A method to reconstruct missing data in sea surface temperature data using a neural network is presented. Satellite observations working in the optical and infrared bands are affected by clouds, which obscure part of the ocean underneath. In this paper, a neural network with the structure of a convolutional auto-encoder is developed to reconstruct the missing data based on the available cloud-free pixels in satellite images. Contrary to standard image reconstruction with neural networks, this application requires a method to handle missing data (or data with variable accuracy) in the training phase. The present work shows a consistent approach which uses the satellite data and its expected error variance as input and provides the reconstructed field along with its expected error variance as output. The neural network is trained by maximizing the likelihood of the observed value. The approach, called DINCAE (Data INterpolating Convolutional Auto-Encoder), is applied to a 25-year time series of Advanced Very High Resolution Radiometer (AVHRR) sea surface temperature data and compared to DINEOF (Data INterpolating Empirical Orthogonal Functions), a commonly used method to reconstruct missing data based on an EOF (empirical orthogonal function) decomposition. The reconstruction error of both approaches is computed using cross-validation and in situ observations from the World Ocean Database. DINCAE results have lower error while showing higher variability than the DINEOF reconstruction.

The surface circulation of the Caribbean Sea and Gulf of Mexico is studied using 13 years of satellite altimetry data. Variability in the Caribbean Sea is evident over several time scales. At the annual scale, sea surface height (SSH) varies mainly by a seasonal steric effect. Interannually, a longer cycle affects the SSH slope across the current and hence the intensity of the Caribbean Current. This cycle is found to be related to changes in the wind intensity, the wind stress curl, and El Niño–Southern Oscillation. At shorter time scales, eddies and meanders are observed in the Caribbean Current, and their propagation speed is explained by baroclinic instabilities under the combined effect of vertical shear and the β effect. Then the Loop Current (LC) is considered, focusing on the anticyclonic eddies shed by it and the intrusion of the LC into the Gulf of Mexico through time. Twelve of the 21 anticyclonic eddies observed to detach from the LC are shed from July to September, suggesting a seasonality in the timing of these events. Also, a relation is found between the intrusion of the LC into the Gulf of Mexico and the size of the eddies shed from it: larger intrusions trigger smaller eddies. A series of extreme LC intrusions into the Gulf of Mexico, when the LC is observed as far as 92°W, are described. The analyses herein suggest that the frequency of such events has increased in recent years, with only one event occurring in 1993 versus three from 2002 to 2006. Transport through the Straits of Florida appears to decrease during these extreme intrusions.

We present an extension to the Data INterpolating Empirical Orthogonal Functions (DINEOF) technique which allows not only to fill in clouded images but also to provide an estimation of the error covariance of the reconstruction. This additional information is obtained by an analogy with optimal interpolation. It is shown that the error fields can be obtained with a clever rearrangement of calculations at a cost comparable to that of the interpolation itself. The method is presented on the reconstruction of sea-surface temperature in the Ligurian Sea and around the Corsican Island (Mediterranean Sea), including the calculation of inter-annual variability of average surface values and their expected errors. The application shows that the error fields are not only able to reflect the data-coverage structure but also the covariances of the physical fields.