I thought I would use a different procedure for evaluating cloud datasets: EOF analysis. My hope is that this will provide insight into what sort of problems are present in the data.

I began by calculating monthly anomalies in ISCCP total cloud at each grid point. After doing so, I removed cloud variability related to El Nino via linear regression on a time series of Nino3.4 SST. Although not necessary, the removal of El Nino cloud variability makes the subsequent plots more clear since some El Nino cloud variability happens to project onto apparently spurious artifacts. I then normalized the time series at each grid point so that I could carry out the EOF analysis on the correlation rather than the covariance matrix. I think the correlation matrix will be better for finding artifacts since they seem to have similar relative effects irrespective of absolute cloud amounts. My next step was weighting the monthly anomalies by the square root of the cosine of latitude (this is the appropriate grid box area weighting for EOF analysis). After calculating the EOFs, I applied varimax rotation to the first ten in order to bring more clarity to the EOF patterns. Rotation is not necessary, but without it the patterns end up being somewhat more "mixed."

I've attached time series and spatial patterns for the first five rotated EOFs (see related tar file link to left). The time series and spatial patterns are scaled to have unit variance. Note that REOFs 1, 4, and 5 are clearly artificial with large trends or jumps in the time series and obvious satellite features in the spatial patterns. My hope is that this information may be useful for tracking down causes of artifacts (i.e., satellite changes, etc.). REOFs 2 and 3 appear to be natural tropical variability that was not fully removed by linear regression on El Nino or that happens at shorter time scales.