Support Vector Machines for Polar Cloud Detection

• Activity Number: 62
• Type: Invited
• Date/Time: Monday, August 9, 2004 : 8:30 AM to 10:20 AM
• Sponsor: ENAR
• Author(s): Tao Shi and Bin Yu*+ and Eugene Clothiaux and Amy Braverman
• Companies: University of California Berkeley and University of California, Berkeley and Pennsylvania State University and California Institute of Technology
• Keywords: cloud detection ; Support Vector Machine ; feature selection ; MISR

Abstract

Cloud detection is crucial to many problems in climate research, but is notoriously difficult in polar scenes where the background scene has similar spectral signatures to those of clouds. The Multi-angle Imaging SpectroRadiometer (MISR), launched in 1999 as part of NASA's Earth Observing System, provides the next generation of high-resolution datasets for climate studies including those related to clouds and their effects on the radiation budget of Earth. MISRs view the Earth at nine view angles and four spectral wavelengths, providing 36 dimensional data at 1.1 km resolution. This is the first instrument in Earth orbit to provide both multi-angle and multispectral information simultaneously. We apply support vector machines with careful feature selection that incorporates our physical understanding of the radiative properties of ice, snow, and clouds to this problem. We compare various methods to develop a practical strategy for obtaining training data to feed the SVM. We then apply the Gaussian kernel support vector machine to a representative area over Greenland and compare the results to an expert-labeled image.