[A-CG36] Satellite Earth Environment Observation

[ACG36-14] Development of a sky camera images analysis algorithm using machine learning for EarthCARE/MSI Level2 Cloud Product data validation

*Minrui Wang¹, Kanta Shimizu², Takashi Y. Nakajima¹ (1.Tokai University Research and Information Center, 2.Internet Initiative Japan Inc.)

Keywords:Satellite obversion, Cloud cover, Whole-sky camera image, Machine learning

Cloud properties such as cloud optical depth, cloud droplet size, and cloud top height (cloud top pressure) are planned to be retrieved from the Multi-Spectral Imager aboard the EarthCARE which will be launched in early 2024. For this analysis, we have developed two algorithms, one is the CLAUDIA (Cloud and Aerosol Unbiased Decision Intellectual Algorithm) that discriminates cloud pixels from clear pixels, the other is the CAPCOM (Comprehensive Analysis Program for Cloud Optical Measurements), a retrieval algorithm for cloud properties. These algorithms have been used for JAXA’s major satellite mission such as the GOSAT and the Shikisai (GCOM-C). In order to validate the CLAUDIA we have set three ground-based sky camera system at Shibuya (Tokyo), Kumamoto (Kyushu), and Iriomote island (Okinawa). In this presentation, we are going to mention about a machine learning algorithm for analyses of the sky camera (2-pi camera) images to discriminate cloud cover with clear area. Logistic regression method and XGBoost are used in our algorithm, and our results from the algorithm have shown very high accuracy on images of both day and night. Besides, our results of confusion matrix, Receiver Oprating Characteristic (ROC) and Area Under the Curve (AUC) also have shown high precision of the algorithm.