Ice water path retrievals from Meteosat-9 using quantile regression neural networks

Abstract. The relationship between geostationary radiances and ice water path (IWP) is complex, traditional retrieval approaches are not optimal. This work applies machine learning to improve the IWP from Meteosat-9 observations, with a focus on low latitudes, training models against retrievals based CloudSat. Advantages of include avoiding explicit physical assumptions data, an efficient use information all channels, easily leveraging spatial information. Thermal infrared (IR) used as input achieve performance independent solar angle. They compared including reflectances well subset IR channels for compatibility historical sensors. accomplished quantile regression neural networks. network type provides case-specific uncertainty estimates, compatible non-Gaussian errors, flexible enough be applied different architectures. Spatial incorporated into through convolutional (CNN) architecture. choice outperforms architectures that only pixelwise. In fact, CNN shows good by using channels. makes it possible compute diurnal cycles, problem CloudSat cannot resolve due its limited temporal sampling. These compare favourably in CLAAS, dataset approach. results highlight possibilities overcome limitations physics-based while providing efficient, probabilistic methods. Moreover, they suggest this first can extended higher latitudes data considered complement upcoming Ice Cloud Imager mission, example, bridge gap sampling respect space-based radars.