Dylan Epstein-Gross ’25
Deep Learning for Interpolation of Sea Surface Temperature
Certificate(s): Statistics and Machine Learning
My project focused on developing deep learning models to interpolate missing values of sea surface temperature (SST) in gridded satellite data. SST is an extremely important variable for climate models and a key input for predicting ocean currents, but current high-resolution observations are incomplete due to cloud cover. To improve these valuable SST maps, I filled in the clouded data using state-of-the-art data processing pipelines and neural networks. I developed, trained and tested several neural network architectures in the lab and compared them with the deterministic statistical method used for interpolation in the literature. I found that certain models that process both spatial and temporal features of data were more effective than the baseline in reducing the deviation of interpolated values from actual temperature measurements. In the process, I researched many different neural network designs and learned a lot about the fascinating complexities of spatiotemporal interpolation as it relates to climate science. This research experience confirmed my passion for machine learning and inspired me to pursue it further during my time at Princeton, especially in conjunction with real-world problems like those found in oceanography.
Oceans and Atmosphere
School of Oceanography, University of Washington - Seattle, Washington
Georgy Manucharyan, Assistant Professor, School of Oceanography, University of Washington