Deep Learning of Sea Surface Temperature Patterns to Identify Ocean Extremes

We perform an out-of-distribution analysis of ~12,000,000 semi-independent 128x128 pixel^2 SST regions, which we define as cutouts, from all nighttime granules in the MODIS R2019 Level-2 public dataset to discover most complex or extreme phenomena at ocean surface. Our algorithm (Ulmo) is a probabilistic autoencoder, combines two deep learning modules: (1) trained on ~150,000 random cutouts 2010, represent any input cutout with 512-dimensional latent vector akin (non-linear) EOF analysis; and (2) normalizing flow, maps autoencoder's space distribution onto isotropic Gaussian manifold. From latter, calculate log-likelihood value for each outlier be those lowest 0.1% distribution. These exhibit large gradients patterns characteristic highly dynamic surface, many are located within larger complexes whose unique dynamics warrant future analysis. Without guidance, Ulmo consistently locates outliers where major western boundary currents separate continental margin. Buoyed by these results, begin process exploring fundamental learned Ulmo, identifying several compelling examples. Future work may find that algorithms like hold significant potential/promise learn derive other, not-yet-identified behaviors archives satellite-derived fields. As important, see no impediment applying them other large, remote-sensing datasets science (e.g., sea surface height, color).