Outlier detection in satellite data using spatial coherence

Satellite data sets often contain outliers (i.e., anomalous values with respect to the surrounding pixels), mostly
due to undetected clouds and rain or to atmospheric and land contamination. A methodology to detect outliers
in satellite data sets is presented. The approach uses a truncated Empirical Orthogonal Function (EOF) basis.
The information rejected by this EOF basis is used to identify suspect data. A proximity test and a local median
test are also performed, and a weighted sum of these three tests is used to accurately detect outliers in a data
set. Most satellite data undergo automated quality-check analyses. The approach presented exploits the spatial
coherence of the geophysical fields, therefore detecting outliers that would otherwise pass such checks.
The methodology is applied to infrared sea surface temperature (SST), microwave SST and chlorophyll-a concentration data over different domains, to show the applicability of the technique to a range of variables and
temporal and spatial scales. A series of sensitivity tests and validation with independent data are also
conducted.