Support vector machines for cloud detection over ice-snow areas

In polar regions, cloud and underlying ice-snow areas are difficult to distinguish in satellite images because of their high albedo in the visible band and low surface temperature of ice-snow areas in the infrared band. A cloud detection method over ice-snow covered areas in Antarctica is presented. On account of different texture features of cloud and ice-snow areas, five texture features are extracted based on GLCM. Nonlinear SVM is then used to obtain the optimal classification hyperplane from training data. The experiment results indicate that this algorithm performs well in cloud detection in Antarctica, especially for thin cirrus detection. Furthermore, when images are resampled to a quarter or 1/16 of the full size, cloud percentages are still at the same level, while the processing time decreases exponentially.