Surface Height Interpolation

Rising sea levels are one of many consequences of anthropogenic climatechange. Over the past few decades, several global observational records havebecome available that give a more detailed picture of the increasingimpacts. Nevertheless, there continue to be data challenges, such assparsity or signal to noise ratio, that need to be dealt with. Machine Learning (ML)and specifically, Deep Learning (DL) approaches have presented themselves as valuabletools for such large-scale and complex data sources. To this end, the OceanBenchBenchmark suite was recently developed to provide astandardized pre-processing and evaluation framework for Sea Surface Height(SSH) interpolation tasks involving nadir and Surface Water and Ocean Topography(SWAT) Altimetry Tracks. From the methodological perspective, a reoccurringissue is the lack of uncertainty quantification for DL applications in EarthObservation. Therefore, we extend the suite of metrics provided by OceanBenchto probabilistic evaluation metrics and test state-of-the-art uncertaintyquantification models from the DL community. Specifically, we focus onConditional Convolutional Neural Processes (ConvCNP) andInpainting Diffusion models as methodologies to quantifyuncertainty for the interpolation task and demonstrate their viability andadvantages over other ML methods for both accuracy and probabilistic metrics.