Abstract
This study investigates the use of a masked autoencoder (MAE) to address the challenge of filling gaps in high-resolution (1 km) sea surface temperature (SST) fields caused by cloud cover, which often result in gaps in the SST data and/or blurry imagery in blended SST products. Our study demonstrates that MAE, a deep learning model, can efficiently learn the anisotropic nature of small-scale ocean fronts from numerical simulations and reconstruct the artificially masked SST images. The MAE model is trained and evaluated on synthetic SST fields and tested on real satellite SST data from the Visible Infrared Imaging Radiometer Suite (VIIRS) sensor on the Suomi NPP satellite. We demonstrate that the MAE model trained on numerical simulations can provide a computationally efficient alternative for filling gaps in satellite SST. MAE can reconstruct randomly occluded images with a root mean square error (RMSE) of under 0.2 °C for masking ratios of up to 80 %. A trained MAE model in inference mode is exceptionally efficient, requiring 3 orders of magnitude (approximately 5000×) less time compared to the conventional approaches of cubic radial basis interpolation and Kriging tested on a single CPU. The ability to reconstruct high-resolution SST fields under cloud cover has important implications for understanding and predicting global and regional climates and detecting small-scale SST fronts that play a crucial role in the exchange of heat, carbon, and nutrients between the ocean surface and deeper layers. Our findings highlight the potential of deep learning models such as MAE to improve the accuracy and resolution of SST data at kilometer scales. This presents a promising avenue for future research in the field of small-scale ocean remote sensing analyses.