Ajitha Cyriac: Validating satellite-derived data for sea surface temperature downscaling using machine learning
Validating satellite-derived data for sea surface temperature downscaling using machine learning
Ajitha Cyriac a, Chaojiao Sun a, Richard Matear b, Jim Greenwood a, Fabio Boschetti a
aCSIRO Environment, Crawley, bCSIRO Environment, Hobart
Machine learning (ML) has become a powerful tool to produce robust high-resolution climate information by its ability to rapidly downscale coarse-resolution climate models. However, the effectiveness of ML approaches in climate applications is often constrained by the availability of high-resolution training data. This study explores the applicability of satellite-derived observations in training super-resolution deep learning architectures to downscale sea surface temperatures (SST) along the Australian coastline. We evaluate two models: one trained on observational satellite data and another on high-resolution numerical ocean simulations. Our results indicate that the observationally trained framework achieves comparable performance to its simulation-trained counterpart, establishing empirical satellite data as a robust, accessible alternative to computationally demanding model outputs. By validating these satellite-derived training sets, we provide a scalable framework for high-fidelity climate downscaling in data-sparse regions, offering a critical tool for assessing climate impacts on remote World Heritage sites such as the Great Barrier Reef and Ningaloo Coast.