Polar cloud detection using satellite data with analysis and application of kernel learning algorithms

Clouds play a major role in controlling Earth's climate, and studying global cloud distribution is the first step to advance our understanding and improve our prediction of global climate changes (such as global warming). However, cloud detection is particularly challenging in polar regions because of the snow and ice coverage. Collaborating with Dr. Eugene Clothiaux and Dr. Amy Braverman, we propose new algorithms to detection polar clouds using data collected by the Multi-angle Imaging SpectroRadiometer (MISR) and Moderate Resolution Imaging Spectroradiometer (MODIS).; We devise an Enhanced Linear Correlation Matching Classification (ELCMC) algorithm to improve the MISR polar cloud detection based on the multi-angle information and the ELCMC algorithm provides a 92% average accuracy rate. The accuracy is further improved when the ELCMC results are used as labels to train nonlinear classifiers such as Quadratic Discriminate Analysis (QDA) or Gaussian kernel Support Vector Machines (SVM). Combining MISR data with the hyper-spectral information provided by MODIS, we propose an algorithm to improve the polar cloud detection. The highly accurate (97%) consensus pixels are used to train QDA on all MISR and MODIS features. Then the resulting classifier provides a 94% accuracy rate, higher than both the MISR rate (88%) and the MODIS rate (90%) over "partly cloudy" scenes.; Because SVMs provide better classification results but require much more computation than QDA in the cloud detection problem, we use binning to reduce the computation of Gaussian kernel regularization methods. In regression we show that binning keeps the same minimax rates of the unbinned estimator and reduces the computation from O(n 3) to O(m3), with n as the sample size and m as the number of bins. To achieve the minimax rate in the k-th order Sobolev space, m needs to be in the order of O(kn 1/(2k+1)), which makes the computation of order O(n) for k = 1 and even less for larger k. On a particular polar scene, the SVM trained on 966 bins has a comparable test classification rate as the SVM trained on 27,179 samples, but significantly reduces the training time from 5.99 hours to 2.56 minutes.