Reconstruction of sea surface temperature under clouds using masked autoencoders

This paper presents a methodology for reconstructing high-spatial-resolution sea surface temperature (SST) fields under cloud cover using masked autoencoders (MAE). The MAE model is trained on high-resolution SST maps from the ECCO forward simulation, LLC4320, and reconstructs missing data by masking out a portion of the input pixels. The impact of masking ratios and methods, as well as network architecture variations, is investigated. Preliminary results show that MAE can reconstruct global SST under a random 80% mask to within 0.3. C root mean squared error (RMSE). Applying this methodology to SST data with significant cloud contamination can enhance dataset quality, uncovering details hidden by clouds and expanding the use of high-resolution SST images.