The Impact of Thresholds in Cloud Detection Uncertainty

Abstract:

The Moderate Resolution Imaging Spectroradiometer (MODIS) on the NASA Terra and Aqua satellites provides cloud properties with daily, global coverage from its broad spectral range (36 bands between 0.415-14.235 micron) at high spatial resolution (250 m for two bands, 500 m for 5 bands, 1000 m for 29 bands). These cloud properties are aggregated to produce climatologies and histograms that are used widely in observational studies, modeling applications, and further data production. Providing a measure of the cloud detection uncertainty is essential to its quality and proper use.

 

The MODIS cloud mask algorithm includes several domains defined according to latitude, surface type, and solar illumination, including land, water, snow/ice, desert, and coast for both day and night. A series of spectral tests are applied to identify the presence of clouds. There are several groups of tests, with differing numbers of tests in each group depending on the domain. A clear-sky confidence level ranging from 1 (high) to 0 (low) is returned for each test. The minimum confidence from all tests within a group is taken to be representative of that group. The Nth root of the product of all the group confidences (Q) determines the final confidence, where N is the number of groups. 

 

Our objective is to quantify the sensitivity of the MODIS cloud detection to various factors included in the cloud mask algorithm– specifically surface type, seasonality and viewing geometry, as well as, traditional identified uncertainties like cloud cover heterogeneity. The challenge is that two or more of these factors frequently interact, producing combined uncertainties in the retrievals that cannot be quantified by calculating retrieval sensitivities to each factor separately. This presentation will present an assessment of uncertainty sources within the MODIS cloud mask products due to uncertainties in the cloud thresholding approach used in the algorithm. This is accomplished through statistical comparison of cloud detection results of many realizations of the same scenes with appropriate modifications to the various thresholds.