The Impact of Different Support Vectors on GOSAT-2 CAI-2 L2 Cloud Discrimination

Greenhouse gases Observing SATellite-2 (GOSAT-2) will be launched in fiscal year 2018. GOSAT-2 will be equipped with two sensors: the Thermal and Near-infrared Sensor for Carbon Observation (TANSO)-Fourier Transform Spectrometer 2 (FTS-2) and the TANSO-Cloud and Aerosol Imager 2 (CAI-2). CAI-2 is a push-broom imaging sensor that has forwardand backward-looking bands to observe the optical properties of aerosols and clouds and to monitor the status of urban air pollution and transboundary air pollution over oceans, such as PM2.5 (particles less than 2.5 micrometers in diameter). CAI-2 has important applications for cloud discrimination in each direction. The Cloud and Aerosol Unbiased Decision Intellectual Algorithm (CLAUDIA1), which applies sequential threshold tests to features is used for GOSAT CAI L2 cloud flag processing. If CLAUDIA1 is used with CAI-2, it is necessary to optimize the thresholds in accordance with CAI-2. However, CLAUDIA3 with support vector machines (SVM), a supervised pattern recognition method, was developed, and then we applied CLAUDIA3 for GOSAT-2 CAI-2 L2 cloud discrimination processing. Thus, CLAUDIA3 can automatically find the optimized boundary between clear and cloudy areas. Improvements in CLAUDIA3 using CAI (CLAUDIA3-CAI) continue to be made. In this study, we examined the impact of various support vectors (SV) on GOSAT-2 CAI-2 L2 cloud discrimination by analyzing (1) the impact of the choice of different time periods for the training data and (2) the impact of different generation procedures for SV on the cloud discrimination efficiency. To generate SV for CLAUDIA3-CAI from MODIS data, there are two times at which features are extracted, corresponding to CAI bands. One procedure is equivalent to generating SV using CAI data. Another procedure generates SV for MODIS cloud discrimination at the beginning, and then extracts decision function, thresholds, and SV corresponding to CAI bands. Our results indicated the following. (1) For the period from November to May, it is more effective to use SV generated from training data from February while for the period from June to October it is more effective to use training data from August; (2) In the preparation of SV, features obtained using MODIS bands are more effective than those obtained using the corresponding GOSAT CAI bands to automatically extract cloud training samples.