Optimal vertical localization for the assimilation of cloud-affected satellite observations

Localization is essential for any ensemble-based data-assimilation system for numerical weather prediction, and most localization approaches are distance-based. For example, in the observation-space localization used by the Deutscher Wetterdienst (DWD), the localization is a function of the distance between a model grid point and an observation location. Observation-space localization for satellite observations is especially challenging because they do not have a constant or well-defined observation location. Instead, the observed signal may originate from various vertical levels and is affected by the presence of clouds. We derive an optimal localization for all-sky visible and infrared satellite observations over Germany by minimizing the difference between the DWD operational analysis and radiosonde profiles in a 1-month cycled assimilation experiment that excluded radiosondes. We use reconstructed partial analysis increments (PAI) to approximate a wide range of localization settings without needing to rerun the costly month-long experiment. We find that visible satellite observations require no localization, but that infrared observations deteriorate the analysis if they are not localized carefully.