The overall goal of RealPEP project P3 is a significant improvement in quantitative precipitation forecasts (QPF) by numerical weather prediction (NWP) for up to a day, as well as the achievement of a seamless precipitation prediction framework. This will lead us to improvements in predicting discharge and potential flash floods in small- to meso-scale catchments. We plan to achieve the overall goal of P3 by using the extended information on cloud and precipitation states and processes hidden in satellite and radar polarimetricobservations for extending data assimilation in NWP along different pathways. These pathways include a) the direct assimilation of polarimetric moments, b) assimilation of a 3D observation-based micro-physics composite with e.g. hydrometeor mixing ratios, c) preconvective information obtained from satellites, and d) nowcasted fields.

Contribution of University of Bonn

In addition to the assimilation of 3D reflectivities and radial winds, RealPEP’s QPF subproject P3 also aims at the assimilation of polarimetric radar observations from the German C-band radar network. The development of DWD's polarimetric radar forward operator (EMVORADO-pol) has not yet been completed, so RealPEP is (as a first step) attempting the assimilation of polarimetric observations indirectly via 3D microphysical retrievals of liquid and ice water content (LWC and IWC). In the previous status report, the skill of various individual polarimetric LWC estimators adapted to the German C-band radar network based on disdrometer-measured LWC and corresponding T-matrix code simulated radar moments was shown (see Fig. P3-1). In the meanwhile, P3 developed a hybrid LWC estimator (Reimann et al., 2021) using LWC(Z, ZDR) for all rays with a total differential phase shift < 5° and LWC(A) for all other rays if Z<45dBZ and LWC(KDP) when Z>=45dBZ. When applied to real radar observations, the hybrid retrieval resulted in an improved correlation coefficient compared to the newly developed, individual retrievals (see Fig. P3-2) and existing retrievals from the literature (not shown). Observations within the melting layer are not used, while for observations in the ice phase the hybrid IWC(Z, ZDR, KDP) estimator proposed by Carlin et al. (2021) is utilized. The latter showed a good performance when evaluated with airborne in-situ measurements (Blanke et al., 2023).
We compare three assimilation configurations: A) assimilation of conventional observations only (e.g. SYNOP, AIREP, and TEMP data), B) assimilation of conventional and 3D-Z observations as performed in the operational routine, and C) like configuration B but with LWC/IWC as alternatives to Z where possible. Assimilation of radar observations (configurations B and C) significantly improves deterministic first-guess (one-hour) precipitation forecasts in terms of Fractions Skill Score (FSS; Roberts and Lean, 2008) for a convective squall-line event in July 2017 (brown, green, and blue curves are above the black curve in Fig.P3-3). While the assimilation of 3D-LWC as an alternative to Z systematically improves the deterministic precipitation forecast for the considered period (green curve), the assimilation of IWC is less successful (blue curve). Possible reasons include the use of an IWC retrieval not adapted to Germany, the increased observation errors above the melting layer, inadequacies of the model’s ice module, pronounced spatial degradation and uncertainties of KDP derived from the C-band volume scans with low radial resolution (1 km until May 2021). Optimisation of data assimilation parameters for LWC and IWC (e.g. observation localisations and observation errors) and evaluation of forecasts with longer lead times (up to 9 hours) are ongoing. In collaboration with the RealPEP subproject P4, these precipitation forecasts will then be used to evaluate the impact of polarimetric data on flash-flood forecasts in small- to mesoscale river catchments.

Figure P3-2: Histograms of disdrometer-derived LWC vs. LWC estimated with newly developed a) LWC(Z), b) LWC(Z,ZDR), c) LWC(A), d) LWC(KDP), e) LWC(KDP,Z), and f) hybrid LWC retrievals with corresponding root-mean square deviations (RMSD), correlation coefficients ®, and mean bias deviations (MBD). The hybrid retrieval (f) yields the best r.

Figure P3-3: Time series of fraction skill score (FSS; Roberts and Lean, 2008) for first-guess (one-hour) forecasts of accumulated precipitation for a threshold of 2mm/h produced in KENDA hourly assimilation cycles with a) the assimilation of only conventional observations (black), b) the additional assimilation of 3D-Z (brown), c) the assimilation of data as in b) but with LWC instead of Z where possible (green), and d) the same as c) but with IWC (blue).

References

• Blanke A., Heymsfield A.J., Moser M., Trömel, S., 2023: Evaluation of polarimetric ice microphysical retrievals with OLYMPEX campaign data. Submitted to Atmospheric Measurement Techniques. DOI: 10.5194/egusphere-2022-1488.
• Carlin J.T., Reeves H.D., Ryzhkov A.V., 2021: Polarimetric Observations and Simulations of Sublimating Snow: Implications for Nowcasting. In Journal of Applied Meteorology and Climatology 60 (8), pp. 1035–1054. DOI: 10.1175/JAMC-D-21-0038.1.
• Reimann L., Simmer C., Trömel S., 2021: Dual-polarimetric radar estimators of liquid water content over Germany. In Meteorologische Zeitschrift 30 (3), pp. 237–249. DOI: 10.1127/metz/2021/1072.
• Roberts N.M., Lean H.W., 2008: Scale-Selective Verification of Rainfall Accumulations from High-Resolution Forecasts of Convective Events. In Monthly Weather Review 136 (1), pp. 78–97. DOI: 10.1175/2007MWR2123.1.

Contribution of Deutscher Wetterdienst Offenbach

1 Overview

P3 continues in Phase 2 to work on the NWP component of the RealPEP flash flood prediction system, which includes as a central objective the assimilation of observations with relevance for the improvement of quantitative precipitation forecasts (QPF). We are working on the assimilation of polarimetric radar, CML, and satellite observations (prepared in part by projects C1 and P1) as well as information derived from them. Furthermore, we contribute to the seamless prediction aspect by testing and implementing the assimilation of nowcasted states and evaluate the impact of assimilating observations and/or derived microphysical variables re-gridded into composites as an alternative of assimilating the observations or derived variables given in their proper grids. The results of our developments are evaluated both with the quantitative precipitation estimates developed by project P1 and via their impact on flash-flood prediction (FFP) developed by project P4.

2 Assimilation of Commercial-Microwave-Link (CML) Observations

The use of Commercial-Microwave-Link (CML) observations for observing rainfall strongly improved in the past years. Erratic fluctuations of the raw CML data and biases arising from wet antenna attenuation could be corrected and led to stable and reliable rainfall estimates with resolutions of seconds or minutes. Accordingly, CML data are becoming an important complement to rainfall measurements based on RADAR observations. Within the exploratory experimental framework (EEF) of KENDA, we are implementing a setup for CML observations in order to test the assimilation of path-integrated attenuation. In our current setup we are employing the radar forward operator EMVORADO for obtaining simulated attenuations. After having tested the system in a single-observation setup, we will move to the assimilation of all available CMLs.

3 Convective Initiation (CI) Gridded Fields Assimilation

Forecast lead-times for precipitation can potentially be extended, when information on observable key quantities of a precipitation-inducing environment is available and can be assimilated in NWP. Project C1 - besides integrating observations and retrievals within RealPEP into the POLARA framework - did promising work on the derivation of total water vapor fields from satellite observations. Water vapor content and its spatial variability is considered an important indicator of the probability of convection development; thus its relation to yet non-precipitating clouds may provide usable information on convection before it has matured and already started to produce observations by the polarimetric radar network. Currently, we are in the process of exchanging the relevant data and preparing first experiments for assimilating this kind of data.

4 Targeted Covariance Inflation

First experiments performed within this project using Targeted Covariance Inflation (TCI) to initiate precipitation in areas void of any convection in all ensemble members (see Fig. P3-4) have already shown very promising results (Vobig et al., 2021). In our current implementation the TCI approach employs correlations between simulated reflectivities and simulated water vapor fields for increasing the spread of simulated reflectivities prior to the assimilation. The TCI achieves this spread increase by adjusting the simulated reflectivities within all ensemble members provided that the area where the TCI is applied to has observed reflectivities but missing simulated reflectivities. Thus, we are able to integrate an advantage of the rather bruteforce Latent Heat Nudging (LHN) approach into the LETKF framework. Note that, even though TCI has currently only been employed in the context of radar reflectivity assimilation, the use of TCI is likely to become a key ingredient to enable successful CML and CI assimilation. This is because—as already observed when assimilating radar reflectivity—missed precipitation cells in ensembles will most probably cause problems here. Currently, we are investigating the use of process-based correlations (e.g. correlations only associated with convective environments) and study how to mitigate negative effects of the TCI on observation error statistics for the humidity.

Figure P3-4: Radar reflectivity composites for different forecast times (row-wise) where the reference date is 2019-06-03 12 UTC. The first column shows the observed reflectivities and the other columns show (from left to right) the simulated reflectivities (using the ICON-D2 model) using default settings without TCI, non-default settings without TCI, and non-default settings with TCI. The two arrows indicate the region where the TCI successfully produces two cells that are consistent with observed cells but are missing without an application of the TCI. Note that the TCI has only been applied once here during the assimilation at 12 UTC.

5 Implementation of seamless prediction with KENDA

Furnished with the theoretical concept of assimilating (nowcasted) future states into KENDA developed during Phase 1 (Potthast et al., 2022) we are in the process of implementing and testing a first version of a seamless prediction with KENDA for the test cases. This experiment is based on the assimilation of the 2D (or 3D) composite including its nowcasted extensions (either deterministic or as a nowcast ensemble). With the nowcasted information, we can generate analyses for future times (quasi-analyses), which should have a higher quality than the predictions starting with the normal analysis based on already observed states, because they contain besides the latter also their projections. Given a correct estimation of the respective error covariances, the impact of the nowcasted information will fade with the length of the assimilated nowcasted data. It still needs to be proven, however, that the success found for our simple play systems also holds for the much more complex and multidimensional weather system. Thus, we will use the 3D composite and its nowcasted extensions into the future for the test period agreed on in RealPEP and compare the nowcast- analyses of the future with the normal predictions.


References

• Vobig, K. et al., 2021: Targeted covariance inflation for 3D-volume radar reflectivity assimilation with the LETKF. RMetS, 147(740), 3789–3805.
• Potthast, R. et al., 2022: Data Assimilation of Nowcasted Observations. Monthly Weather Review, 150, 969–980.

Contribution of University of Bonn

The RealPEP subproject P3 is currently focusing on the identification of polarimetric liquid water content (LWC) retrievals suitable for the subsequent indirect assimilation of polarimetric observations from the German C-band radar network into the ICON model. For this purpose, existing LWC relations are evaluated and several new LWC estimators are developed and tested on the basis of a large disdrometer data set from Germany and corresponding T-matrix scattering calculations. Among the newly developed LWC estimators, the LWC(KDP, Z, ZDR, AH), LWC(KDP, ZDR, Z), LWC(AH, ZDR, Z), and LWC(KDP, Z, AH) relations achieve comparable and best performances in terms of root mean square error (RMSE), standard deviation and correlation coefficient when applied to the simulated radar moments (see Fig. P3-1). In contrast, the LWC(Z) relation shows the worst performance, while the LWC(KDP) relation yields the second worst performance, which can be explained by noise in the simulated KDP data in the lower LWC range (not shown). The comparison of the new LWC estimators with existing relations from the literature, their application to real radar observations from the German C-band radar network, and the development of a hybrid LWC retrieval are ongoing.

Figure P3-1: Taylor diagram comparing the root mean square error (RMSE), the standard deviation, and the correlation coefficient between the measured LWC from a large German disdrometer data set and the LWC estimated from newly developed polarimetric retrievals using radar moments simulated with T-matrix scattering calculations based on the same DSD data set. Used radar moments are horizontal reflectivity (Z), differential reflectivity (ZDR), specific horizontal attenuation (AH), and specific differential phase shift (KDP). Retrievals based on KDP are only developed and evaluated for data with Z>28dBZ.

Contribution of DWD Offenbach

Current work focuses on the assimilation of radar reflectivities (Z) and radial winds (RW) within the KENDA system using the ICON model of DWD. While assimilation of RW will be in operational use soon, remaining scientific tasks are currently under investigation regarding the assimilation of Z, i.e. situations where the model predicts the emergence of a convective cell not observed in nature or vice versa. It is important to note that these discrepancies are often accompanied by small ensemble spreads w.r.t. Z and, therefore, the ensemble kalman filter (LETKF) is not able to produce adequate increments for eliminating the aforementioned discrepancies. We are working on overcoming this problem via a so-called covariance inflation for reflectivities where the spread of Z is increased by employing a correlation between Z and the humidity QV. As indicated in Figure P3-2 our approach for the covariance inflation of reflectivities leads to a larger spread for Z, especially in regions with large differences of observed and simulated reflectivities. An in-depth study of the subsequent assimilation using these inflated reflectivities shows an overall positive impact, however, there are still several open questions that we have to investigate.

Figure P3-2: Example of observed reflectivities Zobs, simulated reflectivities Zsim, the difference Zsim- Zobs, and the ensemble spread of Z (s(Z)) for 30 May 2016 at 15 UTC at the radar station in Dresden (left column). The left column shows according variables without and the right one with application of a covariance inflation for the simulated reflectivities.