Cloud fraction errors caused by finite resolution measurements

The standard method of deriving cloud fraction from space simply finds the fraction of total image pixels that contains some cloud. The sensitivity of this method to sensor resolution has been examined by Shenk and Salomonson [1972], assuming the cloudy pixels are detected perfectly. Their experiment is reexamined to show that the sensitivity is much more complex than they predicted. We derive the upper and lower bounds of the true cloud fraction, A(t), given the standard method estimate, A(e), to show that the range of possible values for A(t) can be, in general, very wide. By including the fraction of apparent cloud edge and cloud interior pixels, the bounds can be reduced and improvements to the standard method can be obtained. However, this improvement is also resolution limited by the misidentification of partially cloudy pixels as cloud interior rather than cloud edge. A potentially better technique for estimating the true cloud fraction is therefore explored using a pattern recognition approach. A nearest neighbor classification rule is used in two sets of experiments: one using 684 simulated cloud fields as a training set, the other using 370 cloud fields based on Advanced Very High Resolution Radiometer (AVHRR) measurements. Given the underlying distribution of A(t), A(e) overestimates A(t) with an overall average bias of 32% and standard error of 11% for the simulated training set, and a bias of 35% and a standard error of 3% for the AVHRR training set. The pattern recognition estimate, A(p), is essentially unbiased and has a standard error of 12% for both training sets. The relevance of these training sets to new scenes and the importance of imperfect cloud detection have yet to be investigated, but the pattern recognition technique shows considerable potential advantage over the standard technique in providing unbiased estimates of cloud fraction that are less sensitive to the effects of sensor resolution.