Recovering missing data in satellite images. An application to adriatic sst and comparison with in situ data

Satellite images are very useful for many applications in oceanography and other environmental sciences. They offer a great coverage both in time and space, not attained by in situ measurements. Clouds are responsible for missing data on images provided by receptors working in the visible and IR range receptors. In some seasons the cloud coverage can reach an important percentage. Many data analysis techniques do not need a total coverage, although it is always desirable. Some applications, such as Empirical Orthogonal Function (EOF) analysis, or wavelet decomposition need a complete set of data, and a technique for recovering these missing data is indispensable. In this work we present DINEOF (Data INterpolating Empirical Orthogonal Functions), a method for the reconstruction of satellite data, based on an EOF decomposition. DINEOF reconstructs the missing data from an optimal set of EOFs. The optimal number of EOFs is determined by cross-validation. This method has shown to obtain robust results. DINEOF has been applied to a series of 105 AVHRR SST images of the Adriatic Sea, in a period ranging from May to October 1995. The mean cloud coverage of this data set is 52%. The error obtained by the cross-validation is of 0.6°C, and a total of 10 EOFs were necessary to reconstruct the data. A comparison with in situ data obtained form the MEDAR/Medatlas database is made. A total of 452 stations are examined. The RMS error between MEDAR/Medatlas and the reconstructed data is of 0.95°C. The error between MEDAR/Medatlas data and the points that are not missing in the Adriatic data set is of 0.67°C, which can be considered as the inherent error between the in situ and remote sensed data sets.