The International Satellite Cloud Climatology Project (ISCCP) data set of cloud amounts and other products [Rossow and Schiffer, 1999] is a more than 20 year archive of daily global observations. Recently, this record has been used to study long-term trends in surface solar radiation [Pinker et al., 2005], with studies concluding that changes in cloudiness seen in the ISCCP record are evidence for a widespread increase in surface solar heating [Hatzianastassiou et al., 2005] and a decrease in planetary albedo [Pallé et al., 2004; Pallé et al., 2005], having implications for global climate models [Pallé et al., 2006]. Still others have used the ISCCP multi-decadal trends to suggest that long-term changes in cloudiness are causing widespread changes in outgoing longwave radiation [Cess and Udelhofen, 2003], are evidence of recent global brightening [Wild et al., 2005], and result from feedbacks associated with global warming [Ding et al., 2004]. However, these trends in total cloudiness have not been observed in surface [Norris, 2005] and other satellite [Jacobowitz et al., 2003; Wylie et al., 2005] cloud records. While a lack of corroboration with other data sets does not imply a deficiency in the ISCCP data, it has been suggested that the ISCCP cloud amounts may be affected by satellite related artifacts [Campbell, 2004; Norris, 2000].
 The ISCCP data set utilizes radiance information, with a nominal resolution of 8km at nadir, from a series of geostationary satellites to create 3-hourly maps of cloudiness and other associated products. The spatial coverage of the data is extended by utilizing Advanced Very High Resolution Radiometer (AVHRR) data from polar orbiting satellite roughly at latitudes above and below 60N and 60S, respectively. The exception to this is a segment of the Indian Ocean where geostationary satellite coverage did not exist until the late 1990s and data from the AVHRR was employed. We explore the long term variability of the ISCCP data using the D2 monthly mean cloud product for the years of 1983–2006 [Rossow and Schiffer, 1999], the most recent release (data from the British Atmospheric Data Centre, http://badc.nerc.ac.uk, and the Langley Research Center EOS DAAC, http://eosweb.larc.nasa.gov/). While we only consider the infrared total clouds product from this data set, which uses radiance measurements at 11 μm, this field is directly tied to the other cloud and clear-sky products available from ISCCP [Rossow and Schiffer, 1999].
 In order to minimize the effect of ENSO on the analysis, we regressed out the monthly Nino 3.4 index from each 2.5-degree grid cell (index data from the Climate Prediction Center at http://www.cpc.noaa.gov). In order to analyze the low frequency variability of the ISCCP data we created a low pass filtered data set by applying a 13-month boxcar filter to the time series at each cell in order to remove any intra-annual signal from the time series. In addition, we created a deseasonalized data set for analysis by removing the monthly mean from each month's data at every grid cell. Mean time series of the unprocessed, low frequency, and deseasonalized data was made by averaging the area weighted pixels from 60S–60N. This latitude range was chosen in order to emphasize the contribution to the global cloudiness time series from geostationary satellites and not that from the polar orbiting platforms. The dotted line in Figure 1 is the mean time series of the unprocessed data, the thin solid line is the deseasonalized series, and the thick line is the mean series of the low pass filtered cloudiness. All three time series show low frequency variability that is characterized by a 1% increase in clouds from the beginning of the record until about 1987, followed by a 4% decrease in cloud amounts for the next 13 years, and again a 1% increase from late 2000 until the end of the record.
 To determine the regions that contribute the most to the interannual patterns of cloudiness from Figure 1, we regressed the ISCCP low pass filtered mean time series (thick line in Figure 1) back onto the low pass filtered ISCCP cloud data. We used this time series in order to minimize the effect of any intra-annual variability on the resultant regression maps. Figure 2 is a map of the coefficients from this linear regression; regions of high values signify areas that contribute the greatest to the low frequency variability in the ISCCP time series (Figure 1), and regions with low or negative values contribute the least or negatively to that same mean pattern. The black contour bounds areas where the resultant time series from the regression can explain enough variance in the low pass filtered data to be statistically significant at the 0.95% level based on a two-tailed t-test and 22 degrees of freedom. The most striking features in Figure 2 are the circular patterns centered in the Atlantic, Western Pacific and Eastern Pacific Oceans. These circles correspond to so-called geostationary “footprints” that describe the area observed by each satellite. At the center of these footprints a satellite sensor's viewing angle is perpendicular to the surface, corresponding to a satellite zenith angle of 0 degrees. At the footprint edges the satellite zenith angle is much higher, corresponding to a longer path length through the atmosphere that light must travel before it is detected by the sensor. The band of negative regression coefficients over the Indian Ocean corresponds to the area where geostationary data was not available until 1997, and AVHRR polar orbiting satellite data was used in the record.
 The patterns in Figure 2 reflect different satellite viewing geometries and how those angles alter the record of cloudiness on a long term basis, a so-called ‘limb darkening’ effect, which is well known [Joyce et al., 2001; Minnis, 1989]. Furthermore, as the equatorial edges of the geostationary footprints happen to fall over oceanic regions, the low frequency ISCCP time series in Figure 1 (solid black line) better represents over-water than over-land cloud patterns. This distinction in the ISCCP long-term trends between oceanic and land covered regions has been noted in other studies [Norris, 2005].
 We propose that if the ISCCP low frequency time series was caused by a large-scale signal that is simply enhanced at the edges of the satellite view, the regression coefficients and associated time series would be more spatially uniform across those footprints. To test this theory we separated the deseasonalized ISCCP data into regions with high and low regression coefficients from the map in Figure 2. Here we use the deseasonalized data since the low pass filtered one is too much smoothed to reveal sudden changes in cloudiness, and the unprocessed record masks those changes in the seasonality. The cutoff for this distinction was one standard deviation above the mean regression coefficient value (a value of roughly 2). The areas with high regression coefficients comprised about 20% of the global surface between 60S–60N, and is bounded by the gray contour in Figure 2. We then created two area weighted mean time series of cloud amounts using this deseasonalized data averaged over 60S–60N for the areas with regression coefficients above and below our cutoff. The resultant time series correspond to areas with high regression coefficients (which we refer to as the HR series), and low regression coefficients (the LR series).
 The black dashed line in Figure 3 is the LR series, and the black solid line is the HR series. There is an almost 20% reduction in cloud amount for the HR time series from the beginning of the record through mid 2001. This is very different from the LR series, which shows a small amount of long term variability; a decrease in clouds from the late 1980s to the early 1990s. This separation is remarkable considering we choose a somewhat arbitrary dividing line between the LR and HR regions. The HR series also shows several abrupt increases (1984–1985, 1989–1990, 2000–2002) and decreases (1983–1984, 1987–1989, 1992–1993, 1995–1996, 1998–1999) in cloud amounts, and many periods of little to no interannual variability (1985–1987, 1993–1995, 1996–1998, 2002–2005).
 These observations would be consistent with the theory that changes in the number of geostationary satellites are altering the global mean cloud amounts [Campbell, 2004]; as the number of observing satellites increase, as was the case from the mid 1980s through the 1990s, more regions that had previously been at the limbs of a satellite's field of view would be closer to nadir for a newer imager. However, the abrupt changes in global cloud amounts that result from sudden changes in the geostationary viewing platforms would be somewhat hidden in the annual cycle, leading to what would appear to be a steady downward trend in global cloudiness. The increase in cloud amounts around 1985 and again 1990 could result from failures or repositioning of some operational instruments. Also, the increase in cloudiness during 2001 and 2002 may be physical, but it also corresponds to the switch to the new generation of AVHRR satellites (KLM series). The impact of this change in the reference AVHRR satellite on the ISCCP products are currently under investigation by the ISCCP processing team (B. Rossow, personal communication, 2006).
 To explore how changes in satellite geometries may be altering cloud amounts in the ISCCP data we created a monthly time series of 1/mu, the inverse of the cosine of the satellite zenith angle (ISCCP D1 data from the British Atmospheric Data Centre, http://badc.nerc.ac.uk, and the Langley Research Center EOS DAAC, http://eosweb.larc.nasa.gov/). The satellite zenith angle describes the angle between the local zenith and the line of sight to the satellite, assuming a non-curved surface. The cosine of the satellite zenith angle (mu) varies between 1 (at nadir) and 0 (looking out tangentially from the satellite). 1/mu is therefore a relative measure of the distance a photon must travel from a plane surface to be detected by the satellite sensor. A value of 2 for 1/mu corresponds to a path length that is double than that at nadir.
 We then separated the 1/mu data by regions corresponding to high and low regression coefficients from the map in Figure 2, similar to what was done for the deseasonalized cloud data. The solid gray line in Figure 3 is a time series of 1/mu that is area weighted, averaged over the HR regions, and smoothed using 2 recursive 5-month boxcar filters. The sudden drops in 1/mu likely result from the introduction of new geostationary satellites into the ISCCP data, or repositioning of existing satellites to fill in gaps in the global coverage. Table 1 lists a few key changes in the ISCCP satellite constellation that may have resulted in the pattern of 1/mu for the solid gray line in Figure 3, including the introduction, failure, or movement of various viewing platforms. This list is not exhaustive and only highlights a few changes in the ISCCP observational platforms.
|Event||Approximate Date||Longitude||Effect on 1/mu and Retrieved Cloud Amounts|
|Launch of GOES 7||Feb-87||75 West||Decrease|
|Failure of GOES 6||Jul-89||136 West||Increase|
|Repositioning of METEOSAT 3||Aug-91||75 West||Decrease|
|Launch of GOES 8||Apr-94||75 West||Decrease|
|Repositioning of METEOSAT 5||Jul-98||63 East||Decrease|
|AVHRR KLM Calibration Error||Jun-01||Global||Increase|
 The correlation between the HR and 1/mu time series is remarkable. Clearly there is a deterministic relationship between cloudiness at the tropical limbs of the geostationary satellites, and the satellite zenith angle, which is supported by the theory that limb darkening causes observations of increase cloudiness at the edges (especially the tropical ones) of the geostationary satellites [Joyce et al., 2001; Minnis, 1989]. It is also evident that were this not the case, the ISCCP data would exhibit no significant long term upward (during the early 80s) and downward (over much of the 1980s and 1990s) trends in cloudiness. If artifacts at the equatorial limbs did not exist, it is likely that the low frequency variability in mean global cloud amounts would be more like that in the LR time series (black dashed line, Figure 3), which is an average of cloudiness over 80% of the global surface between 60S–60N. Again, the LR series will show some variability that is similar to the HR series since we choose a somewhat arbitrary dividing line between the two regions, but this choice is sufficient for our illustrative purposes.
 To contrast the time series of 1/mu for the regions of high regression coefficients, we also created a time series of 1/mu (also area weighted and 5-month recursively boxcar filtered) for the LR regions as specified in Figure 2. The gray dashed line in Figure 3 is the LR time series of 1/mu, and shows much less variance than that for the HR values of 1/mu. This is not surprising since 1/mu is more sensitive to change at larger satellite zenith angles.
 To again emphasize this effect of viewing geometry on the long term trends from ISCCP, we have also created two time series of cloud amounts for a region of the equatorial Indian Ocean where the regression coefficients are strongly positive (20S–20N and 55W–65W) and negative (20S–20N and 65W–75W). Until a geostationary satellite was moved to over this area in the late 1990s, the region that is further west corresponded to an area that was at the edges of a satellite footprint, and the region further east was one covered by AVHRR data. Figure 4 is a time series of those regions, which have been smoothed using two recursive 13 month boxcar filters. The black line represents the area further west and shows roughly a 20% drop in cloud amounts when the new geostationary instrument was introduced. The gray line is of the region that used the AVHRR data for the first 15 years of the record, and shows a steady time series.
 While polar orbiting platforms also suffer from artifacts due to viewing geometry, these are averaged out over several days as these instruments view the same location with the same satellite geometry roughly once a week. In contrast, the effect on the geostationary platforms is cumulative, and therefore can lead to a completely non-physical signal, like that of the black line in Figure 4. This cumulative effect of the viewing geometry likely explains why none of the regions in Figure 2 that are covered by the AVHRR data are well characterized by the low frequency ISCCP series in Figure 1, despite the fact that the same processing algorithm is applied to both the geostationary and polar orbiting instruments.
 Although the ISCCP data is very appropriate for many applications, clearly its use in global multi-decadal studies is troubling. Especially as we find similar results, to varying degrees, by repeating this analysis for ISCCP high, mid, and low cloud amounts. However, it may be possible that new methodologies can be developed to process the ISCCP data and account for these types of artificial variations in cloudiness. One simple post-processing method would be to simply remove those areas at the tropical geostationary limbs from the calculation of global mean cloud amounts. Another may be to regress out the time series of 1/mu from the cloud data. However, clearly the best way to deal with the effects of the satellite viewing geometry on global cloud amounts would at the data processing stage, especially since the ‘limb darkening’ effect is non-linear and therefore a regression technique will not be able to completely account for the effect of changing viewing geometry on the cloud data.
 We have demonstrated that the long-term global trends in cloudiness from the ISCCP record are influenced by artifacts associated with satellite viewing geometry. Results from earlier studies based on these trends may be influenced by these non-physical artifacts, and we therefore suggest that development of a correction for the data is warranted. As the number of publications on the subject of climate change continues to grow [Stanhill, 2001], this paper highlights the need to critically explore the source of any trends in global, multi-decadal satellite data sets.