Ryan Holmes: Coastal downscaling of Australian seasonal sea level and sea surface temperature forecasts using neural networks
Operational seasonal forecasts of sea level and sea surface temperature (SST) from the Bureau of Meteorology’s seasonal forecasting system, ACCESS-S2, support decision-making across a wide range of sectors. However, the system’s coarse ocean resolution (~25 km) limits its ability to resolve coastal variability and processes, where forecast information is often most valuable to end users. This study presents an initial investigation into the potential for low-cost machine learning methods, including a super-resolution convolutional neural network (SRCNN), to dynamically downscale seasonal forecasts of sea level and SST from 25 km to 8 km resolution. Several SRCNN architectures are trained and evaluated on the GLORYS global ocean reanalysis within a perfect-model framework. Preliminary results indicate that the use of spatially weighted loss functions, together with mask-aware convolutions that account for missing data over land, substantially improves downscaling performance near the coast. I discuss the implications of these results for improving ACCESS-S2 coastal forecast skill.