Landsat thematic mapper (TM) data are used to provide a very high spatial resolution source of cloud ''truth.'' TM atmospheric window channels at wavelengths of 0.83 and 11.5 μm are used to simulate the visible and infrared channels on meteorological satellites. The TM data are spatially averaged to provide spatial resolutions (that is, pixel sizes) ranging from the full resolution 28.5-m data to the 1-, 4-, and 8-km resolutions typical of AVHRR and GOES data. These data are then used to examine the sensitivity of satellite-derived estimates of cloud fractional coverage to changing sensor spatial resolution. Seven different cloud retrieval algorithms are examined, including the International Satellite Cloud Climatology Project (ISCCP), new cloud for Earth (NCLE) radiation budget, hybrid bispectral threshold method (HBTM), spatial coherence, box counting based on fractal theory, reflectance threshold, and temperature threshold methods. The analysis is carried out for 24 cloud fields of varying cloud type: cumulus, stratocumulus, altocumulus, cirrus, and multilayered cloud. Results indicate that estimates of cloud fraction vary greatly as a function of sensor spatial resolution and cloud algorithm assumptions. Even for 28.5-m spatial resolution data, current cloud algorithms give cloud fractions that vary by as much as 0.25.

In general, cloud algorithms are sensitive to either sensor resolution (threshold methods such as ISCCP or HBTM) or to assumptions about cloud optical depth (NCLE, spatial coherence). When present, the effect of sensor resolution is small for satellite pixel sizes less than 0.25 km. The effects are large for pixel sizes of 1 km or larger. Results are shown to be a strong function of cloud type. The effect of sensor resolution is strongest for boundary layer clouds and is very weak for cirrus clouds. At the 4- to 8-km spatial resolutions typical of meteorological satellite data, the ISCCP algorithm overestimates cloud fraction for boundary layer cloud by about 0.05 but underestimates thin cirrus by 0.05. The NCLE algorithm underestimates all cloud types by an average of 0.32, and spatial coherence underestimates boundary layer cloud fraction an average of 0.18. For boundary layer clouds the errors are traced to the assumption of cloud-filled pixels for ISCCP, the assumption of optically thick clouds for spatial coherence, and the assumption of a typical cloud albedo for the NCLE method. Using 8-km data, the HBTM and ISCCP methods provide the most accurate cloud fraction, although the HBTM method underestimates whin cirrus by 0.20 and ISCCP overestimates all cloud types but cirrus by about 0.05.

The 24 cloud fields examined were chosen to explore some of the more difficult cloud retrieval cases, so the results should not be extrapolated to global average conditions. Nevertheless, the results suggest a critical need for a clearer understanding of the performance of satellite-derived cloud properties. ¿ American Geophysical Union 1992