Evaluating the Performance of SVM in Dust Aerosol Discrimination and Testing its Ability in an Extended Area

The Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observation (CALIPSO) has been run more than five years, and one of its payloads-the Cloud-Aerosol Lidar with Orthogonal Polarization (CALIOP)-can offer global atmospheric vertical profiles, resulting in it being widely used in aerosol research. Because the Lidar ratio is an important parameter for aerosol parameter retrieval and the choice of the Lidar ratio needs confirm the type of layers (aerosol or cloud), the data process cannot do without accurate and efficient aerosol and cloud classification. Initially, we found the classification errors existed in PDF version 1 data, so we introduce the SVM classifier to improve the accurate of thick dust and cloud classification. Despite NASA issued the PDF version 2 and improve the correct rate greatly, but we cannot ignore SVM advantages. Therefore in this paper, to continue the former study, we will not only validate the feasibility of this method (especially in dust source areas), but also make a thorough study of classification result effected by different feature vectors and samples. Through the accuracy testing we found, with the increase of the number of samples, SVM have the better result, and the feature space which include depolarization ratio have the more stable result. Though we validated the advantage of SVM, but if we want to use this algorithm in global scale or the whole process of dust deposition, there are still some improvements we need to do. To obtain the thinking of algorithm revision, we using MODIS, HYSPLIT, and CALIPSO products in a dust storm and observe its transmission, then the deficiencies of SVM can be found when compare with PDF version 2. Indeed PDF considered the regional variation, with this transmission of dust storm we need adjustment training samples and hyperplane of classifier. The diversity between SVM and PDF will be applied in the future research as the reference to make the new algorithm more robust and accurate.