Creation of a high resolution precipitation data set by merging gridded gauge data and radar observations for Sweden

Journal of Hydrology xxx (2015) xxx–xxx

Contents lists available at ScienceDirect

Journal of Hydrology journal homepage: www.elsevier.com/locate/jhydrol

Creation of a high resolution precipitation data set by merging gridded gauge data and radar observations for Sweden Peter Berg ⇑, Lars Norin, Jonas Olsson SMHI, Folkborgsvägen 17, 601 76 Norrköping, Sweden

a r t i c l e

i n f o

Article history: Available online xxxx Keywords: Radar Climatology Bias correction

s u m m a r y Hydrological forecasting systems require accurate initial conditions, particularly for real time precipitation data, which are problematic to retrieve. This is especially difficult for high temporal and spatial resolutions, e.g. sub-daily and less than 10–20 km. Forecasting fast processes such as flash flood are, however, dependent on such high resolution data. Gridded gauge data produces too smooth fields and underestimates small scale phenomena, such as convection, whereas radar composites contain the small scale information, but suffer from inconsistencies between individual radars and have poor long term statistics. Here, we present a method to merge a radar composite with daily resolution gridded gauge data for Sweden for the time period 2009–2014 to produce a one hourly 4
4 km2 data set. The method consists of a main step where monthly accumulations of the radar data are scaled by those retrieved from the gridded data for each month. An optional quantile mapping based bias correction step makes sure that the daily intensity distribution agrees with the gridded observations. Finally, the data are disaggregated to an hourly time resolution. This results in a data set which has the same long-term spatial properties as the gridded observations, but with the spatial and temporal details of the radar data. Validation of the method is performed with high resolution gauge data, and shows a high quality of the derived product. Ó 2015 Elsevier B.V. All rights reserved.

1. Introduction Weather radars have been supporting meteorologists in their short term forecasts for many years now. However, their use as forcing data sets for impact models have been limited by the often variable and unreliable information that the radar provides. Radar composites mostly produce reasonable spatial information about precipitating systems at short accumulations, but when aggregating the data for longer time periods non-physical patterns often appears as a result of small systematic errors (Olsson et al., 2013). Blocking or deflection of the radar beam, nonprecipitation echoes, and changes in the radar hardware or software are just a few examples of problems that present a challenge for constructing longer time series of forcing data from radar records (Michelson et al., 2005). Recently, hydrological forecasting in Sweden has started to extend its focus from large scale precipitation to also include smaller temporal and spatial events. Extreme small scale events are generally of convective type, e.g. thunderstorms, and can produce large amounts of precipitation over a small area in a short ⇑ Corresponding author. Tel.: +46 (0)114958436. E-mail address: [email protected] (P. Berg).

time. Such events can have severe impacts on smaller catchments and lead to devastating flash floods. Forecasting of convective events is progressing rapidly through higher resolution meteorological models in combination with ensemble methodologies. To take full advantage of these high-resolution meteorological forecasts in flood forecasting, the hydrological model needs to be of similar high resolutions. In Sweden, the HYPE model (Lindström et al., 2010) is used for flood forecasting, with the country divided into almost 40,000 sub-catchments with a mean size of about 7 km2. This is essentially sufficient for resolving also the small, steep sub-catchments in hilly terrain and the urban and peri-urban subcatchments that are most prone to fast flooding. The time step used is, however, one day which prevents an accurate representation of rapidly increasing flows. Currently, the system is being developed for hourly simulation and a main challenge in this work is to achieve a continuous hourly model initialization, i.e. a proper description of the initial state of the soil and water storages leading up to the forecast. This initialization requires nation-wide hourly updated high-resolution precipitation fields and to construct this forcing data set is one key objective of the present study. Precipitation gauges are generally considered the most reliable measure of precipitation, but the network density puts constraints on the possibilities to produce gridded data sets. A typical

http://dx.doi.org/10.1016/j.jhydrol.2015.11.031 0022-1694/Ó 2015 Elsevier B.V. All rights reserved.

Please cite this article in press as: Berg, P., et al. Creation of a high resolution precipitation data set by merging gridded gauge data and radar observations for Sweden. J. Hydrol. (2015), http://dx.doi.org/10.1016/j.jhydrol.2015.11.031

2

P. Berg et al. / Journal of Hydrology xxx (2015) xxx–xxx

midlatitude convective event has a spatial extent of about 10 km and a lifespan of about 30 min (Berg et al., 2013), and the spatial scales of precipitating systems decrease with shorter temporal scales (Eggert et al., 2015). Therefore, the density of the gauge network puts constraints on the attainable spatial as well as temporal resolution when gridding the station data. Hourly data is for that reason not possible, except for extremely dense networks, and daily or monthly time scales are selected for most products. Radar data provides good spatial coverage at a very high spatial (1–2 km) and temporal (5–15 min) resolution. Radars measure the backscattered power from its targets which are then converted to a radar reflectivity factor (e.g. Doviak and Zrnic´, 2006). The reflectivity (Z) is converted to precipitation rate (R) using a Z–R relationship (e.g. Battan, 1973). Converting reflectivity to precipitation rate is theoretically based on the drop size distribution, which is, however, unknown and must be assumed. Futhermore, the Z–R relationship depends on the type of precipitation and shows strong variability from event to event and even within a single precipitation event. Other sources of errors, such as blockage of the radar beam or attenuation, may lead to an underestimation of precipitation, whereas the detection of non-precipitating targets, such as ground echoes or clear air returns (e.g. echoes from insects or turbulence), leads to overestimation of precipitation. Near the melting layer, water coated snow particles give rise to strong radar echoes (a so-called ‘‘bright band”) which can lead to an overestimation of the precipitation rate. Measurements from a network of radars can be merged into composite images. However, non-homogenized calibration between neighboring radars may result in a patchy precipitation field across the composite image. Substantial efforts have been put into producing radar composites of high and homegenous quality, and several different methods have been proposed (e.g. Krajewski et al., 2010; Berne and Krajewski, 2013). The best applicable method is case-dependent, and for places with low density gauge networks, the so-called ‘‘mean field bias” method can be applied, where all gauges covered by the radar are averaged and used to correct the radar derived intensities (e.g. Wilson and Brandes, 1979). With denser gauge network, numerous methods have been proposed, e.g. ‘‘conditional merging” (Sinclair and Pegram, 2005; Goudenhoofdt and Delobbe, 2009; Yoon and Bae, 2013) or different methods of first deriving a gridded precipitation field from the gauges and then use this for the radar corrections (e.g. Krajewski, 1987; Haberlandt, 2007; Paulat et al., 2008; Zhang et al., 2014). Paulat et al. (2008) combined a gridded gauge network of daily data with an hourly resolution radar composite to produce a high resolution (hourly, 7 km) merged data set. Their method basically consists of scaling the hourly radar intensities by the ratio of the gridded gauge data and the daily sum of the radar data, thus disaggregating the gridded data to a higher temporal resolution. This method requires a very high resolution gauge network in order to capture all spatial details at the daily time scale. In Sweden, and many other regions, the gauge networks are much coarser and do not support such high detail at the daily timescale. Precipitation patterns become more spatially homogenous with temporal aggregation, and here we investigate a method that is principally similar to the Paulat et al. (2008) method, but base it on monthly gridded data together with a distribution based bias correction algorithm. Radar data from the operational system used in Sweden, called NORDRAD, are used to provide the spatial information and the high temporal resolution. The new data set is referred to as HIPRAD (HIgh-resolution Precipitation from gauge-adjusted weather RADar). The data are presented in Section 2, the method of combining the data sources in Section 3, and evaluation results in Section 4. We close the paper with discussion and conclusions in Section 5.

2. Data The NORDRAD radar composite is the product used operationally for Sweden. NORDRAD (Carlsson, 1995) is a close collaboration between the Swedish Meteorological and Hydrological Institute (SMHI), the Norwegian Meteorological Institute, the Finnish Meteorological Institute, the Estonian Meteorological and Hydrological Institute, and the Latvian Environment, Geology and Meteorology Agency and has an additional agreement with the Danish Meteorological Institute. Within the NORDRAD collaboration, horizontal cross sections of radar reflectivity (pseudoconstant altitude plan position indicator, PCAPPI) are exchanged in real time. There are currently 35 operational weather radars in Sweden, Norway, Finland, Estonia, Latvia, and Denmark. The Swedish weather radar network consists of 12 C-band Ericsson Doppler radars. These radars perform scans at ten different tilt angles (from 0.5° to 40°) every 15 min. Echoes with radial velocities less than ±1 m/s are suppressed by a built-in clutter filter. A more detailed description of the Swedish weather radars can be found in Norin (2015). At SMHI, composite radar images covering the NORDRAD countries are generated using radar data from the nearest radar from as many available weather radars as possible. The NORDRAD composite image has a spatial resolution of 2
2 km2 and is generated every 15 min. In addition to ground clutter filtering, performed by each individual radar, several corrections are made to the radar composite in the post-processing. A beam blockage correction, based on the method by Bech et al. (2003), is applied to correct for the reduction in reflectivity due to topography. After generating the radar composite image, satellite cloud observations are used for removing radar echoes in regions where no clouds are visible (Michelson, 2006), and the systematic range-dependent bias is corrected using measurements from rain gauges (Michelson and Koistinen, 2002). NORDRAD composite images have been archived at the SMHI since 2005. During 2007, a change in both hardware and software were introduced to the Swedish radars to enable Doppler processing for all scans, and since September 2014 Swedish radars are being upgraded with dual-polarization. Between 2009 and 2014 we have a relatively robust composite, and we restrict this study to that period. We further restrict the current study to southern Sweden in order to avoid influence of a region in northern Sweden with no radar coverage. The study area and the influencing radar locations are presented in Fig. 1. It also shows the borders between the influence regions of different radars used for the algorithm that computes the composite fields. The presented case is when all radars are active, and the borders will change if a radar becomes inactive. PTHBV is the name of a gridded precipitation gauge data product originally developed by Johansson and Chen (2003). It is in active use by the SMHI, and is continuously extended in time as new gauge data are collected. At the base of the data set are around 700 gauges, which are interpolated using optimal interpolation, and then corrected for orographic effects by applying a climatological wind direction. This leads to a 4
4 km2 spatial resolution daily data set for all of Sweden, including catchments extending beyond the political borders. The gauges that go into the PTHBV calculations have been investigated for under-catch problems, i.e. how much precipitation the gauge is not catching depending on e.g. turbulence around the gauge (Alexandersson, 2003). A catchcorrection, depending on the climatological wind conditions as well as precipitation type (rain or snow), is applied in seven different classes of wind exposure of the gauges. The corrections range between 0.15% and 12% for rainfall, and 4–36% for snow, and result in 10–18% increases on average, depending on the gauge type.

Please cite this article in press as: Berg, P., et al. Creation of a high resolution precipitation data set by merging gridded gauge data and radar observations for Sweden. J. Hydrol. (2015), http://dx.doi.org/10.1016/j.jhydrol.2015.11.031

P. Berg et al. / Journal of Hydrology xxx (2015) xxx–xxx

+

60

x

+

x

x

+ x

x

x

x

x

56

12

x

x

x x

x

x x

+

200

x x

+

x

+

14 16 Longitude

100

x

x

+

+

+

x

x

x

x

x

x

xx

x

x

x

+

x x

x

x

x

x x x

0

x

xx x

x

x

x

x

x

x Latitude 58

x

x

x

x

x

300

18

400

500

Topography [m] Fig. 1. Topography of southern Sweden, with gauge locations (red x’s), radar locations (orange pluses), and the borders for the radar composite (dashed orange lines). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Because of the catch correction applied in PTHBV, it is expected to have larger mean values compared to the non-catch corrected gauges used here. 3. Method The goal is to construct a high resolution precipitation product with high, i.e homogenous, quality both in time and space. This is carried out by merging information from a radar composite at high temporal resolution, with a comparably high spatial resolution gridded gauge data set at a coarser temporal resolution. Due to the varying quality of precipitation in the radar composite, it is necessary to first apply different filters and corrections to enforce a more homogenous radar product. Second, the radar composite is corrected at each grid point to the gridded gauge data set, to improve the spatial homogenity over long time periods. The method to do this follows five fundamental steps: (i) A first correction consists of identifying precipitation measurements that only occur in single 15 min periods in the radar composite, and setting these to zero as they are likely spurious signals. The radar echoes are then converted to intensities through the Z–R relationship:


1 Z b R¼ ¼ a

z

1010 a

3

(iii) In this step, the main spatial inhomogeneities are corrected. First, both radar and PTHBV data are accumulated over a 30-day period. Second, each time step of the radar is scaled by the ratio of monthly mean PTHBV and radar at each grid point. This adjusts the mean value of the radar data to be identical to that of PTHBV, and by applying this for each grid point also spatial inhomogeneities are removed from the radar data. The assumption is that PTHBV correctly represents the precipitation data at the monthly time scale. The 30-day period is used because it accumulates enough data to give robust results, while also being short enough to capture possible transient changes in the composite structure, e.g. when a single radar is not available for a shorter time period (for example during maintenance). (iv) The radar composite potentially has a bias in the precipitation intensity probability distribution function (PDF), due to errors in measurements and also due to the simplified conversion of echoes to intensities. Furthermore, if the scaling applied in the previous step is large, i.e. much different from 1.0, the intensity PDF will be affected more for higher than for lower intensities and potentially affect the PDF negatively. A bias correction method called quantile mapping (QM), that is commonly applied in climatological analyses (Berg et al., 2012; Piani et al., 2010) and has also been used to adjust radar measurements to gauge data (Calheiros and Zawadzki, 1987; Rosenfeld et al., 1993), can then be applied. It basically calculates a transfer function which transforms the source PDF (here the radar) to be equal to a reference PDF. In our case, we only have access to daily PTHBV data with sufficient spatial coverage, so a grid point specific transfer function is computed and applied at the daily time scale, and for each grid point individually. QM aims at correcting the distribution as a whole, and might therefore retain some bias in the mean moment. A separate scaling of the data is therefore performed after the QM step to assure a perfect correction of the mean. (v) Finally, the daily data are dis-aggregated back to hourly with the same method used to correct the monthly data in step (iii), i.e. by scaling each hour of the radar data from step (ii) by the ratio of the daily corrected data and the daily radar data. Step (iv) requires a calibration period for the correction factors, which must cover at least a few years. This makes the QMcorrection sensitive to changes in the radar configuration. Here, a calibration period from 2009 to 2011 was applied and the method was evaluated for 2012–2014. In Section 4, the method is presented both with and without step (iv), referred to as HIPRADQM and HIPRAD, respectively.

4. Results

!1 b

ð1Þ

where z is the measured reflectivity in dBZ, a = 200, and b = 1.6. At times, there are missing 15 min radar observations, and such cases are filled by interpolation of neighboring time intervals. The data are then aggregated to hourly resolution by averaging the 15 min intervals. (ii) The radar data are bi-linearly interpolated to the 4
4 km2 PTHBV-grid which is on a slightly different projection. This step acts also as a simple smoothing filter that reduces the impact of single radar pixels with very high intensities.

In the following analysis, we compare the newly derived products HIPRAD and HIPRAD-QM, to the original NORDRAD radar composite. The latter has, however, gone through the first step (i) of corrections presented in Section 3, where some spurios signals and missing data are treated. The time period 2009–2014, for which the radar composite is stable regarding both the radar network and the composite structure, is used for all analyses. Thus for HIPRAD-QM, both the calibration and validation periods are presented together. Analyzing the two periods separately in comparison to observations shows only small deviations, indicating robustness in the length of the calibration period and in the consistency of the radar composite over time (not shown).

Please cite this article in press as: Berg, P., et al. Creation of a high resolution precipitation data set by merging gridded gauge data and radar observations for Sweden. J. Hydrol. (2015), http://dx.doi.org/10.1016/j.jhydrol.2015.11.031

4

P. Berg et al. / Journal of Hydrology xxx (2015) xxx–xxx

NORDRAD − PTHBV HIPRAD−QM − PTHBV

HIPRAD − PTHBV

JJA

DJF

PTHBV

<100

180

240

300

360

420

480

<−25

−20

Mean seasonal sum [mm]

−15

−10

−5

0

5

10

15

20

>25

Relative difference [%]

Fig. 2. Seasonal averages for DJF (top) and JJA (bottom). Absolute values are shown for PTHBV (left), and relative differences to PTHBV for NORDRAD (2nd column), HIPRADQM (3rd column) and HIPRAD (right).

Fig. 2 shows the seasonal average deviations of NORDRAD, HIPRAD-QM and HIPRAD from PTHBV. The problems with the NORDRAD product for long term averages is striking, with clear blocking features and radial biases of well over ±25%, as well as large differences between the different radar influence regions (compare to Fig. 1). These biases are almost completely removed for HIPRAD, as almost per definition. For HIPRAD-QM, there are remaining biases of a few percent in some regions, with a stronger bias remaining for Hallandsåsen, which is a horst in southern Sweden (seen in Fig. 2 as a marked positive bias for JJA). This horst causes mainly clutter in the radar echoes. The bias over the horst is evident for both the calibration and validation periods, whereas other biases in southern Sweden are related to smaller differences between the two periods. Ideally, a longer time period should be used for the calibration in order to reduce the impact of natural oscillations in climate, but with the changes in radar set up and calibration it is not possible to attain this. Spring and autumn show similar biases to the summer and winter ones, respectively. PDFs of precipitation intensity, pooled for the complete domain, are presented in Fig. 3. Overall, the PDFs compare well, although some larger deviations occur for higher intensities. These deviations are large for NORDRAD, but significantly reduced for HIPRAD and especially HIPRAD-QM. In JJA, the deviations for the high extremes are less dramatic, and HIPRAD-QM is very close to PTHBV. Furthermore, NORDRAD and HIPRAD are underestimating moderate intensities, which are improved to some extent in HIPRAD-QM. A skill score for the PDFs is calculated by comparing the overlapping areas under the PDFs between each radar data

set and PTHBV (Perkins et al., 2007). A value between 0 and 1 is obtained, with 1 being a perfect score. The skill scores are generally high, with values for DJF of 0.97, 0.97, and 0.99 for NORDRAD, HIPRAD-QM and HIPRAD, respectively. In JJA, the performance of HIPRAD-QM stands out due to the large weight of the performance for high probability intensities, as noted above. The scores for JJA are 0.94, 0.98, and 0.95. In the following analysis, we focus on hourly statistics. The PDFs for hourly intensities are shown in Fig. 4. There are clear biases for both low intensities and for extremes. Wet hours, i.e. hours with at least 0.1 mm of precipitation (see Table 1), as well as low intensites are overestimated by NORDRAD and the derived products. A fundamental issue with radar derived measurements is that the radar measures precipitation up to a few kilometers height in the atmosphere and the corrections towards surface elevation are not always sufficient. Another issue that plays some role is the spatial inconsistency between the radar (here at 4
4 km2) and the point source gauge data, which causes a shift of the PDF towards lower intensities. HIPRAD, and especially HIPRAD-QM reduce the low intensity bias somewhat. Moderate intensites are simulated well by all three data sets. NORDRAD is overestimating extreme intensities in DJF, but is slightly underestimating the intensities in JJA. The HIPRAD algorithm reduces the extreme intensities to some extent for both HIPRAD and HIPRAD-QM, but they remain fairly close to NORDRAD. Table 1 presents additional statistics for the grid point comparison to the gauge data. As noted above, the wet hours are

Please cite this article in press as: Berg, P., et al. Creation of a high resolution precipitation data set by merging gridded gauge data and radar observations for Sweden. J. Hydrol. (2015), http://dx.doi.org/10.1016/j.jhydrol.2015.11.031

5

P. Berg et al. / Journal of Hydrology xxx (2015) xxx–xxx

Prob.density [day/mm]

10 10 10 10 10 10

DJF

−1

JJA

−2

−3

−4

−5

PTHBV NORDRAD HIPRAD−QM HIPRAD

−6

1

2

5

10

20

50

200

1

Intensity [mm/day]

2

5

10

20

50

200

Intensity [mm/day]

Fig. 3. PDFs of daily data pooled for the complete domain. Note the logarithmic axes.

Prob.density [h/mm]

10

10

10

10

10

10

DJF

−1

JJA

−2

−3

−4

Gauge NORDRAD HIPRAD−QM HIPRAD

−5

−6

0.1

0.5

2.0

5.0

20.0

Intensity [mm/h]

0.1

0.5

2.0

5.0

20.0

Intensity [mm/h]

Fig. 4. PDFs of hourly data at gauge locations, pooled for all stations in southern Sweden. Note the logarithmic axes.

Table 1 Hourly statistics averaged for all stations and nearest grid points for DJF/JJA.

Wet hour (%) Mean (mm/h) Q99.9 (mm/h) Max (mm/h) SD (mm/h) Cor (–)

Gauge

NORDRAD

HIPRAD-QM

HIPRAD

9.2/7.5 0.061/0.11 2.5/8.3 5.9/23.0 0.24/0.64 –

13.6/10.8 0.075/0.11 3.2/7.4 8.0/22.1 0.28/0.61 0.69/0.73

12.9/10.9 0.075/0.12 3.2/7.2 6.4/15.5 0.28/0.59 0.71/0.70

12.9/10.6 0.076/0.12 3.2/7.7 7.0/21.9 0.29/0.62 0.73/0.73

overestimated by NORDRAD by three to four percentage units, and this is more or less retained also in HIPRAD-QM and HIPRAD. The mean intensity is overestimated in DJF by all three data sets compared to the gauge data, although the JJA values are similar. This is explained by the undercatch correction in PTHBV, which is used as reference in the HIPRAD-corrections. The undercatch correction has a larger effect in winter due to problems with measuring snowfall with gauges. The 99.9th percentile pf precipitation intensity is presented in Table 1. It follows the same pattern as that presented in Fig. 4, with a slight overestimation in winter and underestimation in summer. The values are similar for NORDRAD, HIPRAD-QM and HIPRAD. Also the maximum recorded intensities follow this bias pattern, but are generally rather close to the gauge records.

Regarding the temporal characteristics, the standard deviation (SD) is similar between the three radar data sets, with a slight overestimation in DJF and underestimation in JJA. The temporal correlations are generally around 0.7, which is a high value considering that hourly data are used. The diurnal cycle of precipitation intensity is an interesting feature of hourly precipitation that is not yet possible to describe satisfactorily with hourly disaggregation methods based on climate model simulations, see Berg et al. (2015). Fig. 5 presents the diurnal cycles for DJF and JJA for the different data sets and the gauge data. In DJF, there is no pronounced diurnal cycle as the area is dominantly affected by stratiform cyclonic systems. NORDRAD, HIPRAD-QM and HIPRAD all show an offset due to the different mean values, due to the undercatch correction in PTHBV as noted above, but also strange lows in the morning and evening. The reason for the lows might be a processing error in the construction of the NORDRAD composite, and is being investigated. The JJA diurnal cycle is more pronounced with a primary afternoon peak, and a secondary morning peak. The former is due to increased convective activity in the afternoon, whereas the latter is the result of a peak in relative humidity in the cold morning hours. NORDRAD captures the diurnal cycle well, besides a too strong peak in the morning hours. This might be due to reflections from raindrops not reaching the ground, or bright band problems. A curious peak occurs slightly after the primary afternoon peak.

Please cite this article in press as: Berg, P., et al. Creation of a high resolution precipitation data set by merging gridded gauge data and radar observations for Sweden. J. Hydrol. (2015), http://dx.doi.org/10.1016/j.jhydrol.2015.11.031

6

P. Berg et al. / Journal of Hydrology xxx (2015) xxx–xxx

0.16

DJF

0.08

0.10

0.12

0.14

Gauge NORDRAD HIPRAD−QM HIPRAD

0.06

Intensity [mm/h]

JJA

00

03

06

09

12

15

18

21

00

03

06

Time [h]

09

12

15

18

21

Time [h]

Fig. 5. Diurnal cycle at gauge locations for DJF (left) and JJA (right) for non-catch corrected gauge data, NORDRAD, and the HIPRAD products that were corrected to the catch corrected PTHBV data.

We speculate that this might be due to remaining elevated droplets that are interpreted as precipitation by the radar. HIPRADQM and HIPRAD has an offset compared to the gauge data similar as in DJF. Defining an event as a temporal sequence of contiguous precipitation above a set threshold, here 0.1 mm/h, we can compare the radar products to the gauge data at the locations of the gauges, see Table 2. On average, the duration of the events is around 3 h for all three data sets. The maximum duration is clearly overestimated for NORDRAD, whereas HIPRAD-QM and HIPRAD both underestimate it. Both the mean event intensity, and the mean peak intensity, i.e. the highest recorded intensity during an event, are similar for all three data sets, although the gauge values are consistently slightly lower. Fig. 6 shows mean and peak intensities as a function of duration. Clearly, the mean event intensity is lower for the radar products compared to the gauge data. This is an expected effect from the spatial coarsening. There are no significant differences for HIPRAD-QM or HIPRAD compared to NORDRAD. Similar characteristics are also seen for the peak event intensities, where also an additional reduction of intensities for the shortest durations can be seen for HIPRAD-QM and HIPRAD. Two illustrative hourly snapshots from NORDRAD, HIPRAD-QM and HIPRAD are shown in Figs. 7 and 8. The spatial details are clearly retained in HIPRAD-QM and HIPRAD, and the main changes are in the magnitudes. Fig. 7 shows a large scale event where there are clear deviations in magnitudes over various regions. On snapshots like this, it is difficult to see the change in magnitude that occurs for different parts of the radar composite, which are so clearly illustrated in the long term average (Fig. 2). HIPRAD-QM and HIPRAD have corrections for this, which are manifested as certain problematic radar pixels as well as larger regions with changed magnitude. Table 2 Event statistics averaged for all stations and nearest grid points.

Mean duration (h) Max. duration (h) Mean intensity (mm/h) Mean peak int. (mm/h)

Gauge

NORDRAD

HIPRAD-QM

HIPRAD

2.9 82.0 0.53 0.90

3.1 123.0 0.60 1.02

3.1 58.0 0.64 1.07

3.0 58.0 0.64 1.07

Fig. 8 shows a snapshot of a convective storm that caused severe local flooding in the city of Malmö in the south-west corner of Sweden. HIPRAD-QM and HIPRAD have some slight changes to this extreme event, but are largely retained closely to the original radar image. 5. Discussion Radars are accurate in detecting precipitation, and advanced methods are available for correction of atmospheric effects on the beam path, measurement errors, etc. However, the radar networks are still largely under development regarding both hardware and software, and it is difficult to find consistent radar data for long time periods, especially for climatological research. The main purpose of the current study was to construct a data set that can be used to calibrate a hydrological model for running with hourly time steps. Clearly, the NORDRAD composite is not of sufficient quality for such applications, as can be seen in Fig. 2. The proposed method takes advantage of the high spatial and temporal details of the radar data, and also the long term stability and ground truth precipitation measurements of gauge networks, and was found to work satisfactorily for several investigated characteristics of precipitation. Most importantly it corrects the spatial patterns in the long term averages, while at the same time retaining a good agreement with station data at shorter sub-daily temporal accumulations. However, some deviations occur for the most intense events. Such events are problematic with the presented method when they dominate the monthly total precipitation and were not captured at all, or only partly, by the gauge network. A more flexible selection of the period for calculating the scaling factor could help resolve this issue, but the quality and underlying station density of the gridded gauge data are the main limiting factors for such events. Two versions of the corrections were presented; HIPRAD and HIPRAD-QM, with the additional step (iv) of corrections of the intensity PDF at the daily time scale. We find almost no significant differences between the two versions in our study, which is likely because of mostly small ratios between PTHBV and the radar data at the monthly time scale, as well as an originally good performance of the radar for the daily intensity PDF. The simpler method (HIPRAD) is then the obvious choice, especially considering the

Please cite this article in press as: Berg, P., et al. Creation of a high resolution precipitation data set by merging gridded gauge data and radar observations for Sweden. J. Hydrol. (2015), http://dx.doi.org/10.1016/j.jhydrol.2015.11.031

7

P. Berg et al. / Journal of Hydrology xxx (2015) xxx–xxx

HIPRAD−QM

HIPRAD

5

10

15

NORDRAD

30 20 10 0

Peak Intensity [mm/h]

40

0

Mean Intensity [mm/h]

20

Gauge

0

10

20

30

40

50 0

10

Duration [h]

0 7e−05 8e−04

20

30

40

50 0

10

Duration [h]

0.009

0.09 0.7

<−500

20

30

40

50 0

10

Duration [h]

−100

−50

Norm. event count [−]

−10

0

20

30

40

50

Duration [h]

10

50

100

Relative difference [%]

Fig. 6. Event based statistics for the average event intensiy (top) and the peak intensity (bottom) for gauges (1st column), and differences to the gauges for NORDRAD (2nd column), HIPRAD-QM (3rd column), and HIPRAD (4th column). Note the non-linear color scales.

20140812 − 09:00 NORDRAD

0

HIPRAD−QM

0.1

0.2

0.5

1

2

HIPRAD

5

10

20

50

100

Intensity [mm/h] Fig. 7. Snapshot from the 12th of August 2014 with a migrating large scale system over southern Sweden.

need for longer time series and stationarity of the bias for the QM method. Although the proposed method works well, it is a blunt tool for correcting the radar composite, and time would be well spent working on corrections directly on the composite, and calibrations between the individual radars. For climatological applications, it is

crucial to apply also new methods to past data and thereby create the basis for long consistent records of high quality. Also the Z–R relationship used to derive intensities from radar echoes was here used with the same parameters, independent of the precipitation phase (rain or snow) as well as of the drop size distribution in different types of precipitation processes. This has been shown to

Please cite this article in press as: Berg, P., et al. Creation of a high resolution precipitation data set by merging gridded gauge data and radar observations for Sweden. J. Hydrol. (2015), http://dx.doi.org/10.1016/j.jhydrol.2015.11.031

8

P. Berg et al. / Journal of Hydrology xxx (2015) xxx–xxx

20140831 − 05:00 NORDRAD

0

HIPRAD−QM

0.1

0.2

0.5

1

2

HIPRAD

5

10

20

50

100

Intensity [mm/h] Fig. 8. Snapshot from the 31st of August 2014 with a severe convective storm that caused serious flooding in the city of Malmö in the south-west corner of Sweden.

produce errors to a magnitude of around 20% for some locations (Gerstner and Heinemann, 2008), and is a topic for future development. The HIPRAD data set is currently applied to 1 h simulations with the hydrological model HYPE (Lindström et al., 2010), with a focus on short-term extreme precipitation. An operational version of the method, which will provide close to realtime updates of the current precipitation state to the hydrological model, is under development. Further extension of the data set to the Swedish radar archives to the early 2000s are ongoing, together with updated and temporally consistent corrections of the radar data. Acknowledgements The study was funded by the Swedish Civil Contingencies Agency (MSB), through project Grant No. 2011-3774. We acknowledge the work by colleagues at SMHI for the development and maintenance of the NORDRAD precipitation composite product, as well as for the PTHBV gridded data set. References Alexandersson, H., 2003. Korrektion av nederbörd enligt enkel klimatologisk metodik. Meteorologi 111 (SMHI). Battan, L.J., 1973. Radar Observation of the Atmosphere. University of Chicago Press. Bech, J., Codina, B., Lorente, J., Bebbington, D., 2003. The sensitivity of single polarization weather radar beam blockage correction to variability in the vertical refractivity gradient. J. Atmos. Oceanic Technol. 20, 845–855. Berg, P., Feldmann, H., Panitz, H.-J., 2012. Bias correction of high resolution regional climate model data. J. Hydrol., 80–92 Berg, P., Moseley, C., Haerter, J., 2013. Strong increase in convective precipitation in response to higher temperatures. Nature Geosci. 6 (3), 181–185. Berg, P., Bosshard, T., Yang, W., 2015. Model consistent pseudo-observations of precipitation and their use for bias correcting regional climate models. Climate 3, 118–132. Berne, A., Krajewski, W.F., 2013. Radar for hydrology: unfulfilled promise or unrecognized potential? Adv. Water Resour. 51, 357–366. Calheiros, R.V., Zawadzki, I., 1987. Reflectivity-rain rate relationships for radar hydrology in Brazil. J. Climate Appl. Meteor. 26 (1), 118–132. Carlsson, I., 1995. NORDRAD – weather radar network. In: Collier, C.G. (Ed.), COST 75 Weather Radar Systems. European Commission, pp. 45–52. Doviak, R.J., Zrnic´, D.S., 2006. Doppler Radar and Weather Observations. Dover Publications, Mineola, New York, USA. Eggert, B., Berg, P., Haerter, J.O., Jacob, D., Moseley, C., 2015. Temporal and spatial scaling impacts on extreme precipitation. Atmos. Chem. Phys. 15 (10), 5957– 5971.

Gerstner, E.-M., Heinemann, G., 2008. Real-time areal precipitation determination from radar by means of statistical objective analysis. J. Hydrol. 352 (3), 296–308. Goudenhoofdt, E., Delobbe, L., 2009. Evaluation of radar-gauge merging methods for quantitative precipitation estimates. Hydrol. Earth Syst. Sci. 13 (2), 195–203. Haberlandt, U., 2007. Geostatistical interpolation of hourly precipitation from rain gauges and radar for a large-scale extreme rainfall event. J. Hydrol. 332 (1), 144–157. Johansson, B., Chen, D., 2003. The influence of wind and topography on precipitation distribution in Sweden: statistical analysis and modelling. Int. J. Climatol. 23 (12), 1523–1535. Krajewski, W.F., 1987. Cokriging radar-rainfall and rain gage data. J. Geophys. Res. 92 (D8), 9571–9580. Krajewski, W.F., Villarini, G., Smith, J., 2010. Radar-rainfall uncertainties: where are we after thirty years of effort? Bull. Am. Meteorol. Soc. 91 (1), 87–94. Lindström, G., Pers, C., Rosberg, R., Strömqvist, J., Arheimer, B., 2010. Development and test of the HYPE (Hydrological Predictions for the Environment) model – a water quality model for different spatial scales. Hydrol. Res. 41 (3–4), 295–319. Michelson, D., 2006. The Swedish Weather Radar Production Chain. Tech. rep., Proc. ERAD 2006. Michelson, D.B., Koistinen, J., 2002. Gauge-radar network adjustment for the Baltic Sea Experiment. Phys. Chem. Earth (B) 25, 10–12. Michelson, D., Einfalt, T., Holleman, I., Gjertsen, U., Friedrich, K., Haase, G., Lindskog, M., Jurczyk, A., 2005. Weather Radar Data Quality in Europe – Quality Control and Characterization. Tech. rep., COST Action 717, Luxembourg. Norin, L., 2015. A quantitative analysis of the impact of wind turbines on operational Doppler weather radar data. Atmos. Meas. Tech. 8 (2), 593–609. Olsson, J., Södling, J., Wetterhall, F., 2013. High-resolution precipitation data for hydrological modelling: a preliminary study. SMHI Hydrology 116 (SMHI, 601 76 Norrköping, in Swedish). Paulat, M., Frei, C., Hagen, M., Wernli, H., 2008. A gridded dataset of hourly precipitation in Germany: its construction, climatology and application. Meteor. Z. 17 (6), 719–732. Perkins, S.E., Pitman, A.J., Holbrook, N.J., McAneney, J., 2007. Evaluation of the AR4 climate models’ simulated daily maximum temperature, minimum temperature, and precipitation over Australia using probability density functions. J. Climate 20, 4356–4376. Piani, C., Haerter, J.O., Coppola, E., 2010. Statistical bias correction for daily precipitation in regional climate models over Europe. Theor. Appl. Climatol. 99, 187–192. Rosenfeld, D., Wolff, D.B., Atlas, D., 1993. General probability-matched relations between radar reflectivity and rain rate. J. Appl. Meteor. 32 (1), 50–72. Sinclair, S., Pegram, G., 2005. Combining radar and rain gauge rainfall estimates using conditional merging. Atmos. Sci. Lett. 6 (1), 19–22. Wilson, J.W., Brandes, E.A., 1979. Radar measurement of rainfall-a summary. Bull. Am. Meteorol. Soc. 60 (9), 1048–1058. Yoon, S.-S., Bae, D.-H., 2013. Optimal rainfall estimation by considering elevation in the Han River Basin, South Korea. J. Appl. Meteorol. 52 (4), 802–818. Zhang, J., Qi, Y., Langston, C., Kaney, B., Howard, K., 2014. A real-time algorithm for merging radar QPEs with rain gauge observations and orographic precipitation climatology. J. Hydrometeor. 15 (5), 1794–1809.

Please cite this article in press as: Berg, P., et al. Creation of a high resolution precipitation data set by merging gridded gauge data and radar observations for Sweden. J. Hydrol. (2015), http://dx.doi.org/10.1016/j.jhydrol.2015.11.031