Speaker

J. Xavier Prochaska, professor, UC Santa Cruz

Lessons Learned from Applying Machine Learning to Remote Sensing Observations of Sea Surface Temperature

Abstract

I will describe lessons learned from our team’s program to apply machine learning to large remote sensing datasets of sea surface temperature (SST) on submesoscales (i.e. ~10-100km).  Adopting unsupervised techniques, we first identified outliers in MODIS SST imagery on these scales and associated them with the most dynamic regions of the global ocean (e.g. western boundary currents). We next developed a “language’’ (a.k.a., data manifold) to compactly describe the great diversity of patterns and structure in SST at the submesoscale.  This enables one to search for and analyze specific dynamical signatures in SST across the globe and over the past two decades.  Most recently, we used a Large Language-inspired model trained on ECCO outputs to successfully predict SST under masked (e.g., cloudy) data.  I will conclude by highlighting ongoing work for the retrieval of inherent optical properties from hyperspectral images in anticipation of forthcoming NASA/PACE datasets.