[A-CG36] Satellite Earth Environment Observation

[ACG36-P05] Remote sensing of three-dimensional clouds from multispectral measurements using deep learning

*Hana Kato¹, Hironobu Iwabuchi¹ (1.Center for Atmospheric and Oceanic Studies, Graduate School of Science, Tohoku University)

Keywords:satellite remote sensing, cloud , deep learning, machine learning

Clouds are an important factor in determining the Earth's radiation budget, and those properties such as cloud optical thickness (COT) are globally retrieved by satellite observations. A standard COT retrieval method is the bispectral method based on 1D radiative transfer model neglecting horizontal radiative transfer. In addition, this method is based on the independent pixel approximations which assumes that clouds are plane-parallel and homogeneous within each pixel of the satellite image. However, it has been pointed out that not taking into account 3D radiative transfer and cloud inhomogeneity can lead to large errors in retrieval of cloud properties. Here we present a deep neural network (DNN) approach for the retrieval of COT from multispectral satellite observation. In order to account for 3D radiative transfer and cloud inhomogeneity, the training and test data of the DNN are made by a Monte Carlo 3D radiative transfer model whose input is 3D cloud fields from large-eddy simulations. The DNN structure consists of a convolutional neural network which is good at learning spatial features of cloud imprinted in satellite image. When we compare our method with pixel-by-pixel retrieval trained by a 1D radiative transfer model, our retrieval method can represent 3D radiative transfer effects, utilizing spatial features and data from 3D radiative transfer models. We will also report on results of application to the SGLI (Second generation GLobal Imager) measurement data onboard GCOM-C.