Assimilating novel ground-based remote sensing observations into a numerical weather prediction model for improving model predictions and advancing knowledge of atmospheric boundary layer processes
Final Report Abstract
In this project the assimilation of two ground-based remote sensing devices, the passive microwave radiometer and the active Doppler lidar was developed, and the impact on analysis and forecast of the numerical weather prediction model ICON-D2 was investigated. The forecast model is the convection-allowing regional operational numerical weather prediction system of DWD. Data was obtained from a HATPRO-G5 microwave radiometer and a HALO Photonics Doppler lidar, both based at the Meteorological Observatory in Lindenberg operated by DWD. The assimilation system was the Local Ensemble Transform Kalman Filter (LETKF), which is operated by DWD for the regional forecasting system. For the assimilation of the microwave radiometer data the incorporation of a forward operator to transform the model variables into the brightness temperatures to obtain a suitable model equivalent was necessary as a first step. For this the community model RTTOV-gb was installed and an interface to hand over the ICON-D2 model variables was developed. This allowed then the investigation of observation minus first guess statistics. The observation in this case were the 5-minute averaged brightness temperatures in the 14 channels of the microwave radiometer. Due to highly correlated observation errors of the 14 channels, it was found to be beneficial not to assimilate all 14 channels together, but to restrict ourselves to 1-2 channels, one in the temperature-sensitive V-band and optionally another one from the moisture-sensitive K-band. The reason for this finding is probably the fact, that currently the cross-correlated errors of different channels are not accounted for in the LETKF and this aspect should be revisited, once the proper accounting for the cross-correlations is implemented. In the experiments conducted a cloud screening was carried out to screen out cloudy observations and start with clear-sky assimilation first. In later experiments with only a temperature-sensitive channel which is sensitive to the lower boundary layer only, no cloud screening was carried out, and an improvement of the scores was found, probably due to better data availability. Detailed experiments and evaluation of the results was carried out with respect to vertical localization and observation error specification. In a LETKF data assimilation system the observations are usually localized, i.e. their impact is limited to a region around the observation, both horizontally and vertically. However, as microwave radiances are integral measurements are over the whole vertical column, the latter is not trivial. The problem is not the localization radius, but do define a height, around which to localize. In general, one seeks to estimate and localize around that height, at which the observation is most sensitive. Observation errors were iteratively estimated with the so-called Desrozier-technique. Having found optimal settings with a large number of sensitivity experiments and the above-mentioned approach, a positive impact on the lower boundary layer temperature and moisture (up to 750hPa) at the Lindenberg column could be obtained by assimilating the brightness temperatures from the microwave radiometer. The biggest challenge that we were facing was the verification of the experiments. Radiosondes were available only every 6 hours, which lead to the problem, that often the differences between experiments with only slightly modified settings were not significant. The preparation of the assimilation of the Doppler lidar observations was more straightforward, as no complex forward operator was needed. The horizontal wind speed provided was assimilated. Also here a large number of longer experiments were carried out and evaluated carefully. Here, the averaging time of the observations turned out to be a critical aspect. 10 minutes averages gave much better results than a 30-minute average. This however might be different for different models with different resolution. Also here the observation error was estimated based on the Desrozier method. Moreover, the data was vertically thinned. The verification of the experiments was much more robust, than for the microwave radiometer experiments, because as verifying observations data from the nearby radar wind profiler could be used, which delivers continuous horizontal wind profiles as well. We obtained about 10 percent improvement for the boundary layer winds in Lindenberg four a month-long experiment in June 2021. For another period in August 2022, however, the results were more mixed, no clear positive impact could be obtained. The reason probably was that in this experiment very dense MODE-S data was assimilated as well, and the radial winds from the precipitation radars, thus already a lot of wind information was available in the reference. Generally, it turned out to be difficult to obtain robust results when assimilating only a single device. However, with the findings obtained in this project it will be easy to set up follow-up experiments when more data from more devices will become available.
