ISCCP New Algorithm: Summary of Changes
(Highlights of differences between the C-series and D-series cloud products)
Description of Old Algorithm
The ISCCP cloud detection procedure is applied to each month of satellite data at eight times of day and consists of five steps:
- space contrast test (applied to individual IR images),
- time contrast test (three consecutive IR images at constant diurnal phase),
- cumulation of space/time statistics (both IR and VIS images),
- construction of clear-sky composites for both IR and VIS (once every 5 days at each diurnal phase and location),
- radiance threshold (both IR and VIS images).
The first test classifies as cloudy all pixels that are much colder (low IR radiance) than the warmest value in small spatial domains. It is a spatial contrast test because the warmest pixel is not classified as either clear or cloudy. The second test classifies as cloudy all pixels that have sharply lower IR radiances at the same location as compared with values one day earlier or later and classifies as clear all pixels that show little variation of IR radiance over one-day intervals. To avoid confusion with diurnal variations of surface temperatures, this test is done separately for each time of day. The results of these two tests are combined to label image pixels as clear only when they exhibit low variability in both space and time. The third step collects statistics on the variations of the IR and VIS radiances over larger spatial and temporal domains. These statistics are used in the fourth step, along with the results of the first two tests (number of clear pixels and average clear radiance), to estimate values of clear IR and VIS radiances for each location and diurnal phase, once every 5 days. In the final step, the original IR and VIS radiances in each pixel at each time are compared with the inferred clear values. If the observed radiances differ from the clear-sky values (lower IR or higher VIS) by more than the estimated uncertainty of the clear-sky values, they are classified as cloudy. A subset of these pixels, with radiance values close to the values dividing clear from cloudy, are referred to as marginally cloudy. All other pixels are classified as clear.
Once each pixel is classified as clear or cloudy, the measured radiances can be compared to radiative transfer model calculations that include the effects of the atmosphere, surface and clouds. The attributes of the atmosphere, surface and clouds are represented in the model by a large number of physical properties; but the availability of correlative datasets and restriction of the satellite radiances to two wavelengths limit the number of parameters that can be determined from the observations. The analysis strategy used exploits the correlative data to isolate the cloud effects and attributes all remaining radiance variation to changes in two cloud properties; other parameters are assigned climatological average values.
For VIS (0.6 microns) and IR (11 microns) wavelengths, atmospheric effects are small and depend on ozone and water abundances, the temperature profile and aerosol optical thickness. Complete information on ozone, water, and temperature can be obtained from correlative data. Since little information is available concerning aerosol properties, their effects are neglected; however, since most aerosol occurs near the surface1 and affects the upwelling radiation from there, our use of surface properties obtained from clear radiances with the same radiative model effectively incorporates the primary aerosol effects into these surface properties. The two surface properties retrieved from clear radiance values are the "visible" reflectance and the "brightness" temperature. The visible reflectance represents the amount of sunlight reflected by the surface at 0.6 microns and at a particular sun and satellite geometry (the view/illumination geometry is recorded in the data). Since most surfaces reflect sunlight anisotropically, the reflectance varies with satellite position, time of day, and season. The temperature values correctly describe the IR radiances at the surface but are not equal to the actual physical temperature since most surfaces have IR emissivities slightly less than one. Little information exists concerning emissivities for land and vegetated surfaces, so all temperatures are retrieved assuming an emissivity of one.
Clouds are represented in the model as a single, thin (i.e., isothermal) layer, uniformly covering the image pixel and composed of water droplets with a specified average size (10 microns) and size distribution. All variations of cloudy radiances are attributed to changes in "visible" optical thickness (defined at 0.6 microns) and a temperature. The optical thickness parameter determines the amount and angular distribution of sunlight reflected by the cloud layer (the full effects of multiple scattering are included), and the temperature is a "brightness" temperature, like that for the surface, that is interpreted to represent the physical temperature at the top of an opaque cloud layer. At night when only IR radiances are measured, no cloud optical thickness is reported and IR variations are associated with the cloud top "brightness" temperature. During the day, the cloud optical thickness value is related to an IR optical thickness in the model to correct for cloud emissivities less than one; both the original (opaque cloud value) and corrected cloud top temperatures are reported.
Other cloud properties may vary, including particle size, phase (water or ice), and size distribution. Moreover, vertical (sub-pixel-sized) and horizontal inhomogeneities, such as multiple layering or small scale "brokenness", can affect the radiances. If the role of these other cloud properties can be formulated and their value measured, then the retrieved optical thickness and temperature parameters in the ISCCP data set can be corrected by comparing the radiative model radiances calculated with and without the variation of these additional parameters to produce a transformation from one model to the other.
The basic detection and radiative analysis is performed for each image pixel 2. However, the cloud climatology data need to be reduced to a more manageable volume. This reduction requires that the spatial distribution of the pixel values be summarized in the Stage C1 data. This is done by projecting the pixel data into a standard map grid (resolution about 280 km) and calculating the average values of cloud and surface properties, as well as their standard deviations. To maintain approximately equal statistical significance, the map grid is an equal-area grid. In addition, the explicit distributions of cloud optical thicknesses and top pressures (obtained from the top temperature and the atmospheric temperature profile) are reported (see ISCCP Cloud Classification figure below). All of these statistics are first collected for each satellite separately.
The final C1 data sets (one for each 3-h interval in a month) are assembled by "merging" the results from all the satellites. To preserve the statistical character and uniformity of the C1 results, "merging" is done by selection; i.e., only one result is reported for each map grid cell and any other results are discarded. The choice between satellites is made on the basis of two criteria: 1) a strong preference for time records produced from a single satellite and 2) limiting the satellite zenith angle (which also affects the number of pixels available). At low latitudes, the data reported generally come from the nearest geostationary satellite; if the primary data are missing then they are replaced by an adjacent geostationary satellite (if the zenith angle is not too large) or a polar-orbiter (if available at that time). In the polar regions, coverage is provided solely by the polar-orbiters, with the "afternoon" satellite preferred if two are available.
For a more detailed technical discussion, see:
- Rossow, W.B., and R.A. Schiffer, 1991: ISCCP Cloud Data Products. Bull. Amer. Met. Soc., 72, 2-20.
- Rossow, W.B., A.W. Walker, and L.C. Garder, 1993: Comparison of ISCCP and Other Cloud Amounts. J. Clim., 6, 2394-2418.
Footnote 1: Except after a large volcanic eruption. Large dust storms are generally detected as "clouds" and their properties retrieved as those of a water cloud. (Back to text.)
Footnote 2: Each image pixel is about 4-8 km in size, but the Stage B3 data have been sampled to a pixel spacing of about 30 km. Because of navigation uncertainties, the pixel location is not more precise than about 30 km and is variable within this range. Hence, in the cloud analysis, each image pixel is treated as representing a specific scene about 30 km across, referred to as a B3 data pixel. (Back to text.)
Description of Revised Algorithm
- Revised VIS and IR calibrations to eliminate spurious changes between different reference polar orbiters (Brest et al. 1996)
- Revised normalizations of geostationary satellite calibrations to eliminate occasional short-term deviations (Brest et al. 1996)
- Improved cirrus detection over land by lowering IR threshold from 6K to 4K
- Improved polar cloud detection over ice and snow surfaces by lowering VIS threshold from 0.12 to 0.06 and by using threshold test on 3.7 micron radiances
- Improved detection of low clouds at high latitudes by changing to VIS reflectance threshold test
- Improved treatment of cold (top temperature < 260 K) clouds by using ice polycrystal scattering phase function to retrieve optical thickness and top temperature
- Improved retrieval of cloud optical thicknesses over ice and snow surfaces using 3.7 micron radiances
- Improved retrieval of cloud top temperatures by including effects of IR scattering
- Improved retrieval of surface and cloud top temperatures by adopting new treatment of water vapor continuum absorption in IR
Gridded Product Contents
- Better resolved variations of optically thicker clouds by adding 6th optical thickness category
- Added correct cloud water path parameter
- Reported actual average values of cloud top temperature, pressure, optical thickness and water path for each of nine cloud types defined by cloud top pressure and optical thickness in the 3-hourly dataset
- Reported separate cloud properties for liquid and ice forms of low and middle-level clouds
- Provided conversion of cloud top pressures to cloud top heights above mean sea level based on atmospheric temperature profile
- Added cloud amount frequency distribution to monthly dataset
- Archived pixel-level cloud products with resolution of 30km and 3 hr
For a more detailed technical discussion, see:
- Rossow, W.B., and Schiffer, R.A., 1999: Advances in Understanding Clouds from ISCCP. Bull. Amer. Meteor. Soc., 80. Download document (370 Kbytes PDF), 27 pp.