AI based Out-Of-Distribution Analysis of Sea Surface Height Data

Abstract

We performed Out-Of-Distribution (OOD) analysis of 7.8 million Sea Surface Topography Merged Altimeter L4 cdr grid cutouts in an effort to identify rare (possibly unknown) physical phenomenon sea surface height (SSH) data. The algorithm used for the project is Ulmo which is a probabilistic autoencoder (PAE), originally developed for sea surface temperature data. A PAE is made of an autoencoder for taking the extracted images and encoding them into a latent representation of the data, and a normalizing flow which takes the encoding and maps it to a normal distribution for probabilistic interpretation. A Log-Likelihood (LL) value for each cutout was calculated from this normal distribution and we defined the images with the lowest 0.1 percentile of LL values as anomalies. Ulmo successfully identifies outliers and distinguishes the ocean's most dynamic regions being Western boundary currents.