UC Santa Cruz

This thesis presents a new algorithm to mitigate cloud masking in the analysis of sea surface temperature (SST) data generated by remote sensing technologies such as infrared sensor satellites like the Level-2 Visible-Infrared Imager-Radiometer Suite (VIIRS). Cloud coverage interferes with the analysis of all remote sensing data using wavelengths shorter than ≈ 2 microns, significantly limiting the quantity of usable data and creating a biased geographical distribution towards equatorial and coastal regions. Prior studies have led to use of in-painting algorithms like Navier-Stokes but was typically only used up to 5% masking and had limited success. To address this issue, we propose an unsupervised machine learning algorithm called ENKI which uses a Vision Transformer with Masked Autoencoding to reconstruct pixels that are masked out by clouds. We train four different models of ENKI with training mask ratios (referred to as t) set to 10%, 35%, 50%, and 75% on a generated Ocean General Circulation Model (OGCM) dataset known as LLC4320. To evaluate performance we reconstruct LLC 4320 SST images at a patch masking ratio of 10%, 20%, 30%, 40%, 50% (referred to as p) and examine reconstruction qualitatively and statistically by calculating the root means squared error (RMSE) of reconstructed patches. Through our analysis we discover that edge patches contain a higher error rate and that a bias appears in some models when reconstructing images at p masking ratios away from their training mask ratio t. But we consistently find that at all levels of p masking ratios there is one or multiple models that create reconstructions with a mean RMSE of less than ≈ 0.03K which is lower than the estimated sensor error of VIIRS data which is ≈ 0.078 K for daytime, along scan, and ≈ 0.05 K for nighttime, along-scan. We also conclude the complexity of dynamics within an image and the p masking ratio affect RMSE with higher complexity and p masking seeing higher RMSE values. Critically, we also discover at a patch level that despite RMSE having some correlation to complexity, they are not directly proportional, and RMSE increases at a slower rate as complexity within a patch increases. Our analysis concludes that ENKI shows great promise in surpassing in-painting as a means of reconstructing cloud masking, and future research seeks to analyze ENKI’s capabilities in reconstructing real world data.