| Cloud play a significant role in the weather climate change.The cloud characteristics and microphysical characteristics affect the global equilibrium radiation balance.With the continuous development and the improvement of meteorological satellites,the use of meteorological satellites to observe the earth’s atmosphere has become one of the most important means to study the earth’s atmosphere.The launch of a new generation of geostationary meteorological satellites enables researchers at home and abroad to observe a certain region of the earth for a long time with high spatial and temporal resolution.Among them,cloud mask(cloud detection),as the most essential products for remote sensing,is one of the most important products in downstream inversion applications.In this paper,a cloud detection algorithm of satellite spectral imager based on machine learning is developed,which is applied to the Advanced Himawari Imager(AHI)onboard the Himawari-8 geostationary satellite.Collocated active observations from Cloud-Aerosol Lidar with Orthogonal Polarization(CALIOP)are used as reference labels for model development and validation.Based on the observation information with and without solar reflection channels,algorithms for daytime and night were developed respectively,and the effect of artificial neural network and random forest algorithm on cloud recognition was systematically compared.In order to obtain more reasonable cloud detection results under different land surface types,we proposed three different methods(independent modeling,single parameterization,and multi-parameter characterization)to deal with land surface types,and found that more reasonable processing(multi-parameter characterization)could improve the accuracy of cloud detection by ～3%.Using independent verification of results from the CALIOP cloud product,our machine learning day algorithm outperformed the current official AHI product,with the random forestbased algorithm improving cloud pixel detection accuracy by 5% and achieving 94%.The night algorithm using only infrared observation data is also quite stable,and the accuracy of cloud and clear pixel detection is about 87%.Part of the misclassification of AHI cloud detection is due to the lack of strong water vapor channel of 1.38μm,which makes it difficult to identify thin cirrus.Furthermore,the models can be further improved by using better constrained data and additional information(such as cloudy spatiotemporal distributions,observational geometry information). |
