[ACG44-07] Deep learning for mesoscale air-sea interactions
★Invited Papers
Keywords:mesoscale eddy, air-sea interaction, machine learning
In the field of machine learning, convolutional neural networks have enabled breakthroughs in computer vision and natural language processing. Here we explore the use of convolutional neural networks, or deep learning, to study mesoscale air-sea interactions. We train our deep learning model with satellite observations of SST and surface winds, and compare the model performance to conventional methods that utilize linear regression in physical space and Fourier wavenumber space. Given enough data, the deep learning model has the potential to outperform conventional methods based on a mean square error loss metric. This will allow for ocean model simulations forced by observed satellite winds that are adjusted to remove wind effects due to spontaneous mesoscale eddy activity, and thus consistent forcing of ocean models with satellite-observed winds.