[A-AS04] Machine Learning Techniques in Weather, Climate, Hydrology and Disease Predictions

[AAS04-05] Retrieval of ice cloud properties from Himawari-8 measurement with a deep neural network method

*Xinyue Wang¹, Hironobu Iwabuchi¹, Takaya Yamashita¹ (1.Tohoku University)

Keywords:Deep neural network, Himawari-8 satellite infrared measurement, Retrieval of ice cloud property

Ice cloud constitutes a key component of weather and climate systems, and thus plays an important role in modulating the atmospheric radiation budget. Retrieval of ice cloud properties, such as cloud top height (CTH) and cloud optical thickness (COT), still needs further improvement in terms of accuracy and computational efficiency. In this study, a deep neural network (DNN) approach based on Himawari-8 satellite infrared measurement is presented. The DNN model is trained with 6 thermal infrared channels as input, and cloud properties from DARDAR observation as target. Supplementary variables such as the vertical temperature profile, satellite zenith angle, surface elevation, and detect time are used as input parameter. When compared to previous physical models, our DNN model can yield more high clouds with top height larger than 17.5 km and thicker than 20 in convective systems, especially over the tropical area, and closely following the DARDAR truth. For both the ice-CTH and -COT retrieval, bias mainly manifested as underestimation which is further quantitatively attributed to the cirrus cloud (COT smaller than 1) part. Despite of this, the DNN algorithm is validated to have much better performance on estimating thin cloud properties relative to the current JAXA official product that retrieved at solar wavelengths. One full-disk retrieval takes about 13 min and near-real time estimation can be practically applied in both diurnal scale weather monitoring and climate data record.