Using deep supervised machine learning to explore and identify patterns within sea surface temperature data

Presented at the GHRSST XXIII international science team meeting, 27 June-1 July 2022, online and in-person (Barcelona). #GHRSST23

Short abstract 

Sea Surface Temperature (SST) datasets contain vital information on the Earth’s climate and on the ocean’s dynamical processes. This information is difficult to extract due to its complexity and the size of the associated datasets with current oceanographic methods failing to deal effectively with them. A new direction with great promise is the use of deep learning techniques applied to the datasets. Prochaska et al (2021) developed a machine-learning algorithm, known as ULMO, to discover complex or extreme events in the SST fields. Their focus shifted as they noticed that ULMO proved to work for much more, consistently identifying patterns, not only in the extremes but, within larger portions of the dataset as well. In the work discussed in this presentation, we applied the ULMO algorithm to the nighttime portion of the global Visible Infrared Imaging Radiometer Suite (VIIRS) SST dataset, ~2.4x105 granules requiring ~40 TB of storage. VIIRS is the operational version of MODIS and the next-generation visible and infrared sensor flown by NOAA. We compared the machine learning decomposition of VIIRS to that of MODIS and further analyzed the mesoscale patterns within images from the VIIRS dataset. On a global scale, we saw that ULMO was able to consistently recognize similar, but not identical, patterns within the VIIRS dataset and that the output of ULMO is an excellent discriminator over the whole dataset, again not just for extreme events.