The core objective is a most accurate and reliable near-real time quantitative precipitation estimation (QPE). The estimate will be primarily based on observations from a network of polarimetric radars and attenuation measurements of commercial microwave links (CMLs), which provide the backhaul of countrywide cell phone networks. We will improve, newly develop and evaluate novel, generally applicable methods and apply them to observations of the recently fully upgraded dual-polarization C-band radar network of the German national weather service (DWD) and of a CML subnet operated by Ericsson Germany.

QPE for liquid precipitation will be based on specific attenuation R(A) following a method co-developed by the applicants and already successfully applied to S- and X-band radars. The method will be adapted, transferred to and optimized for DWDs radar network. For improving QPE for mixed-phase precipitation, we will exploit polarimetry for hydrometeor classification and develop new polarimetric methods for the quantification of snow fall intensity. To improve QPE at distances far away from the radars, we will develop the first polarimetric correction method for vertical reflectivity profiles, which will significantly improve QPE in regions with measurements only within and above the melting layer.

As summarized in the status report 2021, RealPEP-P1 suggests for precipitation estimates below the melting layer (ML) hybrid retrievals R(A, K_DP), which are based on specific attenuation A combined with specific differential phase shift K_DP for reflectivity Z_H* > 40 dBZ, because R(K_DP) is less sensitive to the variability in dropsize distributions (DSDs) in heavier rain and R(A) is invalid in rain mixed with hail. It is worthy noted that the attenuation parameter required to derive A in the ZPHI method is DSD-dependent and thus adjusted according to the change of differential reflectivity Z_DR with given Z_H (i.e., Z_DR slope) for each scan. As a result, the proposed polarimetric rainfall retrievals outperform the DWD operational QPE products, and have results comparable to R(Z_h, K_DP) but without suffering from partial beam blockage (Chen et al., 2021). Z_H denotes the reflectivity at horizontal polarization in units dBZ, while Z_h refers to the linear reflectivity in mm⁶m^-3 instead.

However, limitations resulting from the radar measurement geometry (i.e., weather radar monitors precipitation at increasing heights with increasing distance from the radar site) may lead to large errors despite the use of the novel polarimetric rainfall algorithms (see status report 2022). This deficiency is amplified during warm-rain processes if vertical precipitation gradients below the observing heights of operational radars are not taken into account and result in rainfall underestimation. For the Ahrtal flooding in 2021, RealPEP-P1 demonstrated, that the joint use of the gap-filling radar, JuXPol, and vertical profile correction including the application of DSD-derived relations measured with Micro-Rain-Radars (MRRs), which relates DSDs measured higher up to a few hundreds of meters to rain rates close to the ground, considerably mitigates large negative biases. The hybrid estimator R(A,K_DP) shows the lowest errors among all compared QPE products (Fig. P1-5, Chen et al., 2023).
For precipitation events with low ML heights, polarimetric vertical profile of reflectivity (PVPR) correction based on the statistics of Quasi-Vertical Profiles (QVP) and simulations is applied. For pure and uniform stratiform rain, this method successfully corrects for both, the enhanced Z_H due to the bright-band contamination and the reduced Z_H caused by beam broadening effects above the ML. The R(Z_h) retrieval based on the restored Z_h beats the one using hydrometeor-type specific relations (Fig. P1-6).
The first polarimetric snowfall estimation follows Bukovcic et al. (2020) but has been adapted to C-band radars and the climatology of Germany. The retrieval based on K_DP and Z_h agrees better with gauge observations compared to the conventional Z_h-based snowfall product by eliminating the impacts of the variability of particle size distribution on QPE.

Figure P1-5: Scatterplots of daily-accumulated QPE against gauge-measured rain totals for the R(Z_h), R(Z_h,K_DP), R(A_H,K_DP), and R(A_V,K_DP) retrievals (from top to bottom) based on data of the DWD radars (left column) and on the retrievals with inclusion of the JuXPol data and the vertical profile (VP) correction (right column). The color of the dots represents the height of radar observations above the ground.


Figure P1-6: Scatterplots of event-accumulated rainfall sums derived from the (a) R(Z_h), (b) hydrometeor type specific R(Z_h) and ( c ) PVPR-corrected R(Z_h) retrieval versus gauge-measured rainfall sums, and (d)-(f) their corresponding rainfall fields based on the HNR radar observations on 23 September 2018.

In addition to weather radars, Commercial Microwave Links (CMLs) are exploited as opportunistic sensors to accurately estimate path-averaged rainfall intensities near the ground at temporal resolutions of 1-minute and higher. The goal of P1 is to complement weather radar estimates with CML QPE near ground, e.g. for a better estimation of extreme events. As a first step towards merging QPE from CMLs and polarimetric weather radars, we implemented a simple additive correction algorithm for hourly aggregations. Figure P1-7 shows the accumulated rainfall measured during the heavy precipitation event in western Germany, in the Ahr valley, which led to severe flooding. An adjustment was only performed for periods and locations where the radar detected precipitation. Our results showed significant improvements of the hybrid retrievals R(Z, K_DP) and R(A, K_DP) in the left column of Figure P1-5 approaching the quality of R(A, K_DP) in the right column of Figure P1-5 which uses MRR-based relations with vertical profile corrections (VP) and gap-filling. Therefore, CMLs have a great potential to provide significant improvements in an operational scenario where MRR-based relations and gap-filling radars are not available.

After achieving this proof of concept our next goal is a more sophisticated merging at an increased temporal resolution. The main issue with comparing weather radar data and on-ground information from CMLs and rain gauges at a resolution of 1 to 5 minutes is that radar measurements are often obtained higher above the ground (increasing with increasing distance from the radar) and a spatial and temporal mismatch can be expected. Our recent work addresses temporal super-resolution, ground-adjustment and advection correction of radar rainfall. We use ResRadNet (Polz et al. 2024), a 3D-convolutional neural network to simultaneously treat systematic errors as well as temporal sampling errors compared to raingauge measurements at a 1-minute resolution. In this deep learning application, we only use radar observations as model input. Our results (see Fig. P1-8) showed that this approach can significantly increase the linear correlation and decrease the root mean squared error between radar estimates and ground sensors without introducing temporal or spatial inconsistencies. As a next step, we aim to extend this approach to additionally use CMLs and polarimetric variables as input for the neural network.

P1 also investigated the robustness/reliability of CML data in heavy precipitation. Therefore, Polz et al. (2023) analyzed attenuation induced complete loss of signal (blackout) in CML data. Blackouts potentially occur during heavy rain events and lead to missing extreme values. We analyzed 3 years of attenuation data from 4,000 CMLs in Germany and compared it to a weather radar-derived attenuation climatology covering 20 years. Figure P1-9 compares the observed and expected occurrence of blackouts during the analyzed period, as well an analyzed relation to temperature effects. On average, a CML experienced 8.5 times more blackouts than we would have expected from the radar-derived climatology and blackouts did occur more often for longer link paths (e.g., >10 km) despite an increased dynamic range. As a next step, we are investigating promising technological, as well as data-driven, measures to mitigate the impact of blackouts.

Figure P1-7: Daily-accumulated rainfall composite maps for 14th of July 2021. The left column shows radar-based QPE without/with CML adjustment. The top row shows results from an R(A, K_DP)v1 retrieval, which is based on R(A) combined with R(K_DP) for Z > 40 dBZ. The bottom row maps were derived from a similar R(A, K_DP)v2 retrieval which additionally uses relations derived from MicroRainRadar-measured dropsize distributions (DSDs) and vertical profile corrections for reflectivity Z and differential phase shift K_DP applied to four overlapping DWD radars and gap-filling X-band radar (JuxPol). Colored dots indicate the daily-accumulated rainfall sums measured by DWD rain gauges. Note that only gauges observing the whole day are shown here.


Figure P1-8: Maps of rainfall intensity for 18:00 on 6 July 2021. The upper row shows the 5-minute rain gauges, RADOLAN-RY, RADKLIM-YW, an aggregation of the neural network predictions from 18:00 to 18:05 (ResRadNet 5min), and the single neural network prediction from 18:00 (ResRadNet 1min). The bottom row shows the selected study area in Germany and the difference between the rain gauge value and the grid cell it is contained in using the product on top of the respective map. Red colors indicate an overestimation compared to the rain gauges and blue colors an underestimation.


Figure P1-9 (from Polz et al. 2023): (a and b) show the observed and expected number of blackouts per day and month between 2018 and 2020. ( c ) shows the mean 2 m temperature along all commercial microwave links (CMLs) derived from ERA-5-land. (d) shows the observed number of blackout minutes per CML per 3 hr compared to the average ERA5-land 2 m temperature along the link path during the same period. The red line in (c and d) indicates the 4°C threshold below which mixed type precipitation is more likely. 17.7% of all observed blackouts occurred below this threshold.


References:

• Bukovčić, P., A. Ryzhkov, D. Zrnić, 2020: Polarimetric Relations for Snow Estimation— Radar Verification, Journal of Applied Meteorology and Climatology, 59(5), 991-1009. Retrieved Feb 17, 2023, from https://journals.ametsoc.org/view/journals/apme/59/5/jamc-d-19-0140.1.xml.

• Chen, J.-Y., S. Trömel, A. Ryzhkov, C. Simmer, 2021: Assessing the benefits of specific attenuation for quantitative precipitation estimation with a C-band radar network, Journal of Hydrometeorology, 22(10), 2617-2631, https://doi.org/10.1175/JHM-D-20-0299.1.

• Chen, J.-Y., R. Reinoso-Rondinel, S. Trömel, C. Simmer, A. Ryzhkov, 2023: A Radar-Based Quantitative Precipitation Estimation Algorithm to Overcome the Impact of Vertical Gradients of Warm-Rain Precipitation: The Flood in Western Germany on 14 July 2021, Journal of Hydrometeorology, https://doi.org/10.1175/JHM-D-22-0111.1.

• Polz, J., M. Graf & C. Chwala, 2023: Missing rainfall extremes in commercial microwave link data due to complete loss of signal. Earth and Space Science, 10, e2022EA002456. https://doi.org/10.1029/2022EA002456.

• Polz, J., L Glawion, H. Gebisso, L. Altenstrasser, M. Graf, H. Kunstmann, S. Vogl, C. Chwala, 2024: Temporal Super-Resolution, Ground Adjustment and Advection Correction of Radar Rainfall using 3D-Convolutional Neural Networks. IEEE Transactions on Geoscience and Remote Sensing. doi: 10.1109/TGRS.2024.3371577, https://publikationen.bibliothek.kit.edu/1000169549.

Contribution of University of Bonn
Since weather radar monitors precipitation at increasing heights with increasing distance from the radar site, warm-rain event always suffers from large underestimation of rainfall due to vertical precipitation gradients. The use of a gap-filling radar, JuXPol, and vertical profile correction including the application of MRR-DSD-derived relations, which relates DSDs measured higher up to a few hundreds of meters to rain rates close to the ground, significantly mitigates large negative biases; especially the hybrid estimators R(A, KDP) using scan-wise adjusted alpha gives the lowest errors.

Figure P1-3: Daily-accumulated rainfall composite maps from the R(Z), R(Z,KDP), R(AH, KDP) and R(AV, KDP) retrievals (from top to bottom) based on the data of four DWD C-band radars (left column), retrievals with the vertical profile (VP) correction (middle column), and retrievals with additional inclusion of the JuXPol data and the VP correction (right column).The black circles indicate the areas where JuXPol provides lower-altitude observations than the DWD radars. The colored dots are the daily-accumulated rainfall sums measured by 306 DWD rain gauges.

Figure P1-4: Scatterplots of daily-accumulated QPE against gauge-measured rain totals for the R(Z), R(Z,KDP), R(AH, KDP) and R(AV, KDP) retrievals (from top to bottom) based on data of the DWD radars (left column) and retrievals with inclusion of the JuXPol data and the VP correction (right column). The color of the dots represents the height of radar observations above the ground.

Contribution of University of Bonn
The method for Quantitative Precipitation Estimation (QPE) based on specific attenuation R(A) has been successfully applied to S- and X-band radars, but still has to be adapted, transferred to and optimized for the C-band radar network of the German Weather Service (DWD). To establish a robust R(A) algorithm, its dependencies on temperature and DSD variability are investigated. In order to mitigate the sensitivity of A to DSD, normalized number concentration (Nw) is used to classify radar data and, accordingly, corresponding optimized alpha and R(A) relations are utilized. For heavier continental rain dominated by large raindrops originating from hail or contaminated with hail, the R(A) algorithm is combined with retrievals from specific differential phase shift R(KDP) (see Fig. P1-1).

Figure P1-1: Preliminary comparison of 9 hours accumulated rainfall on 19 July 2017 using the gauge-adjusted radar rainfall from DWDs RADOLAN-RW product (top panel) and the R(A)+R(KDP) hybrid algorithm (bottom panels) monitored by the radars in Flechtdorf (FLD) and Ummendorf (UMD).

Contribution of Karlsruhe Institute of Technology
A crucial processing step in QPE with CMLs is rain event detection, i.e. to separate rainy and non-rainy periods in the time-series of CML signal levels. We aim to improve QPE by introducing convolutional neural networks (CNNs) to this task. The CNN is trained on 10% of all available sensors and validated on the remaining 90%. The reference data set is RADOLAN RW from DWD. Applying the model to data from September 2018 has shown that the CNN significantly improves on the error rates of previously used methods (threshold for the rolling standard deviation), while generalizing well to the validation sensors. The performance of detecting rain events of different intensities is illustrated in Figure P1-2.

Figure P1-2: The detection accuracy, i.e. the methods ability to correctly classify periods as rainy or non-rainy (rain rate<0.1mm/h), of the CNN-based method versus the standard variability-based method for different rain intensities. The events are grouped by intensities according to RADOLAN RW.