Improving ocean reanalyses of observationally sparse regions with transfer learning

Oceanic subsurface observations are sparse and lead to large uncertainties in any model-based estimate. We investigate the applicability of transfer learning based neural networks to reconstruct North Atlantic temperatures in times with sparse observations. Our network is trained on a time period with abundant observations to learn realistic physical behavior. Evaluating it within a consistent data assimilation framework, this network learns and reproduces its training data's physical patterns. Additionally, the network is able to transfer these patterns towards a historical ocean heat content estimate in times with sparse observations. Consequently, with infrequent input data, machine learning reconstructions exhibit similar physical structures, while correcting for known errors compared to state-of-the-art data assimilation products. In this manner, transfer learning can impact the initialization and evaluation of climate hindcasts. Furthermore, by exhibiting the capability to accurately transfer results from high to low-frequencies, transfer learning based neural networks showcase their relevance in mixed-frequency measurements beyond climate science.