Special Lecture — J. Xavier Prochaska (UC Santa Cruz)

“On the information content of remote sensing datasets”

Geoscientists have, for decades, leveraged satellite remote-sensing products to characterize and study the physical and biological states and dynamics of Earth’s surface. Across the ocean, my focus, this includes twice daily observations of the sea surface temperature (SST), roughly daily views of the ocean’s color (now hyperspectral; NASA/PACE), and less frequent measures of the sea surface height (SSH; ~18 days with NASA/SWOT). From these datasets, we assess mesoscale processes (e.g. eddies, filaments) that govern ocean dynamics on ~100km scales. In this talk, I will describe our AI investigation into the information content of remote-sensing products by examining the latent spaces of so-called “foundation models” constructed from SST, SSH and ocean color data. We find, similar to results on natural images (e.g. cats, dogs, handwritten digits) that the latent spaces of remote-sensing on scales of ~1-100 km (i.e. the sub-mesoscale) show ~100 eigenmodes with significant variance. We then explore these results to examine the information content of SST and SSH relative to ocean model outputs. Last, we use our framework to reveal the fundamental “eigenmodes’’ of remote-sensing, i.e. the principal components of SST, SSH, and ocean color on sub-mesoscales. I’ll conclude with a brief discussion on future work to construct and use multi-modal foundation models for new studies on the physical ocean.

About Our Speaker

J. Xavier Prochaska
Distinguished Professor of Astronomy and Astrophysics, University of California, Santa Cruz, and Simons Foundation Pivot Fellow

As an observational astrophysicist, J. Xavier Prochaska studies the gas within, around and between distant galaxies. His research programs reveal the physical conditions of this gas: its enrichment by heavy elements, the ionization state, the surface density and its dynamics. With the Simons Foundation Pivot Fellowship, he is transitioning to the field of physical oceanography, working under the mentorship of Daniel Rudnick at the Scripps Institution of Oceanography (SIO) at the University of California, San Diego. Prochaska’s primary goals during the fellowship are to apply his expertise in spectroscopy, machine learning and instrumentation to research complex processes in sub-mesoscale dynamics, generate data-driven predictions for the onset of harmful algal blooms, and develop new technologies for in-situ hyperspectral imaging of coastal regions.