The real-time radar product evaluated in this study is based on the 5 min volumetric (3D; i.e., a collection of typically 5–15 scans executed at different elevation angles) data from at most seven magnetron-based polarimetric C-band ground-based weather radars. Doppler clutter filtering is applied to both horizontal and vertical reflectivity factor (Zh and Zv, respectively) data in order to remove non-meteorological echoes. Two radars are located in the Netherlands, two radars are located in Germany, and three radars are located in Belgium. Fig. a shows the locations of these radars including that of an additional radar in Germany (Neuheilenbach) that has been used in the real-time product since 9 September 2024 (i.e., outside the evaluation period). On average 4.8 and 6.1 radars contribute to the real-time product for the old (31 January 2022–31 January 2023) and renewed (31 January 2023–31 January 2024) product, respectively. For both periods, the median number of contributing radars is five. Over a large part of the radar product domain, the distance to the nearest radar is shorter than 150 km (Fig. a). Notable are the longer distances to the nearest radar in the Dutch–German border region. More information on the Belgian radars is provided by , on the Dutch radars by , and on the German radars by .

All radars that are used in this product are calibrated during regular maintenance (1–2 times per year). Monitoring tools are in place at each institute, which can assist to detect drifts in calibration so that action can be taken if this is the case. It is therefore assumed that the calibration of the radars is accurate. For more information on calibration and monitoring, see for the Dutch radars, and for the German radars.

The radar data processing is described in Sect. . The end result is a 2D real-time radar product of 5 min accumulations every 5 min on a ∼1km×∼1km grid with almost full data availability. The 5 min accumulations are aggregated to 1 h (every hour on the hour) and daily (ending 00:00, 06:00, or 08:00 UTC) accumulations for comparison against KNMI automatic rain gauge accumulations from the Netherlands (1 h) and against daily rain gauge accumulations from different networks across the entire radar product domain. Accumulations were only computed when data availability was at least 83.3 %. The annual accumulations were generated without this data availability criterion.

The combined daily radar–gauge availability is generally 95 % or higher (Fig. b and c). Lower availability is caused by missing gauge records. The exception is that for the renewed radar product the lower availability for the (south)western and southeastern part of the product domain is caused by lower radar data availability. This extended coverage with lower availability can be attributed to the Wideumont radar in the southeast that started contributing in August 2023 and the addition of longer range scan data from the Jabbeke radar in the southwest.

For each radar, a 2D radar pseudo-constant-altitude plan position indicator (pseudoCAPPI) of Zh is produced by logarithmic averaging of the Zh (dBZh) data from three scans. This is converted to precipitation intensities R through a Zh=200R1.6 relationship . A constant mean-field bias adjustment factor is subsequently applied to this intensity field . The mean-field bias adjustment factor is based on radar and rain gauge 1 h accumulations from a previous clock-hour (every hour on the hour; as explained in Sect. ). The resulting intensity fields from each radar are transformed to 5 min accumulations and subsequently combined into a single 2D composite via range-weighted compositing.

The processing for the renewed radar product starts with the removal of non-meteorological echoes (clutter) by employing the fuzzy logic algorithm from the wradlib open-source Python library for weather radar data processing . The clutter results from reflections of non-meteorological targets, such as the ground, buildings, trees, cars, sea surface, ships, birds, and planes. This may be exacerbated by anomalous propagation leading to the beam hitting the ground or sea surface. Interference by other sources of radio waves such as wireless networks and the sun also cause clutter . Removal of clutter-affected data is important, since they can lead to severe precipitation overestimation.

The fuzzy logic algorithm uses decision variables to classify clutter for each voxel from each scan separately . The following decision variables are employed for the Dutch radar data, with their weight provided between brackets: texture (spatial variability) of the differential reflectivity (ZDR; 0.20), the copolar correlation coefficient (ρHV; 0.15), texture of ρHV (0.25), depolarization ratio (0.20) , and clutter phase alignment (0.20) . For each decision variable, the degree of membership to the meteorological target class is calculated employing a membership function. A weighted average of the degree of membership to the meteorological target class is computed using all decision variables, each having its own weight. A voxel is classified as clutter if this weighted average is below a threshold value of 0.6. In that case the associated Zh is set to the “nodata” value (i.e., areas void of data). All details regarding the fuzzy logic algorithm and its (parameter) settings are provided by .

The Belgian radar data lack the decision variable clutter phase alignment. Utilizing the same membership functions, only three decision variables are used and with a different weight: texture of ZDR (0.5), texture of the two-way differential propagation phase (0.3), and ρHV (0.2). Moreover, the employed threshold value in the classification is 0.3.

give more information on the Belgian radar processing. For the German radars, no polarimetric data were publicly available for the periods that are used in this paper. However, DWD has already applied several post-processing algorithms, including a (partly polarimetric) fuzzy logic clutter detection and removal algorithm for the two 1-year periods. provide details on the German radar processing.

Rain-induced attenuation can lead to severe precipitation underestimation for C-band weather radars . The two-way path-integrated attenuation (PIA) along the radar beam can be estimated and corrected for. Here, the recipe from is followed. Horizontal specific attenuation is computed from the specific differential phase (Kdp) assuming a linear relation between the two for the Belgian and Dutch radars (kh=0.081Kdp). Because real-time access to dual-polarization data from the German radars was not possible, a Hitschfeld–Bordan type of algorithm was applied there. This method, called the modified Kraemer method, uses the method proposed by with an additional limit on the attenuation correction of 10 dB and on the resulting reflectivity of 59 dBZh to avoid instability of the algorithm . The final step for both methods is to increase the Zh value for each voxel by the computed attenuation. The method using Kdp was shown to outperform the modified Kraemer method using 1 year of data from the two C-band radars in the Netherlands, but the stability of the modified Kraemer method was also confirmed .

For both methods, the wradlib open-source Python library for weather radar data processing is utilized. Since the correction is meant for attenuation caused by liquid precipitation, a voxel is only used to compute the attenuation in case its height is below the forecasted freezing-level height from the numerical weather prediction model HARMONIE-AROME. And for the method using Kdp, voxels classified as clutter are not employed to compute attenuation . Attenuation correction is applied to all voxels, so also those above freezing-level height.

For the German radars, for which no polarimetric data were available, the modified Kraemer method is applied per elevation scan via the kh=αZhβ relation. The values of α and β are reduced from their initial values in an iterative procedure until the constraints Zh,cor≤59 dBZh and PIA≤10 dB are met. In case constraints are not met, no attenuation correction is performed. More details are provided by , , and .

Here, the algorithm of is extended to attenuation correction for vertically polarized reflectivity, employing kv=0.058978Kdp, in order to obtain the attenuation-corrected ZDR. This is important, since it is used in the VPR correction.

Precipitation generally has a non-uniform vertical structure. Therefore, the increased height of the radar sampling volume at a long range can lead to errors in surface precipitation estimates. To address this, the VPR correction is applied as suggested by , extended with the polarimetric melting layer detection from . The algorithm estimates two idealized VPR profiles, one for stratiform, and one for undefined precipitation based on all data that have been classified as one of these two precipitation types. Note that no VPR estimation or correction is applied for convective precipitation because there is much more vertical mixing in convection. The VPR estimates include an uncertainty estimate that is used in subsequent steps of the QPE processing chain. The VPR estimates are then used to extrapolate all reflectivity data down to the ground. For both estimation and correction for the VPR, the shape of the beam is assumed to be Gaussian.

We use quality information to combine data from several scans to derive a 2D surface reflectivity product. Appendix  details the estimation of a quality index QT for each voxel. The N voxels that contribute to a given pixel in the 2D surface reflectivity product are combined as follows into a quality-weighted reflectivity (Zh,Q): 1Zh,Q=∑j=1NQT,jZh,j∑j=1NQT,j, where both reflectivities are expressed in mm6m-3.

A combined quality index is also computed for each pixel in the 2D radar product (Qr). The basic principles are that each added voxel j raises the value of the quality index, and that the resulting quality index ranges between 0 and 1: 2Qr=1-∏j=1N(1-QT,j). The advantage of this approach is that voxels from scans that have a non-zero quality can compensate for the zero quality of voxels from other scans, and that voxels with a lower quality have a much lower weight in the final product.

The Gabella clutter filter from the wradlib open-source Python library for weather radar data processing has been successfully applied for the detection and removal of residual clutter e.g., RADKLIM for Germany and EURADCLIM for Europe;. Here, it is applied to the 2D Cartesian Zh,Q data from each radar and time interval separately. First, large gradients of Zh,Q in space are detected. For each central pixel, the number of pixels within a square lattice of 5×5 pixels is counted that have a Zh,Q value less than 6 dB lower than the central pixel. The central pixel is classified as clutter if this number of pixels is fewer than six. Second, the ratio between the area and circumference for contiguous echo areas with Zh,Q above 0 dBZh is calculated. The central pixel is classified as clutter if this ratio is lower than 1.3 pixels. The central pixels classified as clutter do not contribute to the radar product by setting the Zh to the “nodata” value and the quality index to 0.

The radar quality index (Appendix ) is also used to weigh the reflectivity values from individual radars into a combined composite of Zh,Q. It employs the same algorithms as for the construction of 2D surface reflectivity products for each radar: the quality weighted Zh is computed using Eq. () and the corresponding quality index is computed using Eq. ().

Reflectivities Zh,Q (mm6m-3) are converted to rainfall intensities R (mmh-1) employing the Zh,Q –R relationship : 3Zh,Q=200R1.6. This implicitly assumes that the precipitation is liquid, i.e., rain.

Storms that move more than 1 km (the spatial resolution of the QPE product) in 5 min (its temporal resolution) will lead to artifacts in precipitation accumulations. This is corrected by interpolating radar composites in time. For this purpose, advection vectors are computed using the implementation of in the OpenCV library (http://opencv.org, last access: 4 February 2025). These vectors are then used to interpolate the radar composites in time. The resulting interpolated precipitation intensity becomes a weighted average of the previous (t=t0) and current (t=t1) precipitation composite: 4Rint(x,t)=t1-tt1-t0R(x+v(t-t0),t0)+t-t0t1-t0R(x-v(t1-t),t1), with t the time (t0≤t≤t1), Rint(x,t) the interpolated precipitation intensity, R(x,t) the composite precipitation intensity, x the location of the pixel, and v the advection vector (ms-1). Subsequently, the advection-corrected 5 min precipitation accumulation is computed by aggregating a number of interpolated composites and multiplying by the time interval between the different interpolated composites: 5Racc(x,t1)=t1-t0N+1∑i=0NRintx,t0+it1-t0N, where N-1 is the number of composites that is generated between the previous (t0) and present (t1) composite. Since the horizontal velocity of showers is expected to be lower than 50 ms-1, N=14 is used for this spatiotemporal resolution (1 km and 5 min).

The method for the adjustment of radar accumulations with rain gauge accumulations is a modified version of Barnes' objective analysis . The aim is to have a method that is robust in situations with large precipitation gradients and limited gauge density, and that can use quality information of both the radar and rain gauge precipitation estimates. The spatial adjustment factor for a given pixel is computed by the ratio of distance-weighted sums of the unadjusted radar and corresponding rain gauge precipitation accumulations at the gauge locations. The computation of the spatial adjustment factor field is described in Appendix . Although the adjustment factor field itself is computed based on radar and gauge accumulations from the same 60 min interval, it does not encompass the radar data from the current 5 min time interval due to the 50 min latency of gauge data, that were only available every clock-hour (every hour on the hour). Hence, the adjustment factor field from a previous clock-hour (every hour on the hour) was used. For instance, from 10:50–11:50 UTC the 5 min accumulations were adjusted employing the spatial adjustment factor field based on 60 min radar and gauge accumulations from 09:00–10:00 UTC. As of 11:50 UTC, the applied spatial adjustment factor field was computed based on data from 10:00–11:00 UTC. And possibly, sometimes even an older adjustment factor field was employed.

Maps of the maximum 5 min precipitation accumulation over 1 year as shown in Fig. a and b reveal values of 10 mm and higher. These are mostly associated with shipping tracks and wind farms over the North Sea and the English Channel, and metal cranes and containers of the Port of Rotterdam (“Maasvlakte (2)”) for the old radar product (Fig. a). They are hardly present in the renewed radar product, showing the effectiveness of the fuzzy logic algorithm and the Gabella clutter filter, with only some remaining isolated spots that are probably due to wind farms (Fig. b). Clutter for the land surface is less of an issue than sea clutter, which is clearly visible for the old radar product. Maps of the relative frequency of 5 min precipitation ≥0.01 and >6 mmh-1 for the old radar product (Fig. c and e) reveal clutter for “Maasvlakte (2)”, shipping tracks (much more pronounced for 6 mmh-1) and wind farms in the North Sea, and quite some suspicious areas for the land surface that are likely clutter. Few clutter areas appear in the renewed radar product, except for the radial patterns over the North Sea and the English Channel for ≥0.01 mmh-1, pointing toward the Belgian radar in Jabbeke, which are caused by interference (Fig. d).

Fig.  shows that to some extent the maximum accumulations, but especially the relative frequencies of exceeding a threshold, are generally larger for the renewed radar product. This could be partly due to the much wetter 1-year period, but also indicates lower underestimation due to improved processing. Note the high 5 min maxima in the southeastern part for the old radar product, which seem connected to rain showers. The spatial patterns here are typical for storms that move faster than a single pixel (1 km) in the time between two samples (5 min). The renewed product (Fig. b) does not exhibit such patterns, indicating that the advection correction (see Sect. ) is effective.

The annual precipitation maps (Fig. a and c) reveal all the types of artifacts found in Fig. , confirming that clutter is a large issue for the old radar product, whereas it is mainly limited to a single spoke-like artifact caused by interference for the renewed radar product, resulting in accumulations of 1420 mm or higher.

Fig.  shows the annual precipitation maps for the old and renewed radar products and the corresponding E-OBS interpolated rain gauge product. The old radar product severely underestimates the precipitation, especially outside the Netherlands, where no gauge accumulations have been employed in the adjustment. Underestimation is less severe for the renewed radar product. Precipitation gradients over the Netherlands roughly coincide with the E-OBS gauge product. The large area with high precipitation accumulations in Germany, that is missed in the old radar product, is captured by the renewed radar product. This is likely the result of the much higher weight of data from the nearest radar in the compositing and, to a lesser extent, the added value of the VPR correction, the rain-induced attenuation correction, and the contribution of the Borkum radar.

The effect of severe beam blockage can be noticed in the east and northeast of the Netherlands for both products, pointing towards the Dutch Herwijnen radar (see Fig. ). For the old product, beam blockage becomes apparent only at longer range from the radar, where the precipitation estimation is mainly based on the data from the Herwijnen radar (other radars have much more weight close to the radar). The severe beam blockage in the easterly direction from the Herwijnen radar, mainly caused by a line of trees, underpins the need for radar site protection. The effects of beam blockage over the northeast of the Netherlands, mainly caused by orography by the Utrecht Hill Ridge , is decreased for the renewed radar product due to the mitigating effect of the addition of data from the German radar in Borkum (Fig. a).

Scatter density plots and metrics of radar versus rain gauge precipitation accumulations are employed to assess the overall performance of the radar product for the Netherlands. For both 1 h and daily precipitation accumulations, the renewed radar product performs better than the old radar product (Fig. ). Notably, the underestimation decreases from ∼17 % to ∼7 % for hourly and from ∼24 % to ∼15 % for daily accumulations. But also the values for CV are lower and the values for ρ2 and KGE are higher for the renewed radar product. The improvements for the renewed radar product are more apparent in the daily scatter density plots than in the hourly plots through better alignment along the 1:1 line. The fact that underestimation by the radar is lower in magnitude compared to the automatic rain gauges than in comparison to the manual rain gauges is consistent with the finding reported in , namely that automatic rain gauges are biased low by 5 %–8 % compared to manual rain gauges. Since clutter leads to overestimation and is more severe for the old radar product, it partly compensates for the mean underestimation. This implies that the underestimation for true precipitation events in the older radar product is actually larger than indicated in Fig. . This is confirmed by the many outliers for (near) zero daily gauge accumulations (and a few outliers for 1 h accumulations) that are not visible anymore in the renewed radar product. Hence, the improvement in the bias may be even larger.

An example of the capability of the renewed radar product for real-time precipitation monitoring is shown in Fig. . Summer storm Poly struck the Netherlands on 5 July 2023 with locally more than 25 mm in 24 h (Fig. a and b). Daily precipitation patterns and accumulations from the radar product match quite well with those from an interpolated rain gauge product, based on validated data from 319 rain gauges from the Netherlands. Only the radar product is available in real time and provides much more spatial detail and can hence capture the counterclockwise rotational movement of the low pressure area (Movie S1 in the Supplement). The scatter density plot shows a fairly high correspondence of the daily radar accumulations with the manual rain gauge accumulations from the Netherlands. It is expected that the old product performs less well primarily because of the lack of attenuation correction.

The performance of daily radar precipitation accumulations is quantified for 1008 (old radar product) and 1210 (renewed radar product) rain gauge locations employing independent daily gauge accumulations from networks in Belgium, France, Germany, Luxembourg, the Netherlands, and the United Kingdom at their default measurement interval (Fig. ; note that this does not include the KNMI automatic gauges used for adjustment). Underestimation is severe for the old radar product (Fig. a), especially when the distance to the nearest employed automatic gauge from the Netherlands and to the nearest radar is long (Fig. a). The renewed radar product displays less underestimation over the land surface of the Netherlands, southwest Belgium, and northwestern Germany (Fig. b). The VPR correction, the rain-induced attenuation correction, and the much higher weight of the nearest radar in the compositing, likely play an important role here, as well as the addition of the German radar Borkum in the northeast of the radar product domain. Notable is the overestimation for more gauge locations in Germany for the renewed radar product. The relative bias in parts of Germany is even closer to zero than in the Netherlands. Perhaps this is also related to different radar hardware calibration. Another explanation is that the German Weather Service has already corrected their radar data for attenuation, so it has been corrected for attenuation twice (although this also holds for the old product). This was not done on purpose, but only because it was assumed that attenuation correction had not been applied yet. Since its application was not mentioned in the metadata, this indicates the importance of provenance in the radar metadata. This double attenuation correction will likely result in overestimation in case of (strong) convective rainfall. The overall effect is expected to be limited, mainly because the contribution to the composite of pixels that have undergone relatively strong attenuation correction is limited due to the reduction of the quality index associated with the attenuation correction (Eq. ). Similar to the annual precipitation accumulations shown in Fig. , the influence of the severe beam blockage to the east of the Herwijnen radar is apparent in the relative bias.

Compared to the old radar product, the values for ρ generally increase (Fig. c and d) and the values for CV generally decrease for the renewed radar product (Fig. e and f), confirming the improved performance for precipitation estimation. For the old radar product, values for ρ are typically 0.8 or higher for the land surface of the Netherlands, but rapidly decrease outside the Netherlands to values below 0.75. For the renewed radar product, values for ρ are generally 0.9 or higher for the land surface of the Netherlands, and this is even found for regions in Belgium and Germany that are close to the Dutch border. At farther distances from the Dutch border, values decrease to values below 0.75. Similar behavior is found for the CV. To conclude, the quality of the renewed radar product is much better for the land surface of the Netherlands and Belgian and German regions close to the Dutch border.

The results shown so far are based on all available data, i.e., without thresholding. Here, the quality of the radar precipitation accumulations for a range of rain gauge precipitation threshold values is examined. Figs.  and  show the values for the relative bias, ρ, and CV, and the number of radar–gauge pairs as a function of this threshold value (step size of 1 mm) for both products for 1 h accumulations and daily accumulations, respectively.

The underestimation becomes generally larger for increasing gauge threshold values, but is less severe for the renewed radar product and fairly constant in the 5–15 mm threshold range (Fig. a) or the entire threshold range (Fig. a) for the renewed product. This is possibly caused by the attenuation correction. Underestimation is only fairly constant for the old product for daily threshold values up to ∼20 mm. The value for ρ usually decreases with increasing threshold value, with the renewed product performing best (Figs. b and b), except for 1 h threshold values above ∼13 mm, for which erratic behavior is found, likely related to the small sample size (Fig. d). The value for CV only strongly decreases from 0 to 1 mm, and remains fairly stable for larger threshold values. Performance is similar for the old and renewed radar products for 1 h threshold values ranging from 0 to ∼10 mm, and slightly better performance for the old product for threshold values larger than ∼10 mm (Fig. c). For higher gauge threshold values, the old product outperforms the renewed product. This is possibly related to the fact that both products cover a different 1-year period. In contrast, for daily accumulations, the value for the CV generally decreases with better performance for the renewed product for threshold values ranging from 0 to ∼25 mm, and similar or slightly better performance for the old radar product for the larger threshold values (Fig. c). Finally, the number of radar–gauge pairs is generally (slightly) higher for the renewed product and comparable or lower for the highest threshold values (Figs. d and d).

To summarize, for the renewed radar product the underestimation is approximately 5–20 percentage points less than for the old radar product for the full range of 1 and 24 h threshold values. The values for ρ are (slightly) higher up to the more extreme 1 or 24 h threshold values for the renewed product. The values for CV are equal (1 h) or lower (24 h) up to the more extreme threshold values and become higher (1 h) or equal (24 h) for the most extreme threshold values for the renewed radar product.

To study a possible seasonal dependence of the quality of radar precipitation accumulations, Fig.  shows scatter density plots of the daily accumulations (see also Fig. c and d) for 5-month winter (November–March) and summer (May–September) periods. The following conclusions are drawn:

The precipitation underestimation is clearly higher in the winter period compared to the summer period, with a much larger difference of ∼15 percentage points for the old radar product compared to the ∼8 percentage points for the renewed radar product (Fig. ).

The Dutch real-time gauge-adjusted radar QPE product has been available since 19 December 2018 and provides 5 min accumulations every 5 min on a ∼1 km2 grid covering the Netherlands and the area around it. Major changes in processing became operational on 31 January 2023, such as clutter removal through a fuzzy logic algorithm and rain-induced attenuation correction on Belgian and Dutch radar data, VPR correction on all radar data, quality-based merging of radar reflectivities from different scans and radars, and a spatially variable instead of a mean-field bias adjustment with rain gauge accumulations. The last year of the old dataset and the first year of the renewed dataset were evaluated.

For the renewed radar product, clutter that was present in the old radar product is effectively removed, showing the efficient application of the fuzzy logic clutter algorithm on Belgian and Dutch volumetric radar data and of the Gabella clutter filter on 2D Cartesian data per radar. Applying additional algorithms, such as a satellite cloud (type) mask and a static clutter mask , do not seem to be necessary.

The influence of beam blockage due to obstacles and orography is also apparent in maps of annual accumulations, exceedance probabilities, and metrics at gauge locations. A beam-blockage correction algorithm will likely lead to further improvements of the product.

The evaluation against independent daily rain gauge accumulations generally reveals a clear improvement for the renewed radar product in terms of the relative bias over the Netherlands (10 percentage points less underestimation), Pearson correlation coefficient, and CV. Improvements are also found far away from gauges that are used in the adjustment, likely pointing to the added value of the VPR correction, rain-induced attenuation correction, quality-based compositing, and the use of more low-level data, as well as to the contribution of data from an additional German radar. At the same time, the need for incorporating rain gauge accumulations outside the Netherlands is clear.

Generally, extremes are better captured by the renewed radar product over the Netherlands: underestimations are lower and Pearson correlation coefficient values are higher. Extremes are very important for many of the applications that use radar-based QPE, so this is an especially encouraging result. Precipitation underestimation is stronger in the winter period than in the summer period for both products. Quality improvement in the renewed radar product with respect to the old radar product is strongest in the winter period. This is likely due to the VPR correction and the fact that the quality-based merging algorithm results in using more low-level data.

The importance of the real-time radar precipitation product expands beyond (near) real-time applications, since it forms the basis for the early and final reanalysis products after reprocessing. These two products have a longer latency of ∼1 d and a few weeks, respectively, but their accuracy is higher, since they are enriched with daily manual rain gauge accumulations. Archives of real-time radar precipitation products can also be useful for training and evaluating (machine learning) nowcasting algorithms. The KNMI gauge-adjusted real-time precipitation product is updated every 5 min and is the input for the pySTEPS deterministic nowcasting product and will be used for an upcoming seamless ensemble forecasting product. Hence, any improvements in this precipitation product directly lead to improved nowcasts, forecasts, and early warnings.

Continuous efforts are needed to further improve the real-time radar precipitation product. The KNMI Radar Team consists of both software developers and researchers, and has adopted the Agile way of working . This allows for fast deployment of improvements to algorithms and addition of new data sources. This makes this real-time QPE product a living dataset. The research carried out on product improvement and evaluation is done in a shared environment, which greatly facilitates collaboration within the team. End users are requested to provide us with feedback on the quality of our product for their specific use case (radar@knmi.nl).

The evaluation of the renewed radar product is assumed to be representative for the current radar product, although already some changes or improvements have been implemented in the renewed radar product after the end of the evaluation period (31 January 2024):

Data from the German radar in Neuheilenbach have been added since 9 September 2024, providing better coverage (Fig. a).

In addition, recent (ongoing) research to show the potential for improving the real-time radar precipitation product includes the following:

Use the differential phase shift to improve precipitation estimation .

Future potential developments are as follows:

Beam-blockage correction employing a digital elevation model. Blocked sectors from the affected elevation scan are assigned a lower weight. Rainfall retrieval through the differential phase can also help to further mitigate beam blockage, although it still does not sample the blocked part of the measurement volume.