Updated precipitation series for the UK derived from Met Office gridded data

Authors

  • I. R. Simpson,

    1. Climatic Research Unit, University of East Anglia, Norwich, NR4 7TJ, UK
    Search for more papers by this author
  • P. D. Jones

    Corresponding author
    1. Climatic Research Unit, University of East Anglia, Norwich, NR4 7TJ, UK
    2. Center of Excellence for Climate Change Research/Dept of Meteorology, King Abdulaziz University, Jeddah, Saudi Arabia
    • P. D.Jones, Climatic Research Unit, University of East Anglia, Norwich, NR4 7TJ, UK.
    Search for more papers by this author

Abstract

This paper introduces new, homogenized UK national and regional rainfall series derived from 5-km gridded daily and monthly precipitation data. These new series have the advantage that they are based on readings from a dense network of rain gauges rather than a maximum of seven per region. Series developed by alternative statistical methods are analysed and compared. Error estimates are derived for the series, based on linear regressive residuals. Using individual regressions for each calendar month, the monthly series are extended back to 1766 for monthly precipitation in England and Wales (EW), to 1873 for monthly values for each of the five EW sub-regions, and to 1931 for daily values in all regions. A follow-up paper will provide analysis of how mean precipitation totals and extremes have changed over the respective periods. Copyright © 2011 Royal Meteorological Society

1. Introduction

Since the early 20th century, work has been underway to produce a homogeneous series of rainfall records across the UK, started by G.J.Symons (Wigley et al., 1984). Following the development of a composite EW rainfall series by Nicholas and Glasspoole (1931), continuing Symons's work, Wigley et al. (1984) developed a homogenized monthly rainfall series for EW, based on series for five separate spatially coherent regions (hereafter referred to as HadUKP). Seven well-spaced stations with long homogeneous records were used to obtain rainfall averages from each individual region, though fewer gauges were used until the mid 19th century, but maintaining similar reduced numbers in each region. The regional mean precipitation totals for each month were calculated using the unweighted average of the seven stations within a given region. Gregory et al. (1991) and Jones and Conway (1997) extended this analysis to include three Scottish regions and one region covering Northern Ireland (NI), and also developed daily series extending back to 1931. The national series for EW and Scotland are derived using weighted averages of the individual sub-regions. The weights are given in Table I. Further discussion of the series is given in these papers.

Table I. The weights used to derive the national series for England and Wales (EW), Scotland (S) and Northern Ireland (NI). The Northern Ireland weight is the multiplier to maintain homogeneity with the UK Met Office for the province (Jones & Conway, 1997)
RegionWeight
  • *

    Includes a small part of south-eastern Scotland.

South East England (SEE)0.275
Central and Eastern England (CEE)0.265
South-West England and South Wales (SWE)0.288
North-West England and North Wales (NWE)0.128
North-East England (NEE)*0.158
South Scotland (SS)0.379
North Scotland (NS)0.270
East Scotland (ES)0.529
Northern Ireland (NI)0.955

Updating the regional and national series was time-consuming and complex due to observing stations moving or closing, meaning that substitutes had to be found to maintain the same number of stations to ensure the homogeneity of the series (Alexander and Jones, 2001), since if the number of stations decreases then the variance increases and the mean may change depending on the locations of the remaining sites. Alexander and Jones (2001) subsequently developed a means of updating the regional and national series in near-real time, using data from MIDAS (the Met Office's database). This scaled precipitation totals according to the ratio of the regional normals obtained from the earlier method of producing the series to the stations' monthly normals. Comparisons with the existing Jones and Conway (1997) monthly series for 1961–1995 yielded excellent agreement but there were significant differences relative to the daily series of Jones and Conway (1997) for 1997, especially in regions with variable topography and limited station coverage, most notably northern Scotland (NS) (Table I).

Recently, the Met Office Hadley Centre (MOHC) developed a 5-km gridded daily precipitation dataset covering the UK and spanning 1958–2007 (MOHC07d). The daily dataset used a slightly different interpolation method than the monthly gridded precipitation data spanning 1914–2006 (MOHC06m), which was produced as part of the development of monthly and annual gridded data for 36 climate variables (Perry and Hollis, 2005a, 2005b).

The aim of this paper has been to generate new rainfall series based on MOHC07d and MoHC06m, extended prior to 1914 whenever HadUKP data are available, based on gridded data interpolated from a large number of observing sites rather than just seven within each region as per the HadUKP series, and thus provide a ‘truer’ representation of the areal mean. Trends in absolute and extreme precipitation across the UK, at regional and national level, will be analysed in a follow-up paper. Regressions have been developed for each month and for each UK region, with error estimates for the regressions, assuming that the new MOHC grids approximate the ‘true’ UK rainfall values and have no error. Each calendar month has been considered separately because of the seasonal variability in the mean UK precipitation distribution.

2. Methods of calculating the regional series

Regional daily and monthly precipitation data were generated in the same format as the HadUKP series (Jones and Conway, 1997) through extraction of totals from the MOHC07d and MOHC06m grids. The most accurate representation of the ‘true’ areal value was obtained by the unweighted averaging of all the grid boxes (AVGR) within each region of the UK, and deriving the composite national series (EW, S and NI) using the weighted mean of the individual sub-regions, using the same weights as were used to generate the HadUKP national series (Table I). An alternative method was to calculate areal values by the unweighted averaging of the seven grid boxes (AVG7) within a given region that correspond most closely to the locations of the seven stations used per region in the HadUKP data. This method enabled a test of whether the MOHC grid boxes corresponding to the station locations produced similar values to the corresponding values in the HadUKP data. Coordinates were given for the HadUKP stations in Jones and Conway (1997) and these were used to select the relevant 5-km grid boxes available in MOHC07d and MOHC06m.

Least squares regressions were developed between the values extracted from MOHC07d and MOHC06m and those of HadUKP. The intercepts were fixed at zero, because it would be unrealistic to suggest that the MOHC data gave precipitation on days when there was none at all according to HadUKP. Days for which the MOHC and HadUKP data both produced zero were excluded from the regression analysis as they may bias the results. All regression analysis ended with 1997 due to likely inhomogeneities in the methods used to generate the HadUKP series from 1998 onwards. Correlations with the original HadUKP series were highest when using AVG7 (Table II), as expected, owing to the collocation of the stations and grid boxes. For EW, correlation values exceeded 0.99 taking all ‘wet’ days of all months as a whole. Using AVGR (Table III), correlation coefficients were slightly lower but, significantly, were similar to those between data using the original HadUKP methodology and the revised methodology given by Alexander and Jones (2001). The mean correlation was 0.99 for EW and 0.98 for Scotland.

Table II. Correlation coefficients between HadUKP and MOHC07d, using AVG7 for daily data for the period 1958–1997
RegionJanFebMarAprMayJunJulAugSepOctNovDecAll
NS0.960.950.940.950.920.930.930.960.950.950.940.960.95
ES0.990.980.980.990.990.990.980.990.990.990.980.980.99
SS0.990.990.990.990.990.990.990.990.990.990.990.990.99
NEW0.990.980.980.980.970.980.980.980.980.990.980.990.98
NEE0.990.990.990.990.990.980.980.990.990.990.990.990.99
SWE0.990.990.990.990.990.980.990.980.990.990.990.990.99
CEE1.000.990.990.990.990.990.990.990.990.990.991.000.99
SEE0.990.990.990.990.980.980.980.990.990.990.990.990.99
NI0.980.980.980.980.980.980.970.980.990.990.980.980.98
S0.990.990.990.990.990.990.990.990.990.990.990.990.99
EW1.001.001.001.000.990.990.990.991.001.001.001.001.00
Table III. Correlation coefficients between HadUKP and MOHC07d, using AVGR for daily data for the period 1958–1997
RegionJanFebMarAprMayJunJulAugSepOctNovDecAll
NS0.910.910.900.900.870.880.880.900.900.910.910.900.90
ES0.960.950.940.960.960.960.960.960.970.970.950.950.95
SS0.980.970.980.980.970.970.970.970.970.980.970.980.97
NWE0.980.970.970.970.960.950.960.960.970.970.970.980.96
NEE0.970.970.970.970.970.950.960.960.970.970.980.970.96
SWE0.980.980.980.980.960.970.970.970.970.970.980.980.97
CEE0.990.990.980.980.970.970.960.970.980.990.990.990.98
SEE0.980.980.980.970.970.970.960.970.960.980.980.990.97
NI0.970.980.980.970.970.980.960.980.980.980.980.980.97
S0.980.980.970.980.980.980.980.980.980.980.980.980.98
EW0.990.990.990.990.990.990.990.990.990.990.990.990.99

For EW, using AVG7 (not shown), 90% of daily values predicted by the regressions were within 0.5 mm of the corresponding HadUKP values and 98% were within 1 mm, while for NS, 55% were within 0.5 mm and 72% within 1 mm. Using AVGR (Figure 1), the percentage of values within 0.5 and 1 mm was smaller. EW had 79% of all values within 0.5 mm, and 93% within 1 mm, while the NS region had the lowest accordance with 40% of values within 0.5 mm and 56% within 1 mm. NS had considerably greater differences than any other region (as was also true of the revised methodology detailed by Alexander and Jones, 2001), due to the more variable topography, the more limited network of sites, and greater absolute amounts of precipitation than in the other regions.

Figure 1.

Percentage of daily values within 0.5 mm (top) and 1 mm (bottom) using AVGR over the period 1958–1997, for each region, for each individual month of the year plus all days of all months, excluding dry days. NS (pink) shows less agreement between MOHC07d and HadUKP than the other regions. This figure is available in colour online at wileyonlinelibrary.com/journal/joc

The regression gradients, of the form (MOHC07d = a(HadUKP)), were consistently close to 1 using AVG7 (Table IV) but were substantially in excess of 1 for NWE and east Scotland (ES) when the AVGR method was used (Table V). This strongly suggests that the seven stations used for HadUKP were biased towards the drier parts of those regions. Southeast England (SEE) showed gradients close to 0.9, suggesting that the Met Office gridded data produces less precipitation for SEE than HadUKP, even when the analysis is restricted to the seven grid boxes that correspond to the HadUKP sites for SEE.

Table IV. Regression gradients of the form (MOHC07d = a[HadUKP]) for each region, using AVG7 for daily data for the period 1958–1997
RegionJanFebMarAprMayJunJulAugSepOctNovDecAll
NS1.011.030.991.050.971.041.051.010.991.011.021.001.01
ES1.051.031.031.061.041.031.061.061.031.051.031.021.04
SS1.001.011.001.011.010.980.991.010.990.991.011.011.00
NWE0.971.021.031.041.031.041.071.031.041.021.031.011.02
NEE0.970.950.981.000.980.991.000.980.980.990.970.950.98
SWE1.031.031.001.081.051.031.031.051.000.981.011.021.02
CEE0.971.000.980.950.970.980.940.930.970.960.970.990.97
SEE0.890.890.910.930.930.920.910.890.920.900.900.910.91
NI1.071.061.041.011.050.990.991.011.011.001.021.081.03
S1.031.041.031.051.031.041.051.041.021.031.031.031.03
EW0.981.000.991.011.001.001.000.990.990.980.990.990.99
Table V. Regression gradients of the form (MOHC07d = a[HadUKP]) for each region, using AVGR for daily data for the period 1958–1997
RegionJanFebMarAprMayJunJulAugSepOctNovDecAll
NS1.061.101.031.010.981.010.930.960.981.011.071.051.02
ES1.341.341.311.251.181.131.131.151.231.321.331.301.25
SS1.031.021.031.051.041.011.011.031.001.011.011.031.02
NWE1.201.261.241.211.141.081.101.131.161.171.201.221.18
NEE1.011.021.011.000.980.980.990.991.021.010.991.011.00
SWE1.061.061.041.061.071.061.031.081.091.061.061.071.06
CEE0.980.980.960.950.910.960.920.920.960.970.980.980.95
SEE0.950.950.950.930.890.920.900.890.950.930.940.950.93
NI1.041.031.001.001.010.980.971.010.980.990.981.031.00
S1.161.181.151.161.131.111.081.101.111.141.161.151.14
EW1.071.081.061.061.041.041.031.041.061.071.071.071.06

Figure 2 shows the results obtained when the regression predictions (regression gradient * HadUKP) are compared with the actual values obtained directly from MOHC07d using AVGR, for EW and for Scotland. There is a clear decline in accuracy in the years following 1997, illustrating the inhomogeneities in the real-time methodology described by Alexander and Jones (2001) relative to the original HadUKP methodology.

Figure 2.

Regression accuracy expressed as the difference between the regression gradient * HadUKP and MOHC07d, using AVGR, for England and Wales (top) and Scotland (bottom). There is a clear decline in accuracy after 1997. This figure is available in colour online at wileyonlinelibrary.com/journal/joc

The monthly precipitation values for each region could be generated via either averaging the MOHC06m grids in the same way as for MOHC07d, or by summing the daily values generated from the MOHC07d grids to give monthly totals. Regressions between the monthly precipitation values generated from MOHC06m directly, and from aggregation of MOHC07d daily values, using AVGR, suggested that there was very little difference in the rainfall totals produced across any of the regions (correlation coefficients, to two decimal places, were 1.00 for all months for all regions). Gradients from monthly regressions (Table VI) were mostly a little above 1.00, suggesting that the Met Office monthly grids produce slightly less rainfall than the daily grids, presumably due to differences in the interpolation methods that are used, especially in ES.

Table VI. Regression gradients of the form (sum of daily values = a* (raw monthly values)) applying AVGR over the period 1958–2006 (for which MOHC daily and monthly data are both available)
RegionJanFebMarAprMayJunJulAugSepOctNovDecAll
NS1.031.051.031.071.061.061.061.071.031.031.031.031.04
ES1.051.071.061.081.051.061.071.071.071.061.061.061.06
SS1.021.041.031.041.041.031.031.041.031.021.031.031.03
NWE1.021.021.031.031.021.021.031.031.031.021.021.021.02
NEE1.031.031.031.031.031.031.031.031.031.031.031.021.03
SWE1.021.021.021.031.021.021.021.031.021.021.021.021.02
CEE1.001.001.011.011.001.011.011.011.011.011.001.001.01
SEE1.001.001.001.011.001.001.011.001.001.001.001.001.00
NI0.991.000.991.000.990.991.001.000.990.991.001.000.99
S1.041.051.041.061.051.051.051.061.041.041.041.041.04
EW1.011.021.021.021.011.021.021.021.021.021.011.011.02

3. Error estimates from the regressions

Error estimates were made using the residuals from the regressions for daily, monthly and seasonal time scales using least squares regression. Error bars were also derived for the gradient terms as well as the final values, and scaled by multiplying by the mean precipitation total. These errors are only ascribed to the years for which data are available for HadUKP but not the relevant MOHC dataset (MOHC07d or MOHC06m) because the values taken directly from the MOHC grids were accepted as the ‘true’ values. The regressions enable the MOHC-based series to be extended back as far as the HadUKP series extend, for example, back to 1766 for monthly data for EW, with associated standard error estimates.

Using AVG7 (Figure 3), mean standard error estimates for the monthly precipitation totals were less than 10 mm for all regions with the exception of NS, which produced error estimates closer to 15 mm during the winter half-year. The estimated standard errors of the slopes in the regressions were less than 2 mm in the case of all regions except NS, and were less than 1 mm in the cases of CEE, SEE and EW.

Figure 3.

Error bar estimates (mm) of the regression values obtained by applying the regression gradients to HadUKP monthly data, using the AVG7 method. This figure is available in colour online at wileyonlinelibrary.com/journal/joc

Using AVGR (Figure 4), error estimates were slightly larger, owing to not maintaining collocation. Mean standard error estimates for the monthly precipitation values were again below 10 mm for all regions except NS, where they exceeded 20 mm in the winter half-year. The estimated standard errors of the slopes in the regressions were below 2 mm for CEE, SEE, NEE and EW, but exceeded 5 mm for NS in January, July and September–December. Seasonal and annual totals had larger error bars, with larger error bars for winter and autumn than for spring and summer, with NS again having a much larger error than the other regions. Daily error estimates for AVGR (Table VII) were below 2 mm for all regions with the exception of NS which exceeded 2.5 mm during the winter half-year. The Scottish regions consistently show greater seasonal variation in error magnitudes than the EW regions, with a maximum in winter and minimum in summer, perhaps reflecting the greater precipitation amounts and variability in the Scottish regions in winter.

Figure 4.

Error bar estimates of the regression values obtained by applying the regression gradients to HadUKP monthly data, using the AVGR method. This figure is available in colour online at wileyonlinelibrary.com/journal/joc

Table VII. Mean error estimates for daily precipitation values from regressions based on MOHC07d, using AVGR
RegionJanFebMarAprMayJunJulAugSepOctNovDecAll
NS2.792.622.711.922.111.992.112.282.852.862.842.862.86
ES1.341.301.351.021.071.151.191.261.271.311.401.451.45
SS1.341.231.191.000.971.011.261.351.521.481.481.311.31
NWE1.141.101.130.961.041.251.311.461.411.391.331.341.34
NEE0.840.790.830.730.811.071.041.160.970.850.800.890.89
SWE1.060.950.970.820.981.071.051.271.271.221.141.131.13
CEE0.470.410.450.520.670.800.980.860.780.560.550.480.48
SEE0.660.560.660.710.811.061.091.011.281.010.830.700.70
NI0.970.830.760.820.770.800.950.990.901.020.970.970.97
S1.141.081.140.870.870.870.981.041.171.221.211.161.16
EW0.460.460.470.370.490.550.590.610.610.610.540.550.55

4. Conclusions

On the basis of the data obtained from MOHC07d and MOHC06m, daily and monthly precipitation series have been derived for the UK regions. The MOHC-derived values were assumed to represent ‘truth’ for the period covered by MOHC (spanning 1958–2007 for daily data, and 1914–2006 for monthly data). These data were extended further back by applying the regressions to the HadUKP data prior to 1914 (monthly) and 1958 (daily) to derive estimated daily and monthly precipitation totals, as far back as 1766 for monthly data for EW, 1873 for the individual EW monthly regional totals and 1931 for NI, Scotland and the Scottish regions. Error bars for these estimated precipitation totals are the errors from regression, assuming that the data covered by MOHC has no errors associated with generating the data, with the only potential source of issues being the interpolation methods used to generate the MOHC gridded datasets. Seasonal precipitation totals for EW and Scotland are shown in Figures 5 and 6, respectively. Earlier versions of these series have been discussed extensively in Jones and Conway (1997). Recent seasons are not unusual except for the autumn of 2000, which was the wettest on record. The aim of this study has been to provide a homogeneous rainfall series for the UK which maintains the same data span as the HADUKP series, but is based on gridded data interpolated from a large number of observing sites rather than just seven within each region, allowing precipitation totals to give a closer representation of the ‘true’ areal values.

Figure 5.

Seasonal precipitation totals for England and Wales with associated error bars. The red line is the series smoothed using a 51-point moving average. This figure is available in colour online at wileyonlinelibrary.com/journal/joc

Figure 6.

Seasonal precipitation totals for Scotland. The red line is the series smoothed using a 51-point moving average. This figure is available in colour online at wileyonlinelibrary.com/journal/joc

The results strongly suggested that the revised methodology for updating the HadUKP series described by Alexander and Jones (2001) contained significant inhomogeneities relative to the original methodology. There is evidence that the original methodology (using seven well-spaced sites within each region) led to bias towards the drier parts of a region, particularly in the case of NWE and ES, and that the mean inter-site correlations were not high enough to ensure that it gave a close representation of the ‘true’ areal values.

A follow-up paper will analyse trends in absolute and extreme precipitation based on the revised dataset, updating earlier work, e.g. by Jones and Conway (1997) and Alexander and Jones (2001).

Acknowledgements

* This work was supported by funding from the Natural Environment Research Council grant NE/F006888/1 and a CASE award 23 from the UK Meteorological Office. Matthew Perry and Dan Hollis produced the gridded precipitation data. David Parker, Tim Osborn, Keith Briffa, John Caesar and Elizabeth Good provided useful assistance and comments.

Ancillary

Advertisement