SSHprediction

Sea surface height (SSH) is a critical metric for understanding eddies and currents in the ocean. Successful future prediction of SSH using machine learning could serve a variety of purposes for oceanographers; for instance, accurate eddy predictions would be invaluable to oceanographers doing field research. Furthermore, machine learning can provide an understanding of decorrelation timescales in different parts of the ocean, providing important data for physical oceanographers from a theoretical standpoint. This study expands on the work done by Martin et. al (2023), which used deep learning to improve on optimal interpolation for interpolation of raw satellite data. We used the SimVP and ConvLSTM machine learning architectures to predict observations up to 30 days in the future, using raw satellite measurements of SSH and SST. The loss was determined by interpolating the gridded output to the location of the satellite data, and then using the mean square error of the interpolated data when compared to the ground truth satellite data. In some experiments, a linear weighting scheme was used in the loss function during training, in which day 1 was weighted three times as heavily as day 30; however, there was no observed difference in the model when this weighting scheme was used. Our results were compared to the interpolated data for the predicted time period and baseline metrics such as persistence. Future work could investigate the longevity of eddies observed in the future data, and the inclusion of more satellites that observe SST data. Latitude/longitude access for SST data: https://drive.google.com/drive/folders/1iktgnSyF4cNj9iOS0p5GUATZGexk0W20?usp=sharing