Julien Brajard (NERSC, Norway); Antonio Bonaduce (NERSC, Norway)

Presenting author: Antonio Bonaduce

The goal is to increase the resolution of the DUACS data set consisting of full, Low Resolution (LR) fields of SSH with a neural network using the new SWOT SSH fields consisting of sparse, High Resolution (HR) fields of SSH.

Because the difference in resolution between the two products is high (approximately a factor 12 in both horizontal and vertical dimension), we chose a generative neural network. Indeed many HR images can correspond to one LR one. A standard U-Net architecture tends to average all of these possible realizations and thus struggles to produce high frequency features (new eddies, internal waves …), whereas a generative network can output individual possible HR realizations.

We therefore trained a Conditional Generative Adversarial Network (CGAN) on matching pairs of DUACS and SWOT images. To tackle the sparsity of data available in SWOT, those images are small patches localized along the satellite paths. The full super resolution of the DUACS data is then done patch by patch.

We are able to match the spectral quality of the HR data, i.e. we can produce out of a DUAC image several high resolution images that “look like” SWOT data, although the RMSE with the truth can decrease.