Accurate, timely, and reliable precipitation observations are mandatory for hydrological forecast and early warning systems. In the case of convective precipitation, traditional rain gauge networks often miss precipitation maxima, due to density limitations and the high spatial variability of the rainfall field. Despite several limitations like attenuation or partial beam blocking, the use of C-band weather radar has become operational in most European weather services. Traditionally, weather-radar-based quantitative precipitation estimation (QPE) is derived from horizontal reflectivity data. Nevertheless, dual-polarization weather radar can overcome several shortcomings of the conventional horizontal-reflectivity-based estimation. As weather radar archives are growing, they are becoming increasingly important for climatological purposes in addition to operational use. For the first time, the present study analyses one of the longest datasets from fully operational polarimetric C-band weather radars; these are located in Estonia and Italy, in very different climate conditions and environments. The length of the datasets used in the study is 5 years for both Estonia and Italy. The study focuses on long-term observations of summertime precipitation and their quantitative estimations by polarimetric observations. From such derived QPEs, accumulations for 1 h, 24 h, and 1-month durations are calculated and compared with reference rain gauges to quantify uncertainties and evaluate performances. Overall, the radar products showed similar results in Estonia and Italy when compared to each other. The product where radar reflectivity and specific differential phase were combined based on a threshold exhibited the best agreement with gauge values in all accumulation periods. In both countries reflectivity-based rainfall QPE underestimated and specific differential-phase-based product overestimated gauge measurements.
Detailed surface rainfall information is of great importance in many fields, not only for agricultural or hydrological applications. In the recent past the COST 717 Action entitled “Use of radar observations in hydrological and NWP models” investigated the assimilation of weather-radar-based precipitation in numerical weather prediction (NWP; Macpherson, 2004). Weather radar data have been assimilated in a variety of assimilation systems and models of increasing resolution. At the beginning, latent heat nudging was the most popular technique (Gregorc̆ et al., 2000), while researchers have recently moved towards volume reflectivity assimilation techniques: for example, Schraff et al. (2016) proposed the KENDA (ensemble Kalman filter for convective-scale data assimilation) operator to assimilate reflectivity volume data in the COSMO (COnsortium for Small-scale MOdelling) model. For decades, gauge networks have provided the best reference datasets. The E-OBS 50-year daily European gridded interpolated dataset has been widely used in climatological studies (Cornes et al., 2018). Gauge-based datasets have well-known shortcomings in their low spatial resolution and to a lesser degree temporal resolution. Precipitation data from satellites provide good spatial coverage but still not in very high temporal resolution, especially in higher latitudes (Sun et al., 2018). Polar-orbiting satellites provide better spatial resolution data in higher latitudes, but they are very limited in temporal resolution (Tapiador et al., 2018). What is more, satellite-based precipitation estimates are limited by the accuracy of the estimates. The accuracy of the estimates has a regional dependency and therefore can vary due to the physiography of the study areas (e.g. precipitation climate, land use, and geomorphology) (Petropoulos and Islam, 2017). Now that weather radars have already been used for decades in many countries, their archives are getting long enough to use the data in climate studies (Saltikoff et al., 2019). In the last decade, various studies have used multi-year single-polarization weather radar data successfully in deriving rainfall climatology with high spatio-temporal resolution (Overeem et al., 2009; Goudenhoofdt et al., 2016). However, quantitative precipitation estimation (QPE) with single-polarization C-band radar is strongly affected by attenuation of the electromagnetic wave in heavy precipitation or a wet radome, hail contamination, partial beam blockage, and absolute radar calibration (Krajewski et al., 2010; Cifelli et al., 2011).
All prior shortcomings can be mitigated by the use of dual-polarization
weather radar data. Several studies have shown that rainfall retrieved from
dual-polarimetric radar differential phase measurements outperforms rainfall
estimated from horizontal reflectivity, especially in heavy precipitation
(Wang and Chandrasekar, 2009; Vulpiani et al., 2012; Wang et al., 2013;
Crisologo et al., 2014). Because differential phase measurements tend to be
noisy and less reliable in low-intensity precipitation, Crisologo et al. (2014) and Vulpiani and Baldini (2013) improved the robustness of their
rainfall retrieval technique by employing a combination of horizontal radar
reflectivity
The main aim of this study is to evaluate the potential of using
polarimetric weather radar QPE on long-term warm-season datasets in various
climatological environments. Previous studies in which the benefits of dual-polarimetric radar QPE have been shown are mostly based on selected short
periods or only single events (Wang and Chandrasekar, 2010; Chang et al.,
2016; Montopoli et al., 2017; Cao et al., 2018). While the performance of
the QPE methods can be compared based on short periods as well, only a study
based on long-term data can prove the robustness of a method and suitability
for long-term operational use. The uniqueness of this paper is ensured by
various features. First of all, we have a long 5-year dataset, starting
already from 2011, derived by operational dual-polarimetric C-band weather
radar made by different manufacturers. The dataset is gathered from the
archive of weather radar scans set up for operational surveillance in the
meteorological services. Secondly, the study areas are from heterogeneous
climatologies, the weather radars being located in Estonia and Italy. This is
also the first ever study evaluating weather radar QPE in Estonia. What is
more, we will assess the effect of the radar scan interval as the radar data
scan frequency is 5 and 15 min from Italy and Estonia respectively. The
study analyses result first in a few selected cases. The whole dataset is
analysed at three accumulation intervals of 1 h, 24 h, and 1 month.
Three radar QPE products are generated for comparison: first the horizontal-reflectivity-based product
The paper is organized as follows. Section 2 describes the rainfall estimation datasets from radar and rain gauges and methods used for comparisons. The results are discussed in Sect. 3. In Sect. 4 conclusions are provided.
To estimate the performance of the radar rainfall products, they were compared with gauge accumulations. The study period was limited to the warm season (May–September for Estonia and April–October for Italy). In Estonia, the mean annual precipitation is 649 mm. Precipitation climatology has distinct seasonality, with a maximum in summer (215 mm) followed by autumn (198 mm), winter (128 mm), and spring (108 mm). The summer maximum of seasonal mean precipitation is especially pronounced in the continental part of Estonia (246 mm in Mauri, south-east Estonia) (Tammets et al. 2013).
In Piemonte, close to the radar, the mean annual precipitation is 870 mm having a bimodal distribution with peaks in spring (266 mm) and autumn (255 mm) (Devoli et al. 2018).
Radar-based QPEs have been accumulated to the 1 h duration, and longer durations have been calculated based on these accumulations. Accumulations were calculated by adding subsequent instantaneous radar QPE values without any space–time interpolation. No missing data for radar or gauges were tolerated to prevent underestimation. A threshold of 0.1 mm was set and applied such that both gauge and radar QPE values must exceed this value to make the pair valid.
The quality of the rainfall estimates was estimated by the following
verification measures (where
In Estonia major renewal and automation of the rain gauge network run by the Estonian Environment Agency (EstEA) started in 2003. From 2003 to 2006 the network was updated to automatic tipping-bucket gauges. Starting from 2006 the tipping-bucket gauges were progressively replaced by weighted gauges. This process was finished by the end of the year 2011. By that time there were 33 automatic weighted gauge stations and 27 stations with tipping-bucket gauges. According to the comparative study of parallel measurements of the tipping-bucket gauges and weighted gauges, the latter exhibited much higher quality (Alber et al., 2015). From the end of 2010, the data were recorded with a 10 min interval. Until 2010 the temporal resolution was 1 h. Both 10 min and 1 h data have been saved by EstEA since then, but only 1 h data have been quality-controlled by EstEA staff. Because the 10 min data are not quality-controlled, 1 h gauge data were used in this study as a more reliable ground truth. The offline manned data quality control includes using mainly weather sensor data as an additional source for comparisons but also neighbouring stations and weather radar data on some occasions. Only weighted gauge data were used because of the higher quality of these measurements and to ensure uniformness of the dataset. In this work, eight rain gauges close to Sürgavere, Estonia, are included (Fig. 1). Data have a resolution of 0.1 mm.
Since 1987, Arpa Piemonte, the Regional Agency for the Protection of the Environment, in Piemonte, Italy, has operated a regional automatic gauge network made up of about 380 tipping-bucket gauges. Most of the gauges are heated to avoid solid precipitation accumulation during the cold season. The temporal resolution of the gauges network is 1 min. The Arpa Piemonte weather stations are equipped with CAE PMB2 tipping-bucket rain gauges. Their resolution (0.2 mm) is the amount of precipitation for one tip of the bucket. The working range of measures is from 0 mm to 300 mm/h, with underestimation for high precipitation intensities. Such errors are corrected according to results of the WMO Field Intercomparison of Rainfall Intensity Gauges (Vuerich et al., 2009). An automatic data quality check is run on real-time data, followed by offline manned data validation. In this study, a network subset made of 42 rain gauges close to Turin, Italy, has been considered (Fig. 1). Precipitation measurements range from 2012 to 2016.
Data from C-band dual-polarization Doppler weather radars in Estonia and Italy were used in this study. The weather radars considered in this study are from different manufacturers, in Estonia Vaisala WRM200 and in Italy Leonardo Germany GmbH METEOR 700C radar. Figure 1 illustrates the location of the Estonian radar (Sürgavere) and the Italian radar (Bric della Croce), together with the locations of available rain gauges.
Sürgavere radar, located in central Estonia at an altitude of 128 m a.s.l.,
has been operational since May 2008, but for this study data starting from
2011 were used because the gauge network had been updated by that time. The radar
performs a surveillance volume scan at eight elevation angles (0.5,
1.5, 3.0, 5.0, 7.0,
9.0, 11.0, and 15.0
Study areas (shaded) located in Estonia (upper left zoomed-in area) and in Piemonte, Italy (lower right zoomed-in area). Grey dots denote gauge locations of both the Estonian and Piemonte region and blue dots gauges inside the study area. Blue stars reveal radar locations.
On the Turin hill, at an altitude of 770 m a.s.l., the operational dual-polarization Doppler C-band weather radar Bric della Croce is located. The radar site is in the central part of the Piemonte region: toward west and north at about 20 km the Alps start, with peaks 2500–3000 m above sea level. The radar performs fully polarimetric volume scans, made up of 11 elevations up to the 170 km range, with 340 m range bin resolution. Bric della Croce observations used in the study ranged from 2012 to 2016, whereas observations from 2012 to 2013 have a 10 min interval and from 2013 to 2016 have a 5 min interval. As can be seen from Fig. 1, a circular area around the radar is used in Estonia, but in Italy a rectangular area is used. The reason for this is that orography in Piemonte is very complex, ranging from flat plains in the Po valley (about 100 m a.s.l.) to the Alps up to more than 4000 m a.s.l. The Bric della Croce weather radar is located on Turin hill, which is about 30 km from the Alps. Therefore, the elegant and simple limitation in range by some kilometres from the radar site does not work. To avoid mountainous areas, where there is partial and total beam blocking and where heavy ground contamination increases, a rectangle area, that extends towards flat grounds, has been preferred.
The maximum distance of the gauges to be included in the comparison was limited to a 70 km radius from the radar location in the case of Estonia and up to 30 km distance in Italy. Thus, in Estonia and Italy rainfall data were from 8 and 42 gauges respectively. By limiting data analysis to the warm season, and constraining the maximum radar range, we were able to ensure that radar data were mainly originating from liquid precipitation (hail can also occur), which is required for more reliable rainfall intensity estimation. The possible occurrence of hail was not removed from the data because of the intention to keep additional data processing minimal and to allow for equal comparison of the various QPE methods. In the case of Italy, the applied range limit is also aimed at eliminating uncertainties due to complex orography, like shielding by the mountains, overshooting, and bright-band contamination.
QPEs, based on horizontal reflectivity, are extensively described by
Cremonini and Bechini (2010) and by Cremonini and Tiranti (2018); meanwhile,
The Sürgavere radar specific differential phase product (
To convert reflectivity
Several studies have shown that
In our study rainfall from a combined threshold approach was used for both
weather radars as a third product
Verification results of the test dataset of 1 month (August 2018)
of the radar-based rainfall 1 h accumulation products of Estonia.
The impact of the temporal sampling was analysed using Italian Bric della Croce weather radar second-elevation PPI (plan position indicator) data, which produce a 5 min interval dataset. A degraded dataset of a length of 1 d, 10 October 2020, with a 15 min sampling rate, was created by removing two out of three files. Hourly accumulation was calculated based on both sampling rates, which resulted in a sample size of 253 514. As expected from the comparison of these accumulation pairs, the obtained normalized mean bias was close to zero (0.03), while the correlation coefficient was 0.922, and the normalized mean absolute error was 0.21.
If we compare different skill scores for 1 h QPEs in Estonia and Italy, part of the differences in correlation coefficient and normalized mean absolute error can be explained as being due to different time sampling. Table 2 below summarizes correlation coefficient and normalized mean absolute error values in Estonia and Italy.
Verification of the 1 h accumulation QPE products of Estonia and
Italy and differences without (“Difference”) and with (“Comp. diff”)
compensating for the impact of the temporal sampling. CC and NMAE values are
obtained from Tables 3 and 4.
The 1-month 1 h rainfall cumulative accumulations for Sürgavere
radar data and Jõgeva station gauge data.
Compensating the values obtained in Estonia for loss of correlation (0.078)
and increased NMAE (0.21) due to 15 min time sampling with values
estimated in Italy, it is visible that CC and NMAE are comparable in Estonia
and Italy (last row in Table 2). It is worth noting that after the
compensation is applied, QPE estimated by
In this section radar QPE products are compared with single-location gauge measurements of selected short periods from Estonia and Italy. This allows for evaluating the performance of the radar QPE against gauge measurements from a time-series viewpoint.
The 1-month 1 h rainfall cumulative accumulations for Sürgavere
radar data and Tartu-Tõravere station gauge data.
Figure 2 shows 1 month of precipitation at the Jõgeva station location (60 km away from the radar site) in Estonia with 1 h temporal resolution.
Overall, radar products follow the gauge measurements well, but there are
considerable differences among them. Reflectivity-based product
Gauge and radar accumulations are not always so well correlated as Fig. 3
demonstrates. In this accumulation period, there are rainfall events which
show that gauge values can be both under- and overestimated by radar
products. Rainfall around 11 June 2016 is overestimated by all
radar QPE products, with the smallest overestimation by
Figure 4 illustrates a case from Italy, a comparison of a gauge located
within 30 km distance from the radar to Bric della Croce radar
precipitation estimation products. At the end of the 34 h period, the
specific differential-phase-based product
The 1 h rainfall cumulative accumulations from Verolengo gauge,
located 29 km from the radar, and co-located Bric della Croce radar QPE.
In all selected cases the general behaviour of QPEs is similar. Weather
radar estimations, even when sampled by 15 min interval observations,
follow gauge measurements with good agreement, although the second case from
Estonia illustrated well that a longer scan interval increases the scatter and
particularly with small-scale convective precipitation, for which a minimal
sampling interval is the most beneficial. From Italy, the example case was
much shorter, but the precipitation intensity was higher. In both cases,
The quality of the rainfall estimates is compared at various accumulation intervals. Comparing different intervals can also be useful to point out representativeness issues caused by low radar scan rates. The investigated period covers the years 2011–2018 in Estonia and 2012–2016 in Italy.
First, in this section hourly accumulations are analysed. Hourly accumulations are especially important for small basins and in extreme precipitation climatology analysis. Hourly rainfall maxima can provide valuable data for flash flood nowcasting and other hydrological applications.
Verification of the radar-based rainfall 1 h accumulation
products of Estonia.
Verification of the radar-based rainfall 1 h accumulation
products of Italy.
Scatter plots of radar-based rainfall estimates against rain gauge
observations for 1 h accumulation intervals in Estonia 2011–2018. The
corresponding verification measures are presented in Table 3. The number of
radar–gauge data pairs with eight gauges and accumulations
Table 3 presents the verification results for the hourly accumulation
interval in Estonia. Figure 5 shows the corresponding scatter plots. As can
be seen, the
Nevertheless, it can be seen from the scatter plots that there is a lot of
scatter in the hourly radar accumulations with all products. Mostly, it can
be linked to the low spatial representativeness of the point measurements of
rain gauges. This effect is more pronounced on a short timescale, and it
originates from a scarce gauge network and insufficient radar scan rate.
Small-scale effects like wind drift might also be more influential on a
shorter accumulation period (Lauri et al., 2012). The reason why
From Italian hourly accumulation scatter plots in Fig. 6, it can be seen that
the overall behaviour of the radar products is similar to Estonia, although
from Fig. 6 it can be noticed that of the four highest 1 h accumulations
measured by the gauge, three of them have significantly higher radar
estimates for
Less random scatter is visible in Italian hourly data due to the more
frequent scan strategy.
The 1 h accumulations for Italy, 2012–2016. The corresponding
verification measures are presented in Table 4. The number of radar–gauge
data pairs with 42 gauges and accumulations
Verification of the radar-based rainfall 24 h accumulation
products of Estonia.
Verification of the radar-based rainfall 24 h accumulation
products of Italy.
Table 5 shows the verification results for the daily accumulation interval
in Estonia, while Fig. 7 presents the corresponding scatter plots. As
expected, much less scatter can be seen than on the daily level, but overall,
the results are consistent with the hourly interval verification outcomes.
Using longer accumulation intervals leads to less severe errors as the
longer period compensates for both underestimates and overestimates.
The reflectivity-based product,
The 24 h accumulations for Estonia, 2011–2018. The corresponding
verification measures are presented in Table 5. The number of radar–gauge
data pairs with eight gauges and accumulations
Table 6 shows the verification results for the daily accumulation interval
in Italy, while Fig. 8 presents the corresponding scatter plots.
The 24 h accumulations for Italy, 2012–2016. The corresponding
verification measures are presented in Table 6. The number of radar–gauge
data pairs with 42 gauges and accumulations
Table 7 shows the verification results for the monthly accumulation interval
in Estonia, while Fig. 9 presents the corresponding scatter plots. Compared
to shorter timescales, overall on a monthly scale the correlation of all the
products with gauge accumulations is higher.
Verification of the radar-based rainfall monthly accumulation
products of Estonia.
Verification of the radar-based rainfall monthly accumulation
products of Italy.
Monthly accumulations for Estonia, 2011–2018. The corresponding
verification measures are presented in Table 7. The number of radar–gauge
data pairs with eight gauges is 179.
Table 8 shows the verification results for the monthly accumulation interval
in Italy, while Fig. 10 presents the corresponding scatter plots.
Scatter plots reveal similar characteristics to the daily level accumulations
of the products.
Monthly accumulations for Italy, 2012–2016. The corresponding
verification measures are presented in Table 8. The number of radar–gauge
data pairs with 42 gauges is 675.
In the present study polarimetric rainfall retrieval methods for the fully
operational C-band radars in Sürgavere, Estonia, and Bric della Croce,
Italy, have been analysed. The study focuses on the warm period of the year,
and a long period of multi-year data is used. From Estonia 5 years of
data from 2011 to 2018 has been included; from Italy, the data interval
ranges from 2012 to 2016. Reflectivity data were calibrated following a
self-consistency theory, and measured horizontal reflectivity (
Three radar rainfall estimation products were computed: a horizontal
reflectivity-based product
Time-series comparison revealed that even with a 15 min scan interval, radar is suitable for QPE, at least with more widespread precipitation like stratiform rain. Still, on the shortest accumulation period of 1 h, the scarcer radar data from Estonia had more scatter than data from Italy, where the scan interval was 10 min on older data and 5 min since 2013. As an overall trend, the longer the accumulation period, the less scattering that was visible. The case comparisons also revealed the shortcomings of analysis based only on selected short periods. The performance of the QPE methods then depends on the representativeness of the chosen cases, and results can easily be skewed. Using a dataset with a length of at least several years without preselection provides more robust results and allows for evaluating the operational usability of the methods.
When the three products are compared to each other based on the full length
of 5 years of data, in the case of Estonia, the
Overall the results show that the combined product
In Estonia, the overestimation of
Synoptic patterns could be used as an additional source for classifying the radar accumulations. This would enable the performance of each radar product to be verified for stratiform and convective events. Moreover, it could be used to investigate if frequent scans play a bigger role in convective events than stratiform events as could be hypothesized and to quantify the effect.
For future studies, it would also be useful to calculate probabilities and return periods of extreme rainfall for weather-radar-based rainfall climatology.
The code used to conduct all analyses in this paper is available by contacting the authors. Gauge and radar data used in this study are available by contacting the authors.
TV, RC, PP, and DM directly contributed to the conception and design of the work. TV and RC collected and processed the various datasets and wrote the original draft with input from PP and DM. All authors reviewed and edited the final draft.
The authors declare that they have no conflict of interest.
The authors are grateful to the Meteorological Observation Department of the Estonian Environment Agency for providing rain gauge datasets. The authors would also like to thank editor Laurent Pfister, reviewer Hidde Leijnse, and two anonymous reviewers for their constructive comments that helped to significantly improve the manuscript.
This research has been supported by the Estonian Ministry of Education and Research (grant no. IUT20-11), the Estonian Research Council (grant no. PSG202), and the European Regional Development Fund within the National Programme for Addressing Socio-Economic Challenges through R&D (grant no. RITA1/02-52-07).
This paper was edited by Laurent Pfister and reviewed by Hidde Leijnse and two anonymous referees.