Cloud Detection over Snow and Ice Using MISR Data
Clouds play a major role in Earth's climate and cloud detection is a crucial step in the processing of satellite observations in support of radiation budget, numerical weather prediction and global climate model studies. To advance the observational capabilities of detecting clouds and retrieving their cloud-top altitudes, NASA launched the Multi-angle Imaging SpectroRadiometer (MISR) in 1999, which provides data in nine different views of the same scene using four spectral channels. Cloud detection is particularly difficult in the snow- and ice-covered polar regions and availability of the novel MISR angle-dependent radiances motivates the current study on cloud detection using statistical methods.
Three schemes using MISR data for polar cloud detection are investigated in this study. Using domain knowledge, three physical features are developed for detecting clouds in daylight polar regions. The features measure the correlations between MISR angle-dependent radiances, the smoothness of the reflecting surfaces, and the amount of forward scattering of radiances. The three features are the basis of the the first scheme, called Enhanced Linear Correlation Matching Classification (ELCMC). The ELCMC algorithm thresholds on three features and the thresholds are either fixed or found through the EM algorithm based on a mixture of two 1-dim Gaussians. The ELCMC algorithm results are subsequently used as training data in the development of two additional schemes, one Fisher's Quadratic Discriminate Analysis (ELCMC-QDA) and the other a Gaussian kernel Support Vector Machine(ELCMC-SVM). For both QDA- and SVM-based experiments two types of inputs are tested, the set of three physical features and the red radiances of the nine MISR cameras. All three schemes are applied to two polar regions where expert labels show that the MISR operational cloud detection algorithm does not work well, with a 53% misclassification rate in one region and a 92% nonretrieval rate in the other region.
The ELCMC algorithm produces misclassification rates of 6.05% and 6.28% relative to expert labelled regions across the two polar scenes. The misclassification rates are reduced to approximately 4% by ELMCM-QDA and ELCMC-SVM in one region and approximately 2% in the other. Overall, all three schemes provided significantly more accurate results and greater spatial coverage than the MISR operational stereo-based cloud detection algorithm. Compared with ELCMC-QDA, ELCMC-SVM is more robust against mislabels in the ELCMC results and provide slightly better results, but it is computationally slower.