Detection of Daytime Arctic Clouds using MISR and MODIS Data

March, 2006
Report Number: 
Tao Shi, Eugene E. Clothiaux, Bin Yu, Amy J. Braverman, David N. Groff

Amongst the 36 spectral radiances available on the Moderate Resolution Imag- ing Spectroradiometer (MODIS) seven of them are used operationally for detection of clouds in daytime polar regions. While the information content of clouds inherent in spectral radiances is familiar, the information content of clouds contained in angular radiances (i.e., radiances emanating to space from the same object but in different di- rections) is not. The Multi-angle Imaging Spectroradiometer (MISR) measures angular radiances to space and its collocation on the NASA Terra satellite with MODIS allows for a comparative analysis of its cloud detection capabilities with those of MODIS. Expert labels from an extensive amount of data are used to compare arctic cloud detection efficiencies of several methods based on MISR radiances and radiance-based features, MODIS radiances and radiance-based features, and their combinations. The accuracy of cloud detections is evaluated relative to 2.685 million 1.1-km resolution expert labels applied to 3.946 million pixels with valid radiances from 32 scenes that contain both clear and cloudy pixels. FisherÕs quadratic discriminate analysis (QDA) with expert labels is applied to MISR radiances, MISR radiance-based features, MODIS radiances, and MODIS radiance-based features. The resulting classification accuracies are 87.51%, 88.45%, 96.43%, and 95.61%, respectively. The accuracies increase to 96.98% (96.71%) when QDA with expert labels is applied to combined radiances (fea- tures) from both MISR and MODIS. These results are indicative of the information content inherent in spectral and angular radiances, but these classifiers are impossible to obtain in practice due to their reliance on expert labels. A second group of classi- fiers, also QDA-based, used automatic training labels from thresholding on combined MISR and MODIS radiance-based features. Training the QDA classifier on the auto- matic labels using MISR radiances, MISR radiance-based features, MODIS radiances, and MODIS radiance-based features led to accuracies of 85.23%, 88.05%, 93.62%, and 93.55%, respectively. For combined radiances (features) from both MISR and MODIS accuracies are 93.74% (93.40%) for the 32 scenes. A scheme that combines training a QDA classsifier with MISR and MODIS automatic labels for the 32 mixed scenes and thresholding of MISR features for classification (with 95.39% accuracy) of an additional 25 pure clear or cloudy scenes produced an accuracy of 94.51% for the 57 scenes, the highest classification rate of any automated procedure that was tested in the study. The accuracy of the MODIS operational cloud mask is 90.72% for the 32 mixed scenes and 93.37% for the 25 pure scenes. Training a QDA classifier on the MODIS mask did not improve classification accuracy. These results suggest that both MISR and MODIS radiances have sufficient in- formation content for cloud detection in daytime polar regions. Together they have slightly more information than separately. The use of an automated, but adaptable, QDA classifier built on a combination of MISR and MODIS data improved classifica- tion accuracy to 94.5% relative to single-value threshold classifiers, based on either sensor separately, with accuracies of 92.0% over all 57 scenes in the study. Classifica- tion accuracy attained by the automated, adaptable QDA classifier is only 2Ð3% short of the best test accuracy achieved from expert training labels. These results imply that analysis of daytime polar cloud masks obtained from MISR and MODIS radiances over much larger spatial and temporal scales is a worthwhile endeavor.

PDF File: 
Postscript File: