An ensemble method for cloud mask calculation based on data from the AHI instrument onboard the Himawari-8/9 satellite using convolutional neural networks

The paper explores a method for calculating a cloud mask based on the use of several convolutional neural network classifiers trained for various observation conditions using a bootstrapping method. An algorithm developed on the basis of this method makes it possible to detect clouds in images obtained from the Advanced Himawari Imager (AHI) instrument onboard the Himawari-8/9 satellite, regardless of the Sun illumination conditions of the territory of the Asia-Pacific region under surveillance in the warm and cold seasons. The accuracy of the obtained results is assessed using a cloud mask provided by the US National Oceanic and Atmospheric Administration, NOAA. Numerical assessment and visual analysis show high accuracy of the developed algorithm, which is higher than the earlier classifier version offered by the present authors. When compared with the NOAA masks, the average F1-measure ranges from 75% at twilight to 85% during the daytime.