Atmospheric scene classification using CALIPSO spaceborne lidar measurements in the Middle East and North Africa (MENA), and India

Abstract

This paper presents a new algorithm based on the support vector machine (SVM) for classifying the Cloud-Aerosol Lidar Infrared Pathfinder Satellite Observations (CALIPSO) data into classes of clean air, cloud, thin aerosol, dense aerosol, surface, subsurface and totally attenuated. The procedure is as follows: At first, the considered features based on CALIPSO data are prepared. Brightness Temperature Differences between 10 and 12 μm (BTD11-12) is then used to better discriminate dense aerosols from clouds. The particle density feature proposed in this research is another feature participating in the classification. Training samples are automatically extracted by applying strict thresholds on the features. A wrapper feature selection is performed to rank the features based on their performance. Four post-processing steps are implemented to correct some misclassified cells e.g. edges of clouds and high-level clouds. The proposed algorithm was implemented on 4 datasets in the Middle East and North Africa (MENA), and India with various types and densities of aerosol. An accuracy assessment based on the comparison between the obtained results and ground truth samples indicated 0.94, 0.96 4, 0.92 and 0.89 kappa coefficients for the datasets. A statistical hypothesis test demonstrated that our SVM classification overcame CALIPSO vertical feature mask (VFM) product. The experimental result indicates the high accuracy of the proposed algorithm for the atmosphere scene classification using CALIPSO data.