The radar datasets are based on the composite product “instantaneous surface rain rate”, as downloaded from the EUMETNET OPERA Data Centre (ODC or Odyssey) from the period 1 September 2019–31 August 2020. They have a grid cell size of 2 km × 2 km and are available every 15 min. The median and average number of contributing radars in the period 1 September 2019–31 August 2020 is 154. Algorithms to take out non-meteorological echoes, including Doppler clutter filtering, and to correct for beam blockage have been applied by NMHSs and/or OPERA. Different kind of radars are employed with, for instance, single-polarisation and dual-polarisation radars, and most radars operate in the C-band frequency range. Radars provide data every 5 or 10 min, where the most recent data are employed in the OPERA 15 min composite product. OPERA converts the horizontally polarised radar reflectivity factors to instantaneous surface rain rates using the Marshall–Palmer Zh–R relation (Zh=200R1.6). The OPERA radar data and the applied algorithms are described in more detail by and .

Here, two OPERA-based clock-hour radar precipitation datasets are utilised. The processing of these datasets and its characteristics are provided by . For each clock hour and grid cell, accumulations are only computed in the case of full availability of the underlying 15 min data. Both datasets have undergone additional removal of non-meteorological echoes by means of two statistical methods and a satellite cloud type mask. The latter sets rain rates to zero in the case of semitransparent clouds or cloud-free areas. Not all of these steps can be applied in real time, but applying them yields a dataset which has undergone the same processing as the EURADCLIM dataset (see below) that is used in comparisons. Moreover, these steps only consider the removal of non-meteorological echoes and not the precipitation retrieval itself. The first dataset is an unadjusted OPERA radar dataset, called “OPERA” in this paper and corresponding to “Gabella + CTM + static filter” in . The second dataset is obtained by merging with disaggregated 1 h ECA&D rain gauge accumulations employing a local mean-field bias adjustment succeeded by a spatial adjustment. For this, the original daily gauge accumulations are disaggregated to 1 h accumulations using the 1 h and 24 h OPERA radar accumulations. An important reason for the disaggregation is the fact that the measurement interval of daily rain gauge accumulations is not the same for the European NMHSs, which is reflected in ECA&D as the NMHSs are the main source of data. A daily adjustment would therefore be less appropriate . This publicly accessible climatological gauge-adjusted product is called “EURADCLIM” . A flowchart of the radar and gauge data processing for OPERA and EURADCLIM can be found in . From these 1 h precipitation accumulations, 24 h (every clock hour) and annual precipitation accumulations are derived in the case that a grid cell has at least 83.3 % availability. The combined radar–gauge availability of daily and 1 h precipitation accumulations is usually high (Fig. a–b).

Daily precipitation time series from 6678 rain gauges were obtained mid-June 2022 from the ECA&D (https://www.ecad.eu, last access: 7 February 2024) project for the period 1 September 2019–31 August 2020. For most regions, the combined availability of radar and gauge data is at least 90 % (Fig. a). Figure c displays the distance from a radar grid cell to the nearest gauge assuming full availability of the 6678 gauges. It reveals a large spatial variability of the gauge network density. The median and mean distances for a radar grid cell to the nearest rain gauge are 42 and 92 km, respectively. The relatively large difference between median and mean distance is primarily caused by areas with low rain gauge network densities (e.g. above sea). Three radar precipitation datasets will be evaluated with the ECA&D rain gauge dataset. It will also be merged with 1 h radar accumulations in case of the EURADCLIM dataset. The ECA&D team has applied quality control on the rain gauge data. Quality control has often also been applied by the NMHS that has provided the data. More details can be found in . The quality of EURADCLIM is lower for areas with low ECA&D gauge network density. Specifically for this 1-year period, some gauge data are not available for the full period. The most notable example of this is that the data from most stations in the United Kingdom end on 31 December 2019 (i.e. only 4 out of 12 months are available).

A Netatmo PWS contains an indoor and an outdoor module measuring air temperature, relative humidity, air pressure, and some additional indoor variables. A wind and rain module can be added. The rain gauge is of the tipping-bucket type containing a collection funnel with a ∼ 13 cm cross-section. A radio connection is used to send the number of tips to the indoor module, which transfers the PWS measurements to the Netatmo platform every ∼ 5 min. Data are visualised at the publicly accessible Netatmo Weathermap (https://weathermap.netatmo.com/, last access: 7 February 2024) but can also be accessed by weather station owners via smartphone or tablet. The default tipping-bucket volume, and hence measurement resolution, is 0.101 mm. This can be altered by a weather station owner using a calibration feature. The manufacturer's specifications report a measurement range of 0.2–150 mmh-1 and an accuracy of 1 mmh-1 . Typical sources of error in (PWS) rain gauge observations are described in , and and references therein.

A 1-year PWS rain gauge dataset of ∼ 5 min accumulations was obtained from Netatmo. Their locations are shown in Fig. b (after applying part of the QC; see Sect. ). The combined radar–gauge availability is generally at least 70 %, averaged over square lattices of nine radar grid cells a side. Based on the individual PWS time series, the median and mean availability is 82.8 % and 69.9 %, respectively. One of the reasons for this relatively low mean could be a strong increase in the number of PWSs over the course of the 1-year period (∼ 23 %). The total number of PWSs is more than 10 times higher than that for the ECA&D rain gauge dataset. The median and mean distances for a radar grid cell to the nearest rain gauge are 28 and 61 km, respectively. As a consequence, the distance to the nearest gauge is generally much shorter compared to the ECA&D dataset but is still quite variable in space (Fig. c and d). It is apparent from this figure that the network density is much higher in regions where the ECA&D gauge network density is (relatively) sparse (e.g. parts of the Iberian Peninsula and eastern Europe), or where rain gauge data are not shared with ECA&D (e.g. Bulgaria). For some regions, the Netatmo gauge network density is lower, though (e.g. in northern Scandinavia). Note that the PWS network tends to be of high density in the areas where the population density is high, whereas it is more difficult to find locations for NMHS gauges in urban areas complying with World Meteorological Organization regulations.

A pan-European gridded dataset of daily mean air temperature was obtained (E-OBS version 26.0e) to evaluate the performance of merged radar datasets for lower and higher temperatures at the location of ECA&D rain gauges. This dataset is derived from ECA&D station observations of 2 m air temperature that are interpolated onto a 0.1∘ grid . Because not all stations with rain gauges have temperature sensors, the E-OBS dataset was employed to derive temperature for each rain gauge location.

The PWS measurements are preprocessed to obtain precipitation accumulations at regular 5 min intervals. This is followed by QC, where for each PWS a neighbour list of PWSs is constructed according to . The neighbour selection has been slightly modified: only the nearest stations (up to a maximum of 20) within a radius of 10 km are selected instead of all stations within a 10 km radius. The 20 nearest stations are expected to be more representative for QC of rainfall at the considered gauge location.

Next, the faulty zeroes and high-influx filters are computed and applied to the 5 min data. Faulty zeroes can be caused by obstruction of the tipping-bucket mechanism or other malfunctioning of the gauge. Unrealistically large accumulations (high influx) can be caused by, for instance, sprinklers, tilting of the gauge or cleaning by pouring liquids into the gauge. The filters and their parameter values have been tested on a 1-year Netatmo rain gauge dataset from the Amsterdam metropolitan area, the Netherlands, from 1 June 2016–31 May 2017 . The basic principle of the faulty zeroes and high-influx filter is that rainfall is correlated in space. If fewer than five neighbouring stations are available, there is insufficient basis to attribute a faulty zero or high-influx flag, and the value is kept (so-called flex filtering). This is done to keep rainfall values in areas with lower gauge network densities. Subsequently, only PWS clock-hour values are computed from the 5 min accumulations in the case that at least 10 out of 12 intervals have valid data. provide a more detailed explanation of the faulty zeroes and high-influx filters, including flowcharts.

Radar–gauge pairs are obtained by collocating the PWS accumulations with the unadjusted OPERA radar accumulations. Only the 1 h PWS accumulations are used for merging if both the unadjusted radar value and the Netatmo gauge value are larger than 0.25 mm, which leads to a strong decrease in the number of radar–gauge pairs. The reason for this is twofold: it strongly reduces the computational time of the merging, and the radar data also act as an additional QC. When precipitation is weak or non-existent according to PWS or radar, the PWS values are considered erroneous. This makes application of the faulty zeroes filter largely redundant, except that it results in ∼ 4 % fewer 1 h PWS values. Both the faulty zeroes and the high-influx filters are applied to 5 min data, which may result in a more refined removal of erroneous values compared to QC on 1 h data. In addition, the 0.25 mm radar threshold can also remove outliers (high-influx) in PWS gauge values. Note that employing unadjusted radar data is justified for the thresholding because it is meant to discern between low and high accumulations, for which high accuracy is not needed. Finally, the effect of removing low values on the outcome of the merging is likely small.

Only radar–gauge pairs with 1 h PWS gauge accumulations ≤ 150 mm are kept, which corresponds with the manufacturer's specifications . Next, the unadjusted OPERA radar data are employed for additional QC (radar QC). Only those 1 h radar–gauge pairs are kept if the ratio of a PWS gauge and OPERA accumulation is lower than 12 and the PWS value is at most 80 mm higher than the OPERA value. Finally, the 1 h PWS gauge accumulations are compared with the gauge-adjusted EURADCLIM dataset, to obtain a default bias correction factor, methodologically similar to . This introduces some dependence, but note that only a constant default bias correction factor is obtained over entire Europe and over a full year. All 1 h PWS gauge accumulations are multiplied by this bias factor of 1.063 to compensate for underestimation.

The merging method is extensively explained by for the EURADCLIM dataset. Here, the same approach is followed, except that the first step of the merging method, the local mean-field bias adjustment, is discarded. The necessity of this is much lower because of the much higher gauge network density. Moreover, it resulted in outliers. The basis of the resulting merging algorithm is Barnes' objective analysis . An adjustment factor is calculated for each radar grid cell and 1 h interval. For a given grid cell, only radar–gauge pairs within a given distance are taken into account for the computation of the adjustment factor only the short range component is employed;. The value for this distance is the range of an isotropic spherical variogram model based on a 30-year rain gauge dataset from the Netherlands from the period 1979–2009 . This range varies as a function of day of year. The adjustment factor field is applied to the OPERA 1 h accumulations to obtain the OPERA-Netatmo merged radar dataset. The median and mean number of radar–gauge pairs used in the merging, so after QC and thresholding, is ∼ 2400 and ∼ 3300, respectively (computed over the ∼ 99.5 % of 1 h intervals in the 1-year period that have at least one valid radar–gauge pair). Note that the median number of radar–gauge pairs is ∼ 6 times larger than for the EURADCLIM dataset, which uses ECA&D rain gauge data. The maximum number of radar–gauge pairs for a given hour is 22 866.

Five statistical metrics are employed to evaluate the performance of radar precipitation accumulations. Here, a residual (Rres) is the radar accumulation minus the ECA&D gauge accumulation. The relative bias is compared to the corresponding ECA&D gauge precipitation accumulations: 1Rel.  bias=R‾resR‾gauges×100=∑i=1nRres,i∑i=1nRgauges,i×100, with n the total number of values and Rres,i the ith residual, “radars” being one of the radar datasets and “gauges” being the point rainfall accumulations from the ECA&D dataset. The residual standard deviation is divided by the mean ECA&D gauge precipitation accumulation (i.e. the coefficient of variation, CV): 2CV=1n-1∑i=1nRres,i-R‾res2R‾gauges.

The Pearson correlation coefficient (ρ) or its squared value (ρ2; the coefficient of determination) is 3ρ=cov(Rgauges,Rradars)s(Rgauges)×s(Rradars), with s the sample standard deviation. The mean absolute error (MAE) is 4MAE=∑i=1n|Rres,i|n.

In some cases, leave-one-out statistics are calculated for an independent evaluation of the EURADCLIM dataset. For these leave-one-out values, the adjusted 1 h radar precipitation accumulation is calculated for a given gauge location without employing it in the adjustment. This is performed for each gauge location separately.

The daily and 1 h radar precipitation accumulations are verified against ECA&D rain gauges in Tables  and , respectively, to assess the effect of the adjustment employing PWS gauge data. No QC implies that the quality control in Sect.  has been omitted, except that the 1 h PWS accumulations are used for merging if both the unadjusted radar value and the Netatmo gauge value are larger than 0.25 mm. When no QC is applied to Netatmo gauge data, the performance of the merged radar dataset is already clearly better than that of the unadjusted OPERA dataset. The underestimation strongly decreases. Values for ρ, MAE and CV improve, especially for larger precipitation accumulations (except for CV for 1 h accumulations). Using quality-controlled Netatmo accumulations in the merging leads to an additional improvement. The relatively good performance when no QC is applied could be attributed to the quality of PWS gauge data and to the merging algorithm acting as a kind of quality control. Only radar–gauge pairs are used in the merging for which radar and PWS observe more than 0.25 mm. Moreover, a spatial adjustment factor is computed by distance-weighted averaging of radar and PWS values, which can average out outliers.

Findings are summarised for daily precipitation with QC (Table ): (1) the strong average underestimation of ∼ 28 % diminishes to ∼ 3 %; (2) more extreme daily precipitation underestimates are greatly reduced but not eliminated; (3) correlation increases, especially for more extreme precipitation; and (4) the values for CV and MAE clearly decrease for all thresholds. The scatter density plots (Fig. a–c) align much better along the 1:1 line compared to the unadjusted dataset. Its performance is lower than for the EURADCLIM dataset (Fig. 3d), though. This is partly caused by the fact that the same gauges used for verification are part of the EURADCLIM dataset, and hence this verification is not independent. The scatter density plots also reveal how quality control has a positive impact: a group of large precipitation accumulations in the case of lower gauge accumulations is removed.

For hourly rainfall (Table ) the underestimation decreases from ∼ 36 % for OPERA to ∼ 12 % for the OPERA-Netatmo dataset with QC, if the ECA&D gauge accumulation is above 0.25 mm. The improvement in the value of ρ is moderate. The value of CV only improves for a gauge threshold of 0.25 mm. The values for MAE show a clear decrease. The scatter density plots for 1 h accumulations (Fig. ) show that for the merged radar datasets the values are closer to the 1:1 line. Again, the QC helps to remove the group of high merged OPERA-Netatmo values for low ECA&D gauge accumulations. Although the impact of QC on overall statistics is small, this is an important improvement. The values for CV and ρ2 for the QC'ed OPERA-Netatmo dataset are (relatively) close to those from EURADCLIM. Note that leave-one-out statistics have been used for EURADCLIM, which is an independent verification. The only dependence for all four datasets is the use of 1 and 24 h unadjusted OPERA radar accumulations to disaggregate daily to hourly gauge accumulations.

For each ECA&D rain gauge location a spatial verification is performed on 1 h accumulations for the unadjusted and Netatmo-adjusted (with QC) OPERA radar datasets (Fig. ). The quality of the unadjusted radar dataset is spatially variable, especially for CV and relative bias. The values for ρ are reasonably high for the OPERA dataset but clearly improve after merging. The improvement appears higher than found in Table  because of the discrete and relatively wide classes used in Fig. . Locally, clear improvements in the values for CV are found, although these can still be quite high. The relative bias strongly improves for the merged OPERA-Netatmo dataset compared to the unadjusted OPERA dataset and its spatial variability diminishes. Regionally, lower Netatmo gauge network densities may be a reason for poorer performance (e.g. Poland, Romania).

The spatial verification reveals severe underestimations for some regions, especially in Scandinavia. This may be related to colder climates. The more frequent occurrence of solid precipitation may result in undercatch or delayed measurements by PWS gauges because precipitation is not melted by a heating device. The manufacturer's specifications even report an operating temperature from 0 to 50 ∘C . To study a possible relationship between performance of merged radar precipitation accumulations and precipitation type, the evaluation is split into two: a group with lower (< 5 ∘C) and higher (≥ 5 ∘C) daily average temperatures from the gridded E-OBS dataset. Because only daily averages of temperature are available, a relatively high threshold of 5 ∘C has been used to account for the possibility of 1 h solid precipitation occurring at temperatures below this daily average. Scatter density plots presented in Fig.  reveal a much more severe underestimation by the OPERA dataset at lower temperatures than at higher temperatures. This is expected given the generally higher influence of the vertical profile of reflectivity in radar QPE and the use of retrieval relations for rainfall in snow conditions. Much better results are obtained for higher temperatures. The same holds for the Netatmo-adjusted OPERA dataset. For lower temperatures, the underestimation is less severe than the unadjusted OPERA dataset, but still a strong underestimation of 27 % is found. This is 21 percentage points lower than the underestimation for higher temperatures, suggesting that the quality of Netatmo rainfall measurements declines for lower temperatures when the probability of solid precipitation is higher. The EURADCLIM dataset still reveals a better performance for higher temperatures, but the difference in relative bias is small between the two groups. Rain gauges are generally expected to have either a heating device (in case of automatic gauges) or instructions to melt precipitation (in case of manual gauges). Moreover, rain gauges operated by NMHSs often have a larger reception cone volume than PWS rain gauges, enabling them to collect more solid precipitation, which can gradually melt, even if the gauges are not heated.

Figure  shows annual precipitation sums for the OPERA dataset without and with Netatmo-adjustment. A comparison against the EURADCLIM dataset is also presented to assess the performance over the entire radar grid. The annual precipitation for the merged OPERA-Netatmo dataset is much larger than the unadjusted OPERA precipitation sum (Fig. a and b). The ratio between the merged OPERA-Netatmo dataset and the EURADCLIM dataset (Fig. d) is between 0.8–1.1, confirming its quality. For some regions, a large underestimation with respect to EURADCLIM is found. This may partly be related to solid precipitation (Scandinavia), which is not expected to be captured well by Netatmo gauges. The limited spatial extent of the employed adjustment method may also result in an underestimation at locations far(ther) away from radars and rain gauges (e.g. above sea around the Iberian Peninsula and Scotland).

Figure  illustrates the potential of merging with PWS gauge accumulations for heavy-precipitation monitoring. It shows precipitation accumulations for a widespread event over northeastern Spain, with large areas with daily accumulations exceeding 50 mm (Fig. c) and locally more than 150 mm (Fig. f). This event with very heavy rainfall led to a strong rise in discharge of many small rivers in the south of Catalonia, causing four deaths and estimated damage of over EUR 700 000 . The large underestimation by OPERA (Fig. a) mainly disappears for the merged OPERA-Netatmo dataset (Fig. b) given its resemblance with EURADCLIM (Fig. c). The underlying Netatmo gauge values (Fig. e) are in reasonable agreement with the ECA&D gauge values, although underestimations are found (Fig. f). Moreover, the density of the PWS network is higher than that of the ECA&D network (Fig. d), and the data could potentially be available in real time, which is usually not the case for gauges in the ECA&D dataset. Note that for this region, the ECA&D network is rather dense compared to most of Europe, and real-time data from these gauges are employed by the Meteorological Service of Catalonia for merging with radar data. Note that only gauge locations (and values) are plotted if more than 0.25 mm of rain has been reported in (at least) 1 h. Zooming in to a clock hour in this 24 h period, a squall line becomes much more pronounced in the two merged radar datasets (Fig. h and i) compared to the unadjusted OPERA dataset (Fig. g). The Netatmo network provides more measurements (Fig. j), and this results in better sampling of the maximum precipitation within this hour (Fig. k and l). The squall line is more severe for EURADCLIM, though, possibly because of the local mean-field bias adjustment, which precedes the spatial adjustment for EURADCLIM only. It is difficult to tell which dataset has the most realistic 1 h precipitation.

provide gauge-adjusted precipitation estimates for the event presented in Fig.  employing data from four local C-band radars, which are not incorporated in the OPERA-based datasets. Hence, the distance to radars is much shorter than for the OPERA datasets, and precipitation is detected at a lower altitude. present a real-time radar product which includes data from the gauges contributing to ECA&D, as well as a climatological product where the gauge data have undergone additional QC. Both the merged OPERA-Netatmo dataset and EURADCLIM dataset underestimate precipitation compared to the local gauge-adjusted radar products. Hence, Fig.  does not provide the best possible precipitation estimates, but it does confirm that merging with PWS gauge data gives a large improvement in OPERA radar precipitation accumulations.

The potential of PWS gauge accumulations for improving radar precipitation accumulations is expected to be higher in urban areas. It can be more difficult to find an appropriate setup for rain gauges in urban environments. Hence, to comply with World Meteorological Organization regulations, NMHSs typically measure in areas with few or lower obstacles. Therefore, the (automatic) weather stations are typically located outside city centres, often in more rural areas, although open areas can also be found in cities. Note that originally NMHS weather stations were often installed at airports, close to large cities, and are present to date. Although urban rainfall monitoring is relevant, cities generally represent only a small fraction of the land surface of a country. Meteorological networks have been designed to ensure regular coverage of countries. In contrast, urban areas generally have the highest PWS network densities. Figure  illustrates a 1 h heavy-precipitation event for the city of Brussels, Belgium. There are more than 80 Netatmo rain gauges in this area, whereas there are only 2 ECA&D gauges. The precipitation accumulations increase for the EURADCLIM dataset (Fig. c) with respect to the unadjusted OPERA dataset (Fig. a). The increase is much stronger for the OPERA-Netatmo dataset (Fig. b), and it has the highest 1 h accumulation, 20 mm. The Netatmo rain gauges measure high accumulations (Fig. e). The ECA&D rain gauges only provide limited coverage, and their values are lower than those from the closest PWS gauges (Fig. d–f). The high network density of PWS gauges helps to capture the localised heavy precipitation (Fig. e), which is hence reflected in the OPERA-Netatmo dataset (Fig. b). The PWS gauges can not be considered to be the ground truth. But the fact that most gauges observe high(er) accumulations gives confidence.

The ECA&D rain gauge dataset is used as a reference dataset because of its established quality and its independence. The disaggregated 1 h data are less independent because the OPERA radar data have been used for disaggregation. This may partly explain that for the OPERA-Netatmo dataset of 1 h accumulations, ρ only improves moderately, and the value of CV only improves for a gauge threshold of 0.25 mm. Since EURADCLIM already uses OPERA, there is also some dependence with the merged OPERA-Netatmo dataset. A disadvantage of using gauges as reference is the mismatch in sampling areas (a few squared decimetres (dm2) for gauges and 4 km2 grid cells for OPERA) between the two datasets. This and other representativeness errors can be large when radar and gauge accumulations are compared , especially for sub-daily accumulations. Hence, differences between (merged) radar and gauge accumulations are not solely caused by sources of error in QPE.

The comparison with the performance of the radar dataset EURADCLIM is provided to put the quality of the merged radar–PWS dataset into perspective compared to a climatological dataset employing official rain gauges. The direct comparison with EURADCLIM for annual precipitation and case studies allows for a high-resolution spatial comparison. The performance of the merged OPERA-Netatmo dataset could also be compared to that of the near-real-time OPERA-based radar rainfall product for the European Rainfall-InduCed Hazard Assessment (ERICHA) system . This would give more insight into the potential of crowdsourcing for improving near-real-time pan-European radar data. This is outside the scope of this paper but could be investigated in future research.

This study should be seen as a starting point for merging PWS and radar precipitation accumulations. As shown in Fig.  and Tables  and , the QC of PWS rain gauge data has a positive effect but with a relatively limited magnitude. A possible explanation for this is that the merging algorithm already acts as a QC by reducing the impact of outliers. Multiple radar–gauge pairs are employed to compute the adjustment factor for a given pixel, thus limiting the influence of a single gauge. Moreover, only radar–gauge pairs are used for which radar and PWS both reported more than 0.25 mm in 1 h, which helps to reduce erroneous values. The quality of the merged product could potentially be further improved by applying additional QC to the PWS gauge data, for instance, by applying a station outlier filter and a dynamically updated bias correction factor . Note that the radar-based part of the QC developed in this study (limits on the PWS / radar ratio and maximum difference between PWS and radar accumulations) implicitly acts as a station outlier filter. Moreover, the performance of this radar-based filter does not depend on the local PWS network density. The suggested station outlier filter compares time series (by computing correlations) from neighbouring stations from at least the last 2 weeks (and longer if there has been insufficient precipitation in this period) to detect deviations in local rainfall dynamics. These deviations could be caused by, for example, incorrect station locations.

Network density can affect the performance of a QC method and whether a QC method is suitable at all. developed a radar version of the QC developed by and applied it to different rain gauge datasets, including a 1-year Netatmo PWS gauge dataset over the Netherlands. This version of the QC algorithm performs well in the case of lower gauge density. Depending on the chosen setting (flex or strict filtering), results are not as good or similar for an area with high network density (the Amsterdam metropolitan area) compared to applying the original PWS-based algorithm. Moreover, this radar-based QC algorithm removes twice as many observations. But its potential for use in regions with lower PWS network densities, such as commonly found in the European dataset, warrants further investigation.

Also note that the default parameter values from for QC, based on Netatmo rain gauge data from the Netherlands, are applied to data from the whole of Europe. Different parameter values could be tested on a European scale or could be regionalised to take into account local precipitation climatology and local network characteristics. For instance, the range with neighbouring PWSs is set to 10 km but could be longer for stratiform rainfall. The same holds for regions with low PWS network density.

Considering other QC algorithms, specifically the spatial consistency test (SCT) against neighbours could be relevant for application to the PWS dataset used in this study because it assesses the likelihood of an observation with stricter testing in case of higher station density . The temporal consistency test and the spatial conformity test from could be interesting to apply on PWS rain gauge data. In particular, the assignment of a quality index to each gauge observation is methodologically interesting. Only observations with a sufficiently high quality index could be selected for merging with radar data, or the quality index could be used as weighting factor in the radar–PWS merging. The QC in this study mainly consists of inter-station checks and the use of an auxiliary source: radar data. The development and application of time series analyses from single stations (intra-station checks) may be explored to improve QC. Finally, NMHS rain gauge data could also be taken into account in the quality control of PWS rain gauge data.

The quality of PWS gauge observations and sources of errors in radar QPE can be highly variable in space and time. This results in spatial variability of the relative bias in 1 h merged precipitation accumulations (Fig. ). The quality of PWS gauge data could be improved by computation of a dynamically updated bias correction factor per gauge, employing neighbouring PWSs. This is intertwined with the station outlier filter , as is the case for the radar version . An extension to this could be to remove a gauge if the bias correction is smaller than 0.5 or greater than 2 . The approach by , who apply bias correction via quantile mapping employing rain gauge observations from the German weather service, is also worth investigating.

A seasonally and regionally variable bias correction factor could be applied to the merged radar–PWS dataset using the methodology of . This bias correction field can be derived from comparing with a reference (the EURADCLIM dataset) from previous years. This approach assumes that the quality of EURADCLIM is high across Europe and stationary in time. Improvements in OPERA precipitation data would require recalculation of bias correction factors. Applying this approach seems challenging, given the large improvement found for daily unadjusted OPERA accumulations for the warm seasons from 2015 to 2018 .

Another possible improvement in the merged radar–PWS dataset could be obtained by the application or development of other adjustment methods see, for example, or, for instance, by taking into account the local gauge network density and local precipitation climatology (instead of the climatology from the Netherlands) to select radar–gauge pairs employed to adjust radar precipitation for a given grid cell. Improving the climatological gauge-adjusted radar dataset EURADCLIM by merging the PWS and the ECA&D rain gauge data at once is another avenue that can be explored. The merging algorithm employed in this study can use quality information about the gauge observations and could give lower weight to lower-quality gauges (in this case the PWS gauges).

In addition to Netatmo PWS gauge data, data from other PWS platforms and companies could also potentially be taken into account for merging with radar data. Examples of such platforms are the Weather Observations Website (WOW; free access; https://wow.metoffice.gov.uk/ (last access: 7 February 2024); https://wow.knmi.nl/, last access: 7 February 2024, Kirk et al., 2021; ), the Weather Underground website (https://www.wunderground.com/wundermap, last access: 7 February 2024) , Weathercloud (https://app.weathercloud.net/map#, last access: 7 February 2024), Meteonetwork (https://www.meteonetwork.it/rete/livemap/, last access: 7 February 2024 and https://meteonetwork.eu/en, last access: 7 February 2024), PWSweather (https://www.pwsweather.com/map, last access: 7 February 2024) and (solely for daily precipitation in the United States) the Community Collaborative Rain, Hail & Snow Network (CoCoRaHS; https://www.cocorahs.org/, last access: 7 February 2024) . The sensor quality and setup of PWSs for these data sources can be better compared to Netatmo PWSs. For WOW, station owners can fill in location attributes related to the surroundings, measurement device and timeliness of reporting. These are employed to compute site ratings (https://wow.metoffice.gov.uk/support/siteratings, last access: 7 February 2024), which could be used as part of the QC on PWS data or as quality index in radar–PWS merging. For Meteonetwork, setup guidelines need to be followed for a station to be accepted. Automatic QC is applied to data from Italy .

PWS air temperature measurements or air temperatures forecasts could be used to estimate whether precipitation is solid or liquid and could hence be used to select which gauges to use in merging. This could prevent the influence of undercatch (particularly in high-wind situations) of snow or delayed measurements due to the fact that frozen precipitation is only registered when it melts (see also Fig. ).

Real-time data from official networks of automatic rain gauges are generally much sparser, and a latency of 5–30 min is not uncommon. Several crowdsourced rain gauge networks potentially disseminate data with similar or even shorter latency. Arrangement of data contracts with commercial companies, preferably on a European or even global level, could help make these data available for everyone (either directly or indirectly through merging in precipitation products). In return, these companies would receive valuable feedback on the quality of their data. And the availability of these data could spur further development of QC algorithms by the scientific community that could then be employed by these companies for their platforms.

Commercial microwave link (CML) data could also be considered as a candidate for merging with radar data. The quality control, rainfall retrieval and merging of opportunistic sensing data, including PWSs, CMLs and SMLs (satellite microwave links), are being studied in the OpenSense (Opportunistic Precipitation Sensing Network; https://opensenseaction.eu/, last access: 7 February 2024) COST Action. OpenSense will address the benchmarking of QC algorithms on datasets, allowing for a fair and standardised comparison of QC algorithms. Development of operational merged products based on radar, PWS or CML data is already being investigated by some NMHSs in Europe . For a specific country or region, it may be efficient to focus on the most promising opportunistic data sources: those for which network densities are highest, obtained experiences and results are favourable, and the probability of continuous data access is higher.

Merging opportunistic sensing data with radar data also involves risks and creates dependence on data beyond the control of NMHSs. For instance, PWS sensors and setups could degrade over the course of time, and network densities may decrease. As a result, the quality of merged products could decrease. Methods that are employed to merge such data with radar accumulations should therefore be designed such that they can deal with the varying quality and availability of opportunistic sensors. Another way to try to prevent the decline in quality and number of opportunistic sensors is social engagement of citizen scientists and outreach to increase the number of citizen scientists. provide an overview of ways to achieve this. Another issue that may play a role with the increased use of opportunistic sensors is the decrease in perceived importance of data from official rain gauge networks, potentially leading to abandoning such networks. However, such high-quality station data are indispensable for climate monitoring. This should be kept in mind by agencies operating these networks.