UC San Diego

Freshwater exchanges between the ocean--ice--atmosphere system play a crucial role in the global climate system. This study provides an analysis of the local impact of freshwater fluxes both off the coast of California and in the Arctic. Studies are carried out using observations and numerical and statistical models. We show that freshwater exchanges between the ocean and atmosphere in the form of precipitation from atmospheric rivers (ARs) over the ocean in the California Current System (CCS) have impacts on the surface ocean salinity on event and seasonal timescales. In the upper ocean, precipitation from ARs can produce long-lasting layers of freshwater, the extent of which are dependent on atmospheric forcing from precipitation and wind. We conclude that upper ocean salinity changes due to ARs are within the limits of detectability of ocean instruments.

We also examine the extent to which wind acts as a driving force for ice motion in the Arctic. To accomplish this, we build a sequence of machine learning (ML) models that make one-day predictions of present-day zonal and meridional sea-ice velocity components from inputs of present-day wind velocity, previous-day sea-ice velocity, and previous-day sea-ice concentration. We analyze the performance of these models, and implement explainable machine learning (XML) methods to understand how they are making their predictions. One of these methods, layerwise relevance propagation (LRP), was developed for ML models that make classification rather than regression predictions. This study is the first known implementation of a global LRP for a regression problem in geosciences. We therefore provide a comparative study of several different XML methods to bolster trustworthiness in the use of LRP for this particular application. A convolutional neural network (CNN) has improved performance compared conventional persistence (PS) and linear regression (LR) models. Outputs from local LRP studies are shown to be consistent with other XML methods. However global implementations of LRP are highly sensitive to choices made during processing. We analyze the coefficients of the LR model to understand the relationship between ice motion and wind speed. We show that the ice is becoming more responsive to wind forcing, and link this to decreasing ice concentration.