A neural-network-based method for generating synthetic 1.6 µm near-infrared satellite images

Abstract

Abstract. In combination with observations from visible satellite channels, near-infrared channels can provide valuable additional cloud information, e.g. on cloud phase and particle sizes, which is also complementary to the information content of thermal infrared channels. Exploiting near-infrared channels for operational data assimilation and model evaluation requires a sufficiently fast and accurate forward operator. This study presents an extension to the method for fast satellite image synthesis (MFASIS) that allows for simulating reflectances of the 1.6 µm near-infrared channel based on a computationally efficient neural network with the same accuracy that has already been achieved for visible channels. For this purpose, it is important to better represent vertical variations in effective cloud particle radii, as well as mixed-phase clouds and molecular absorption in the idealized profiles used to train the neural network. A new approach employing a two-layer model of water, ice and mixed-phase clouds is described, and the relative importance of the different input parameters characterizing the idealized profiles is analysed. A comprehensive data set sampled from Integrated Forecasting System (IFS) forecasts together with different parameterizations of the effective water and ice particle radii is used for the development and evaluation of the method. Further evaluation uses a month of ICOsahedral Non-hydrostatic development based on version 2.6.1 (ICON-D2) hindcasts with effective radii directly determined by the two-moment microphysics scheme of the model. In all cases, the mean absolute reflectance error achieved is about 0.01 or smaller, which is an order of magnitude smaller than typical differences between reflectance observations and corresponding model values. The errors related to the imperfect training of the neural networks present only a small contribution to the total error, and evaluating the networks takes less than a microsecond per column on standard CPUs. The method is also applicable for many other visible and near-infrared channels with weak water vapour sensitivity.