Retrieval of Ice Cloud Parameters from Simulations of the Ice Cloud Imager (ICI) Based on Convolutional Neural Networks

The Ice Cloud Imager (ICI) aboard the second generation European meteorological satellites is the first space-borne submillimeter wave radiometer used for ice cloud remote sensing. Since ice clouds simultaneously contain three ice phase particles - ice, snow and graupel - and the physical distribution characteristics and submillimeter wave radiation properties of the three kinds of particles are different, simultaneously retrieving the physical parameters of the three ice cloud particles from submillimeter brightness temperatures is very challenging. This paper proposes a multi-phase ice cloud particle retrieval algorithm based on convolutional neural networks (CNNs) that can simultaneously retrieve total path and profile of ice, snow and graupel particles from submillimeter brightness temperatures. The retrieval experiments using the simulated ICI brightness temperatures show that the average relative error of retrieved total path of the three kinds of frozen particles is all lower than 57% and those of the content profiles is all lower than 68%.