Assimilating visible satellite images for convective‐scale numerical weather prediction: A case‐study

Abstract

Listen

Translate article

AbstractSatellite images in the visible spectral range contain high‐resolution cloud information, but have not been assimilated directly before. This paper presents a case‐study on the assimilation of visible Meteosat SEVIRI images in a convective‐scale data assimilation system based on a local ensemble transform Kalman filter (LETKF) in a near‐operational set‐up. For this purpose, a fast look‐up table‐based forward operator is used to generated synthetic satellite images from the model state. Single‐observation experiments show that the assimilation of visible reflectances improves cloud cover under most conditions and often reduces temperature and humidity errors. In cycled experiments for two summer days with convective precipitation, the assimilation strongly reduces the errors of cloud cover and improves the precipitation forecast. While these results are promising, several issues are identified that limit the efficacy of the assimilation process. First, the linearity assumption of the LETKF can lead to errors as reflectance is a nonlinear function of the model state. Second, errors can arise from the fact that visible reflectances alone are ambiguous and only weakly sensitive to the water phase and cloud‐top height. And lastly, it is not obvious how to localise vertical covariances as visible reflectances are sensitive to clouds at all heights. For the latter reason, no vertical localisation was used in this study. To investigate the robustness of the results, the horizontal localisation scale, the assigned observation error and the spatial density of observations were varied in sensitivity experiments. The best results were obtained for an observation error close to the Desroziers estimate. High observation density combined with small localisation radii resulted in the smallest 1 hr forecast error. These settings were also beneficial for 3 hr forecasts, but forecasts at that lead time were less sensitive to the observation density and the localisation scale.