Transfer learning used to analyze the dynamic evolution of the dust aerosol. (March 2015)

• Record Type: Journal Article
• Authors: Ma, Yingying; Gong, Wei; Mao, Feiyue
• Is Part Of: Journal of quantitative spectroscopy & radiative transfer. Volume 153(2015:Mar.)
• Page Start: 119
• Page End: 130
• Publication Date: 2015-03
• Subjects: CALIPSO -- CloudSat -- Dust aerosol -- TrAdaboost -- SVM; Spectrum analysis -- Periodicals; Radiation -- Periodicals; Analyse spectrale -- Périodiques; Rayonnement -- Périodiques; Radiation; Spectrum analysis; Periodicals; 543.0858
• DOI: 10.1016/j.jqsrt.2014.09.025
• Languages: English
• ISSNs: 0022-4073

Abstract

Abstract: To keep the advantage of Support Vector Machine (SVM) in analyzing the dynamic evolution of the dust aerosol, we introduce transfer learning as a new method because transfer learning can utilize knowledge from previously collected data and add dozens of new samples, which can significantly improve dust and cloud classification results. It can also reduce the time of sample collection and make learning efficient. In this paper, we receive significant improvement effect using SVM as the basic learner in TrAdaBoost during four consecutive dust storm days, and correct one error classification in PDF. As a result, dust aerosol in high altitude can even spread to stratosphere. Moreover, in the process of dust aerosol transportation, it is highly affected by anthropogenic aerosol, for example, the color ratio (CR) changes from 0.728 to 0.460 and finally reaches 0.466, while depolarization ratio (DR) changes from 0.308 to 0.081 and finally reaches 0.156. It is indicated that the big size and non-spherical aerosol particles reduce obviously after dust aerosol deposition, but small size and spherical anthropogenic aerosol also produce a certain effect, and on March 22, 2010 had a small recovery above the ocean following the reduction of DR and CR. Due to the MODIS resolution not meeting the observation requirement and layer identification being different between CALIPSO and CloudSat, a problem such as stratocumulus cloud in low altitude still exists in aerosol and cloud classification. Lack of ground-based auxiliary data is the main problem which hinders our validation and quantitative analysis. It is pressing for a solution in future. Highlights: We introduce a transfer learning method to identify the dust aerosol. This method can observe the dust storm dynamic evolution. MODIS image, cloud mask and CloudSat data are used to validate the result. We analyze dust aerosol change character, when it transports from west to east China.