A new publicly available daily gridded precipitation data set over mainland Portugal is presented. This data set is also combined with a recent Spanish data set to obtain a high resolution (0.2° × 0.2°) Iberian data set, labeled IB02. This data set covers the period from 1950 to 2003 and is based on a dense network, with more than 2000 and 400 quality-controlled stations over Spain and Portugal, respectively. The ordinary kriging method, applied over Portugal for consistency with the Spanish data set, performs slightly better than simpler interpolation techniques tested over Portugal. Additionally, this paper evaluates four global gridded data sets: two based on rain gauges (Climate Research Unit (CRU) and Global Precipitation Climate Center (GPCC)) and two European Centre for Medium-Range Weather Forecasts (ECMWF) reanalyses (ERA-40 and ERA-Interim), comparing them with the IB02 data set. The main features of the spatial distribution of IB02 mean annual precipitation are reasonably captured by the global data sets, despite their dry biases, mostly in mountainous regions. The four data sets perform better in western Iberia and are able to identify the major drought spells at the Iberian scale. Despite these similarities, GPCC outperforms CRU and ERA-Interim is superior to ERA-40 with respect to several aspects, such as annual cycle and drought detection. The performance of CRU is similar to that of ERA-Interim. The frequency of wet days is overestimated by reanalyses, mainly by ERA-Interim, while heavy precipitation events are underestimated, mostly by ERA-40. At 5 day scales, ECMWF reanalyses reveal difficulties in predicting the magnitude of precipitation, despite their greater ability to estimate the peak locations.
 In the last decades, several gridded precipitation data sets have been developed. These data sets are of major importance for studies of climate variability, for validation of climate and Numerical Weather Prediction (NWP) models and for agricultural and hydrological applications, including drought and floods monitoring. Although observational data sets remain the primary source of precipitation data for hydrologic prediction and climate variability studies, precipitation from global and regional NWP models have been recently used for such purposes [Rodríguez-Fonseca and Serrano, 2002; Dutra et al., 2008].
 Several studies have compared different observational precipitation data sets over Africa [Lamptey, 2008] and over the globe [Fekete et al., 2004]. Precipitation of European Centre for Medium-Range Weather Forecasts (ECMWF) reanalyses has been evaluated over several regions, such as North America [Balsamo et al., 2010], boreal regions [Serreze et al., 2005] and over large river basins [Betts et al., 2003a, 2003b, 2005, 2009]. Herrera et al.  have compared the E-OBS data set [Haylock et al., 2008] with a high-resolution daily gridded data set over Spain. However, a comparison between global data sets, ECMWF reanalyses and a high-resolution data set of precipitation over the Iberian Peninsula is not documented.
 Iberian precipitation is characterized by high spatial and temporal variability because of a complex orography and diverse atmospheric regimes [Serrano et al., 1999]. For instance, mean annual precipitation varies between more than 2000 mm in the northwest coast and less than 200 mm in the eastern coast [Herrera et al., 2010]. In certain regions, the annual precipitation is concentrated in less than 60 days and periods of more than 120 days without precipitation have been registered [Martín-Vide and Gómez, 1999]. Moreover, in eastern Spain precipitation often has a torrential character (with more than 200 mm registered in 24 h), mostly during autumn and late summer, when the high sea surface temperatures from the Mediterranean and the topographic configuration of this region favors the development of convective systems [Romero et al., 1998; Martín et al., 2007].
 Climate model simulations for the next 10 to 60 years indicate that droughts and floods will be more frequent in many regions, including over the Iberian Peninsula [Lehner et al., 2006], an area that is currently vulnerable to droughts. Therefore, over water-stressed regions, such as Iberia, water resource management, hydrologic prediction and consequently precipitation have become a topic increasingly important. On the other hand, there has been an increase in the horizontal resolution of regional climate models, but there is a lack of fine-scale observational data sets needed to evaluate such models [Im et al., 2010]. For all these reasons, the development and assessment of high-resolution precipitation data sets becomes crucial.
 In this paper we describe the development of a new daily precipitation gridded data set over mainland Portugal (denoted PT02) and its merging with a recent Spanish gridded data set [Herrera et al., 2010] to produce a high-resolution (0.2° × 0.2°) Iberian data set, denominated IB02. This new data set spans the period from 1950 to 2003 and is based on a dense network of rain gauges (with more than 2000 stations over Spain and more than 400 stations over Portugal). Station homogeneity was tested applying the standard normal homogeneity test (SNHT) [Alexandersson, 1986]. The ordinary kriging method, applied by Herrera et al.  to create the Spanish data set was also used in this study over mainland Portugal, and compared with other simpler interpolation techniques, namely, inverse-distance schemes.
 In addition, this paper presents a comparison of four global precipitation data sets with the new gridded daily data set IB02. The global data sets include two observation-based monthly data sets: the Climatic Research Unit (CRU) and the Global Precipitation Climate Center (GPCC), and two European Centre for Medium-Range Weather Forecasts (ECMWF) reanalyses, ERA-40 [Uppala et al., 2005] and ERA-Interim [Dee et al., 2011]. These data sets will be compared with regard to several aspects, such as spatial pattern of annual mean, annual cycle and drought identification. The drought assessment is performed using the standardized precipitation index (SPI) [McKee et al., 1993]. Monthly and 5 day precipitation are evaluated using simples scores, such as bias and squared correlation. Finally, the accuracy of spatial patterns of the 5 day precipitation of ECMWF reanalyses is also assessed using the correspondence ratio [Stensrud and Wandishin, 2000], defined as the ratio between the area where both data sets represent precipitation and the area where at least one of the data sets depicts precipitation.
 The present paper is structured as follows. In section 2, the precipitation data sets used in this study are described, including the data and methodology used to develop the new PT02 data set and the joined IB02. The results from the comparison between four coarse-grid precipitation data sets and IB02 are presented in section 3, while section 4 presents concluding remarks.
2. Data Sets and Methods
2.1. ECMWF Reanalyses
 ERA-40 and ERA-Interim are reanalyses of meteorological observations produced by ECMWF. ERA-Interim covers the period from 1989 to present, while ERA-40 spans the period from September 1957 to August 2002. Both use hybrid vertical coordinate and 60 levels from surface to 0.1 hPa, with vertical layer spacing ranging from 0.5 km in midtroposphere to circa 1 km near the tropopause; the height of the lowest level is 10 m. The ERA-40 model uses a spectral resolution of T159 and a reduced Gaussian grid, with a grid spacing of about 110 km.
 The global observing system and consequently the data assimilated in ERA-40 changed considerably between 1957 and 2002. In 1973, satellite radiances started to be assimilated. From 1979 onward, the number of radiosondes decreased, but the number of observations from satellite, aircraft, ocean buoys and others increased substantially, with an additional increase in the number and breadth of remote-sensing data in the 1990s. ERA-40 uses a 3-D variational assimilation system with a 6 h analysis cycle [Uppala et al., 2005].
 The main upgrades of the ERA-Interim (hereafter referred to as ERA-I) relative to the ERA-40 include the use of 4D-Var assimilation with a 12 h cycle, more extensive use of satellite radiances, improved humidity assimilation, improved radiative transfer model and a T255 horizontal resolution (equivalent to 80 km grid spacing).
 Both daily reanalysis precipitation data sets presented in this paper correspond to the 12–24 h forecast interval from initial conditions at 00:00 and 12:00 UTC. An evaluation of the 00–12 h forecast interval was also performed showing an underestimation of precipitation in both reanalyses, with the 12–24 h forecast interval precipitation closer to observations (not shown). These differences are due to initial spin-up in the atmospheric model [Betts et al., 2003b; Andersson et al., 2005]. In addition, it was found that the effect of spin-up is larger in ERA-40 than in ERA-I, in agreement with Betts et al. .
2.2. Global Monthly Data (GPCC and CRU)
 GPCC is the official precipitation Data Center of the World Meteorological Organization (WMO), which receives near real time 7000–8000 Synoptic Ocean Prediction (SYNOP) and ground surface targets (CLIMAT) reports via the WMO's Global Telecommunication System (GTS). The other main data sources of GPCC are the Global Historical Climatology Network (GHCN), Food and Agriculture Organization (FAO) of the United Nations and Climate Research Unit (CRU). The GPCC precipitation data set spans the period from 1901 to 2007 and it comprises 9343 stations with a minimum of 90% data availability during the period from 1951 to 2000 and more than 25,000 stations with at least 10 years of data during this 50 year period [Fuchs, 2009; Schneider et al., 2010].
 The CRU TS3 global data set of monthly precipitation (updated by Mitchell and Jones ) covers the period 1901–2006 and comprises over 10,000 stations worldwide. It includes 428 stations over Europe with a minimum of 70% availability for the period 1961–1990 [Hulme and New, 1997] and more than 19,000 stations worldwide during this period [New et al., 1999]. Both data sets are available at several resolutions and their 0.5 × 0.5 (latitude × longitude) resolution fields are used.
2.3. Spanish Daily Data
Herrera et al.  developed a daily (07:00–07:00 UTC) precipitation gridded data set with a regular 0.2° latitude-longitude resolution for peninsular Spain and Balearic islands, using ordinary kriging [Goovaerts, 1998]. This data set (Spanish daily data, labeled by the authors Spain02, hereafter referred to as SP02) spans the period from 1950 to 2003 and is based on 2756 stations with at least 20 years of data.
2.4. Portuguese Daily Data
2.4.1. The Network
 In this paper we present a daily (09:00–09:00 UTC) precipitation gridded data set with a regular 0.2° resolution for mainland Portugal, labeled IB02. In Portugal, daily precipitation from the current day refers to accumulated precipitation from 09:00 UTC of the previous day to 09:00 UTC of the next day. This data set spans the period from 1950 to 2003 and is based on 806 stations, 188 meteorological stations from the Portuguese Meteorological Service (IM) and 618 rain gauges from the National Water Institute (INAG). Most of these (726) stations have at least 10 years of data. The other stations were kept because they correspond to automatic stations which replaced the conventional ones. Figure 1 presents the spatial distribution of the stations selected and the orography over mainland Portugal, showing a very dense network over the whole territory with the exception of the southwest coast. The mean and maximum distance between nearby stations is 11.7 km and 41.6 km, respectively. However, it is also clear that the network density has suffered considerable changes over the period (see also Figure 2). For instance, about 40% of stations have more than 30 years of data, but around 28% of stations have less than 15 years of data.
 Out of the 561 stations with more than 20 years of data, 132 are located above 500 m of altitude, 10 above 1000 m and 2 stations above 1500 m. The percentage of mainland Portugal with altitudes above 500 m, 1000 m and 1500 m is nearly 23%, 2% and 0.1%, respectively. As a consequence, the altitudinal distribution of stations over mainland Portugal is adequate to represent the influence of orography on precipitation patterns.
Figure 2 shows that the number of stations increased until 1980, when the mean and maximum distance between neighboring stations was about 7.9 km and 24.9 km, respectively. From 1980 onward, the number of stations remained approximately constant until 1992 and decreased sharply after this period, with an apparent recovery from 2001 onward. Similar evolutions of data density can be seen for Europe [Haylock et al., 2008] and for Spain [Herrera et al., 2010].
 Since one goal of the present work is to compare precipitation data sets over the Iberian Peninsula, the Spanish and Portuguese data sets were merged to define IB02. Consequently, the period and spatial resolution of the Portuguese daily gridded data set was chosen to be the same as the SP02 data set. In addition, the grid points of the regular grid were selected to avoid the interpolation of the gridded data over Spain. In other words, the grid of IB02 matches SP02 over Spain.
 It is important to mention that there are differences in the accumulation period between Spain02 [Herrera et al., 2010] and IB02 data sets. For instance, in Spain02 the precipitation for 11 January 2000 is the accumulated precipitation between 07:00 UTC from 11 January 2000 and 07:00 UTC from 12 January 2000, while in the IB02 data set the precipitation for 11 January 2000 refers to the accumulated precipitation between X UTC from 10 January 2000 and X UTC from 11 January 2000, where X is 07:00 and 09:00 in Spain and in Portugal, respectively.
2.4.2. Quality Control Procedures
 The data used to create the PT02 data set were subject to a quality control, which include suspicious data identification and inhomogeneity detection. First, a plausible value check [World Meteorological Organization (WMO), 2008] is done, for instance, values of daily precipitation above 350 mm are rejected. In addition, maximum values above 100 mm are rejected if the following conditions are simultaneously fulfilled for all points within a 150 km radius (denoted in the following by “vicinity”): (1) No other station reports a value greater than 100 mm; (2) the ratio between this value and the largest value in the vicinity is greater than 5; and (3) no station reports rain snowfall, showers, thunderstorms or hail (hereafter referred to as significant weather).
 The data from IM stations were subject to additional consistency checks. The daily (09:00–09:00 UTC) accumulated precipitation was verified to exceed the accumulated precipitation in subintervals (09:00–21:00 UTC and 18:00–06:00 UTC). Internal consistency checks [WMO, 2008] were also performed, in which other meteorological observations at the same station were compared to verify its consistency. In the present work, the daily precipitation is compared with relative humidity, total cloud cover (TCC), past and present weather. For example, if the maximum value of TCC in the 09:00–09:00 UTC interval is zero and no significant weather reported, and the data set indicates occurrence of precipitation, the cloudiness and precipitation observations are flagged as suspect. In general, daily precipitation values inconsistent with two or more variables are rejected.
 Station homogeneity is tested applying the standard normal homogeneity test (SNHT) for a single break [Alexandersson, 1986] to the annual series with at least 30 years. When this method is applied to precipitation, it assumes that the ratio between the series of the station being tested (candidate station) and the reference series is nearly constant in time. A clear explanation of the SNHT is presented by Alexandersson  and González-Rouco et al. .
 The reference series were calculated as a weighted average from a minimum of four nearest stations, using the squared correlation coefficient as weighting factor, for an overlapping period of at least 30 years. The stations used to create the reference series should satisfy at least one of the following conditions: the correlation between the candidate and the selected neighboring series should be higher than 0.75 or higher than 0.5 when the distance between the stations is less than 50 km.
 The quality of the reference series is a key issue for the inhomogeneity detection. Therefore, when a candidate series was classified as inhomogeneous, the SNHT was repeated two more times, with a new set of reference series calculated using only the series identified as homogeneous in the preceding tests. Finally, only those stations passing the final test at a 95% confidence level were used to build the gridded data set. The application of this procedure leads to the rejection of 42 stations.
2.4.3. Interpolation Procedures
 The first method used to build the gridded data set was ordinary kriging [Deutsch and Journel, 1998], assuming a constant term in the regression, similarly to Herrera et al. . This method was chosen to guarantee the consistency between Spanish and Portuguese gridded data sets. We also considered an exponential variogram: γ(d) = ce−d/a, where d is the distance. The parameter c and the correlation length (a) were set to 10 and 40, respectively, after cross-validation.
 Additionally, we use an inverse distance weighting (IDW) method, which involves a distance weighting of stations within an influence radius (Rin), according to:
where Pj represents the precipitation at given grid point, pi represents the station precipitation and dij represents the distance between a given station and a grid point of the regular mesh.
 Greater values of k assign greater influence to values closest to the interpolated point. In the present work, Rin = 45 km and the average and maximum distance between each grid point and its nearby stations is respectively 13.7 km and 42.4 km. On average for each grid point, 14.4, 22.8 and 16.1 stations are used in the periods 1950–1979, 1980–1995 and after 1995, respectively.
2.4.4. Validation of Interpolation Methods
 In order to assess the performance of the interpolation methods, we compare the observed daily and monthly precipitation from 93 stations with the values of precipitation interpolated directly to the coordinates of these stations. For each station, the data of such station are removed before the interpolation is applied. These stations are geographically well distributed over Portugal and have long data series. The interpolation was applied using three methods: the ordinary kriging (labeled OK) and function 1 with values k = 1 and k = 3, labeled IDW1 and IDW3, respectively. As validation measures we compute the correlation coefficient, the root-mean-square error (RMSE) and the RMSE normalized by the standard deviation of observations (RMSEN) for the daily and monthly precipitation only for the subset in which the observed precipitation is at least 0.5 mm. Additionally, the Taylor diagram [Taylor, 2001] is used to compare the three interpolation methods. This diagram enables the comparison of these methods in terms of their correlation, their RMSE (or RMSEN) and their standard deviation.
 The spatial distribution of the RMSEN and correlation coefficient for the kriging method over Portugal for daily and monthly values is presented in Figure 3, for the period 1970–1980. The majority of the stations display a daily correlation larger than 0.7, RMSEN smaller than 0.7 (Figure 3) and RMSE values smaller than 10 mm (not shown). Despite the largest values of RMSE (above 15 mm) found in Serra de Gerês (not shown), in terms of correlation and RMSEN, the worst performance of the kriging technique occurs at one station in Serra de Montesinho and at another station in southern Portugal. In these two stations, r drops to values between 0.4 and 0.6 and RMSEN varies between 0.9 and 1.1.
 In general, there is good agreement between the interpolated and observed monthly series, with r higher than 0.9 (or nearly 0.9) and RMSEN values smaller than 0.5 in most of the stations. The worst behavior of the interpolation occurs in the region of Sintra and Montejunto hills, where r drops to values between 0.8 and 0.9 and RMSEN varies between 0.5 and 0.7. On the other hand, the largest values of RMSE (above 75 mm) of monthly precipitation are found in northern Portugal, in Serra de Gerês and Serra de Montemuro (not shown), regions of complex topography and high variability. In the other regions, RMSE has values smaller than 50 mm and even smaller than 25 mm, mostly in southern Portugal (not shown).
 The comparison between the three interpolation techniques for the 93 stations is summarized in Figure 4, using a Taylor diagram for daily precipitation, including only the days with precipitation with at least 0.5 mm, for the period from 1970 to 1980. The OK and IDW1 techniques behave in a similar way, with correlations around 0.8 and values of RMSEN slightly above 0.6, when averaged over Portugal. The application of IDW3 leads to a smaller underestimation of the observed variability. However, IDW3 performs worse than the other two algorithms, with respect to correlation and RMSEN.
 The performance of the three interpolation algorithms was also analyzed for the other decades, revealing that the OK interpolation is the method less sensitive to the density of observations network. In addition, it was found that kriging outperforms the IDW methods, mostly during the 1950s and 1960s. On the other hand, during the 1980s, when the network density is maxima, IDW1 performs slightly better than OK. These results are consistent with time evolution of the network density (see Figure 2) and with the study by Hofstra et al. , showing that interpolation accuracy depends more on the station network than on the interpolation methods. These results are also in agreement with Dirks et al. , who found that for high-resolution networks there is a small benefit of using kriging methods instead of simpler techniques, such as IDW1.
2.5. Comparison Methods and Metrics
 We compare precipitation data sets by inspection of the geographical distribution of annual mean, annual cycle for different regions and assessment of major drought periods. Several skill scores, including bias, squared correlation (r2) and root-mean-square error normalized by the standard deviation of observation (RMSEN), were computed to evaluate quantitatively 5 day and monthly precipitation. Moreover, the quality of spatial patterns of 5 days of precipitation from ERA-40 and ERA-I was assessed using the correspondence ratio (CR).
 The CR applied to two data sets is defined as the ratio of the area of the region where precipitation occurs in both data sets, AI, versus the area of the region where precipitation occurs in at least one of the data sets, AU:
 Areas of observed and forecasted precipitation are defined for several threshold values of precipitation (1, 5, 10, 20, 30, 50 and 60 mm). For instance, for a threshold of 1 mm, AU is the area expressed by all grid points where observed or reanalysis precipitation exceeds 1 mm, and AI is the area defined by all grid points where both observed and reanalysis precipitation exceeds 1 mm.
 The CR has been used to evaluate outputs from ensembles of NWP models [Stensrud and Wandishin, 2000], but is applied here to provide information about the spatial agreement between observed and reanalyses fields. In the extreme cases, CR = 0 means that there is no overlap between reanalyses and observations, whereas CR = 1 implies an absolute spatial agreement among these fields. Low values of CR can be due to errors in the intensity of precipitation or/and in the position and shape of precipitation systems.
 It is not rare that NWP models underestimate or overestimate precipitation amounts. For instance, in section 3.1.1 we will show that both ERA-40 and ERA-I underestimate the occurrence of days with precipitation above 10 mm. Therefore, it is possible that the spatial pattern of precipitation will be correctly represented, but the value of CR for a certain fixed threshold will be very low only because of the negative bias of the model. To overcome this limitation and evaluate the capability of reanalysis (ERA-40 or ERA-I) to depict the correct position of highest precipitation amounts independently of their biases, CR can be computed with a threshold variable in time (depending on each 5 day period). This threshold depends on the spatial distribution of 5 day precipitation of observations and reanalyses, defined here by the 75th percentile (P75) of the precipitation from all grid points over the Iberian Peninsula, for each data set, IB02 (labeled P75ob) and reanalysis (labeled P75re). Grid points where precipitation exceeds the P75 represent the area of maximum precipitation. The application of this methodology to ERA-I and IB02 data sets, implies that AI is the area expressed by all grid points where precipitation from IB02 and ERA-I exceeds, P75ob and P75re, respectively.
 The skill scores used to assess the performance of the global observational data sets were computed at two different grids: at the IB02 grid and at the native 0.5° latitude-longitude grid (labeled IB05) of GPCC and CRU data sets. Similarly, the ERA-I and ERA-40 precipitation are evaluated at the IB02 grid and at 1° latitude-longitude grid (named IB1), which is close to their native resolution. The precipitation from IB02 was aggregated into the IB05 and IB1 grids using a first-order conservative method [Jones, 1999]. Precipitation from ERA-40 and ERA-I in their Gaussian grids was transformed into IB1 and IB02 grids using a bilinear interpolation.
 Skill scores were computed for each grid point of the referred grids and then spatially averaged over the Iberian Peninsula (labeled IP), over Portugal (labeled PT) and over five regions. These five regions of Iberia were derived using principal component (PC) analysis of SPI-3 (standardized precipitation index, 3 months) [McKee et al., 1993; Vicente-Serrano, 2006a]. The empirical orthogonal functions (EOFs) were rotated using the Varimax method [Kaiser, 1958]. After rotation, the first six EOFs explained 24, 20, 11, 6, 6, and 4% of the total SPI-3. The region definitions were based in the general patterns returned by the EOFs, aggregating the grid points through a maximum loading criterion. Figure 5b shows the five zones over Iberia where the first six EOFs have maximum loadings, representing 37% (Southwest, SW), 30% (Northwest, NW), 17% (Northeast, NE), 9% (Southeast, SE), and 6% (Cantabria, CANT) of the total area. This method has been used in other studies to derive regions in the Iberian Peninsula from SPI [Vicente-Serrano, 2006a] or soil moisture anomalies [Dutra et al., 2008]. In those studies the derived regions are similar to the ones in Figure 5b, with some differences that can be attributed to different observations (modeling) and different periods of analysis. Regions defined in such a way are used to discriminate areas with different precipitation characteristics on the seasonal scale. For this purpose, the preferred time scale is SPI-3, accounting for accumulated precipitation in the previous 3 months. The evaluation of the precipitation data sets is performed separately for each area. This allows the identification of regional limitations in the reanalysis and global data sets that will be addressed in this paper.
3.1. Annual and Monthly Precipitation
3.1.1. Annual Mean and Annual Cycle
 The spatial distribution of mean annual precipitation over the Iberian Peninsula is shown in Figure 6, and the main spatial patterns are coherent with the five areas derived from SPI-3 and are clearly influenced by the Iberia orography (see Figure 5a). All global data sets (Figures 6b–6e) are able to capture the main features of mean annual precipitation. The largest values are found in northern Portugal, Galicia, Cantabrian coast and western Pyrenees and the lowest values are found in the southeastern coast. However, the enhancement of precipitation due to orographic features of the Iberian Peninsula (e.g., Pyrenees, Sistema Ibérico, Sistema Central, Serra da Estrela, Sierra de Ronda) is underestimated by ECMWF reanalyses, leading to substantial dry biases (greater than 650 mm) in the surroundings of these mountainous systems, especially for ERA-40. These biases are expectable, since the horizontal resolution of the models used in ERA-40 and ERA-I is insufficient to correctly represent the complex topography of the Iberian Peninsula. The better performance of ERA-I relative to ERA-40 reflects an improved representation of moist physical processes in ERA-I, broadly described by Dee et al. .
 GPCC performs better than CRU, in particular through a better representation of the local maxima in southwest of Spain and in region of Sistema Central. The higher density of the gauge network used by GPCC (when compared to CRU) over Spain, can possibly explain these differences. The CRU data set has a reduced added value relatively to ERA-I over Spain, except in Galicia. Finally, only the IB02 data set is able to reflect the precipitation enhancement caused by the Sistema Ibérico, by Serra da Estrela and hills in southern Portugal.
 The mean annual cycles of precipitation for the different data sets and the monthly biases for the six regions of Iberia (including PT) are presented in Figure 7. ERA-40 has a considerable dry bias, more marked from October to December and over NE and CANT regions and Portugal. ERA-I reproduces the annual cycle noticeably better than ERA-40. Nevertheless, ERA-I underestimates precipitation in the rainiest months, mainly in Western Iberia and in the CANT region, while it shows a tendency for overestimation in NE and SE regions and over CANT from March to September (not shown).
 The strong annual cycle of precipitation over Iberia is well captured by GPCC and CRU, even though both data sets have a wet bias in NE and SE regions (Figure 7). In the Cantabrian region, the annual cycle is better captured by GPCC than by CRU. In this region, GPCC has a small positive bias, while CRU strongly underestimates precipitation in the rainiest months and slightly overestimates precipitation in summer months (not shown).
 The mean annual cycle of the frequency of days with precipitation greater than or equal to 1 mm (defined as wet days) over the Iberian Peninsula, the CANT and NE regions is presented in Figures 8a, 8c, and 8e. The number of wet days over Iberia is overestimated, mainly by ERA-I, which shows this tendency in all regions. For ERA-I this bias is strongest over NE and CANT regions and smallest over SW region. For ERA-40, this bias is negligible during the summer months, as illustrated in Figures 8a, 8c, and 8e. Despite their biases, both reanalyses are able to capture the intraannual variability of the number of wet days.
 The mean annual cycle of the frequency of days with precipitation greater than or equal to 10 mm (referred to as R10) over the Iberian Peninsula, the CANT and NE regions is presented in Figures 8b, 8d, and 8f. It is visible that the frequency of R10 is underestimated, essentially by ERA-40. This underestimation is strongest during the rainiest months, predominantly in CANT region, where both reanalyses strongly underestimate the amplitude of the seasonal variations of the frequency of R10. Nevertheless, despite these biases, the annual cycle of the number of days with precipitation greater than or equal to 10 mm is reasonably well captured in other areas of Iberia, particularly by ERA-I, as illustrated for NE region in Figure 8f.
3.1.2. Monthly Scores
 The squared correlations between monthly precipitation from IB02 and from the other four data sets (at IB02 grid) were calculated for every grid box and averaged over the Iberian Peninsula and over six areas, for 1990–2001. The correlation from GPCC and CRU was also computed at its native grid, IB05, while ERA-40 and ERA-I were evaluated at IB1 grid, which is closer to their native grids. These evaluations are summarized in Figure 9.
 It is clear that there is a higher agreement between IB02 and the global data sets when the comparison is done at their native grid (or at resolution closer to its nominal resolution) than when it is done at IB02 grid. On average, over the Iberian Peninsula, r2 increases about 5% for observational data sets and more than 10% for ECMWF reanalyses. This is consistent with the fact that the resolution of CRU and GPCC is intermediate between IB02 data set and ECMWF reanalyses.
 For ERA-I and ERA-40, this difference is larger in NE and SE regions, where ERA-I has squared correlation below 0.55 at IB02 and above 0.68 at IB1 grid. In these regions, precipitation has regularly a torrential character originated by severe convective storms, which are favored by interactions between orography and moist fluxes from Mediterranean. This explains the larger impact of the grid resolution (IB02 versus IB1) on the scores. Moreover, the poorest performance of ECMWF reanalyses in these regions reflects the difficulty of NWP models to correctly simulate such processes.
 In NW and SW regions, including Portugal, ERA-40 and ERA-I perform remarkably well, with squared correlations larger than 0.78, as illustrated in Figure 9. In these areas, precipitation is caused mostly by depressions and frontal systems, such good performance reflects the high skill of ECMWF models to represent these systems. The prevalence of such weather regimes also explains the smallest sensibility of their scores to the grid resolution.
Figure 9 also shows that ERA-I performs better than ERA-40. The exception is the Cantabrian region, where ERA-I strongly overestimates the frequency of wet days. In CANT, the use of IB1 instead of IB02 grid also contributes significantly to increase the scores, consistently with the importance of orography for the precipitation in this region.
 Similarly to the precipitation from ECMWF reanalyses, the best agreement between global observational data sets and IB02 is found in Portugal and in NW and SW regions, where precipitation is caused mostly by synoptic-scale disturbances. Consequently, in these areas one station may have a large spatial representativeness. However, in other regions of Iberia where precipitation is frequently associated to convective regimes, being therefore more localized, it is more difficult to obtain a good estimate with a sparse gauge network (such as that used by CRU) or with the coarser resolution (0.5°) commonly used by global data sets.
Figure 9 also clearly shows that monthly precipitation is better represented by GPCC than by CRU. This difference is larger in CANT, NE and SE regions, where the gauge density of GPPC is considerably better than that from CRU. Moreover, an important result is that ERA-I outperforms CRU in all regions and ERA-40 performs better than CRU over Portugal and Cantabrian region.
3.1.3. Drought Detection
 A large part of the Iberian Peninsula is characterized by high temperatures during summer and by a long dry season; more than 75–90% of the annual precipitation occurs from October to May. As a consequence, the droughts that affect frequently the Iberian Peninsula [Vicente-Serrano, 2006b], have strong impacts in water resources, fire risk and ecosystem degradation.
 In this section, IB02 is used to assess the performance of global precipitation data sets when they are applied to drought detection, using SPI, which is one of the simplest drought indicators [Vicente-Serrano and López-Moreno, 2006; Vicente-Serrano, 2006b; Santos et al., 2010]. For each location (grid point or region), SPI is computed monthly, by fitting a probability density function to the frequency distribution of precipitation summed over a time window corresponding to the time scale of interest [Vicente-Serrano, 2006a; McKee et al., 1993]. The SPI with 12 month time window (denominated SPI-12) is analyzed here, since it represents long-term precipitation anomalies with impact on water resources.
 The precipitation data sets are compared in terms of drought detection in Figure 10. There is, in general, a good agreement between the five data sets, in the identification of the major drought spells in the regions illustrated in Figure 10. It is possible to identify nine major drought spells affecting large areas in Iberia: 1953–1954, 1957–1959, 1965, 1976, 1981–1982, 1992–1993, 1995–1996, 1999–2000 and 2002–2003. In 2005–2006, the available data sets (CRU, GPCC and ERA-I) agree on a generalized drought spell in all areas. The squared correlations of SPI-12 between IB02 and the global data sets computed at IB05 grid are presented in Table 1. For NW, SW and Iberia all data sets have a very good performance, with r2 higher than 0.9. In the other regions, GPCC performs clearly better than the other data sets, maintaining values of r2 above 0.9. In NE and CANT regions, CRU performs better than ERA-40 and ERA-I, where the later shows a poor performance in CANT. In SE region, CRU performs only slightly better than ERA-I.
Table 1. Squared Correlation (r2) of SPI-12 From IB02 Versus ERA-40, ERA-I GPCC, and CRU for Five Regions and Iberia, Computed at IB05 Grid, for the Period 1990–2001
3.2. Five Days of Precipitation
3.2.1. Objective Scores
 The squared correlation between precipitation from ERA-I and IB02 data set, computed at IB02 grid, excluding days without precipitation (bars) and with all the days (impulses), for several Iberian regions and seasons is presented in Figure 11 (top). It is clear that when dry days are excluded the performance of ERA-I decreases for all areas and seasons, with largest impact in SE and NE regions and during driest months. On average, over the Iberian Peninsula r2 drops from 0.74 to 0.68 during winter and from 0.5 to 0.4 during summer.
 In general, ERA-I performs better during winter (when r2 exceeds 0.5 for all areas) and worse during summer. In Portugal and NW region, the skill of ERA-I is larger in SON than in MAM months. In the other regions, the opposite is verified. In particular, in SE and NE regions, the performance of ERA-I is poorest during JJA and SON months, reflecting model deficiencies in the simulation of isolated thunderstorms and mesoscale convective systems. On the other hand, ERA-I performs remarkably well over Portugal.
 The comparison between Figures 9 and 11 (top) shows that the explained variance of monthly ERA-I precipitation is larger than the corresponding quantity for 5 day accumulations. On average, over the Iberian Peninsula, it decreases from about 70% for monthly precipitation to about 60% for 5 days of precipitation.
 The squared correlations between the time series of 5 days of accumulated precipitation from IB02 and from ERA-40 and ERA-I are computed for each grid box and averaged over the Iberian Peninsula for each year. The time evolution of this score is illustrated since 1960 for ERA-40 and since 1990 for ERA-I in Figure 11 (bottom). A slight positive trend is noticeable for both ERA-40 and ERA-I. The poorest performance of ERA-40 occurs in 1971 and 1972. These 2 years were characterized by large values of RMSEN (of the order of 0.9) over Iberia, clearly above typical values for all other years (∼0.65–0.75). During these 2 years, positive biases of the order of 3 mm/5 d, on average, are found in the NE and SE regions (not shown). This behavior differs from the remaining period, when ERA-40 tends to underestimate precipitation in all regions of the Iberian Peninsula.
 The better performance of ERA-I relative to ERA-40 for all years is also visible in Figure 11. Moreover, it is evident that computing r2 at IB1 grid instead of at IB02 grid leads to an increase in the explained variance of the order of 15% to 25%, on average over the Iberian Peninsula, for ERA-40 and ERA-I.
 The spatial patterns of 5 days of precipitation from ERA-40 and ERA-I are evaluated using CR. The dependency of this score on several thresholds (1, 5, 10, 20, 30, 50 and 60 mm) is illustrated in Figures 12a and 12b for winter and summer, computed at IB02 and at IB1 grids, respectively. For both data sets and for both grids, as the threshold value increases, the value of CR decreases, indicating that the agreement between reanalyses and observations decreases for larger precipitation amounts. For instance, in winter, CR values decrease by a factor of more than 4, when the threshold increases 5–50 mm/5 d. Moreover, the values of CR increase 20 to 25% when its computation is done at IB1 grid rather than at IB02 grid.
 ERA-I outperforms ERA-40, as demonstrated by the larger values of CR, and both reanalyses performer better in winter than in summer. These seasonal differences are larger than differences in skill between ERA-40 and ERA-I. This is more obvious when using IB02 rather than IB1 grid (Figure 12). The values of CR in autumn and spring are slightly smaller than in winter and are larger than in summer (not shown), which is consistent with results found using the squared correlation (see Figure 11).
 The poor skill of NWP models in predicting precipitation associated with convective systems can explain the ECMWF reanalyses difficulties to predict large precipitation amounts, mainly during summer and for ERA-40, when CR values were found to be lower than 0.1 for a threshold of 50 mm/5 d. However, Figure 12 shows that values of CR computed using P75o and P75e are close to those computed for thresholds of 1–10 mm/5 d, demonstrating a reasonable skill of ERA-I and ERA-40 to depict the position of the highest precipitation amounts, despite their difficulties to predict large precipitation amounts.
3.2.2. Case Studies
 The use of CR to evaluate the skill of 5 days of precipitation of ECMWF reanalysis is illustrated using ERA-I for two specific 5 day periods (Tables 2 and 3). The first period includes a sequence of rainy and no-rain days, and the second period comprises a sequence of days with intense precipitation (including an extreme event). The corresponding patterns of accumulated precipitation are presented in Figure 13.
Table 2. Values of 75th Percentile of the Five Days of Precipitation From All Grid Points Over the Iberian Peninsula, Computed Using IB02 and ERA-I Data Sets, for Two Different Periods: 19–23 May 1997 and 2–6 November 1997
Table 3. Values of Correspondence Ratio Computed for 5 Days of Accumulated Precipitation From IB02 and ERA-I for Two Different Periods: 19–23 May 1997 and 2–6 November 1997
 Between 18 and 23 May 1997, most of the accumulated precipitation over the Iberian Peninsula was caused by a cold front associated with a depression centered south of British Islands between 19 May 1997 and 20 May 1997. ERA-I captures reasonably well the area of precipitation (CR = 0.74), when considering a threshold of 5 mm/5 d (Table 3). However, the precipitation maxima, between 50 and 100 mm, found in Galicia and northern Portugal, were underestimated leading to values of CR equal to zero, when this skill score is computed with threshold values above 30 mm/5 d. For this situation, the value of CR, computed using P75o and P75e, is 0.68 (Tables 2 and 3), showing that the skill of ERA-I to capture the location of maximum precipitation (when neglecting its negative bias) is similar to its skill to depict precipitation amounts above 10 mm/5 d.
 During the second period, between 2 and 4 November 1997, up to 100 mm of accumulated precipitation were recorded in southwest of the Iberian Peninsula, due to a depression crossing the western Iberia from SW to NE. Between 5 and 6 November 1997, a deep cyclone, which reached a minimum center pressure of 981 hPa, produced heavy precipitation (up to 120 mm) over southern Portugal and Badajoz region, causing flash floods and 34 casualties [Fragoso and Tildes Gomes, 2008]. The maximum accumulated precipitation during the entire period was around 250–300 mm. These values were underestimated by ERA-I, which estimates maximum values smaller than 150 mm. This result is not surprising taking into account the small scale of the referred storm and the ERA-I model resolution.
 For the two periods presented in Figure 13, CR values computed using thresholds from 1 to 50 mm, suggest that the performance of ERA-I was better in the period from 2 to 6 November 1997 than in May 1997. However, when using CR computed with thresholds based on P75o and P75e (Tables 2 and 3) the opposite conclusion can be drawn.
 A new high-resolution daily precipitation gridded data set over mainland Portugal, PT02, was presented. This data set is combined with a recent Spanish gridded data set [Herrera et al., 2010] to create a high-resolution (0.2° × 0.2°) Iberian data set, which covers the period from 1950 to 2003. This new data set, labeled IB02, is based on a dense network of rain gauges, with several thousand stations over Spain and several hundred stations over Portugal.
 The ordinary kriging method was compared with two inverse distance schemes over mainland Portugal, showing a small improvement over simpler techniques. For this reason and for consistency with Spain02 data set, the kriging method was applied to create the PT02 data set. Although IB02 is a combination of two different data sets, albeit with a common grid, there is no evidence of artificial features at the border between Portugal and Spain, neither at the monthly scale nor at finer scales, including daily results. When analyzing the monthly observational data sets against IB02, GPCC performs clearly better than CRU in several aspects, including its ability to represent the spatial distribution of the mean annual precipitation and the annual cycle. The higher skill of GPCC relatively to CRU data set is also confirmed by its smaller biases and by the squared correlations between the IB02 and the two global data sets.
 A better agreement between global observational data sets and IB02 was found in NW and SW regions, where precipitation is mainly associated with synoptic-scale disturbances, in contrast with other regions of Iberia where convective events and complex orography play a key role. This result suggests that the accuracy of a precipitation data set is more influenced by the density of the rain gauge network and by its spatial resolution in NE, SE and CANT regions than in other areas, such as NW and SW, where one station appears to have a broader spatial representativeness.
 In general, this study shows that, despite a similar performance in some aspects, ERA-I outperform ERA-40 over the Iberian Peninsula. The northwest-southeast contrasts in the geographical distribution of mean annual precipitation are well captured by the two reanalyses. However, in the vicinity of mountainous regions, the totals amounts are strongly underestimated, mainly by ERA-40. This is consistent with the misrepresentation of the sharp orographic features of Iberia due to the coarser resolution used by ERA-I and in particular by ERA-40.
 The strong annual cycle over the Iberian Peninsula is reasonably well captured by the two reanalyses. However, its amplitude is underestimated by both data sets. ERA-40 shows a systematic underestimation, which is stronger in rainy season and in CANT region. In contrast, ERA-I reduces this dry bias in rainiest months, but it shows a tendency to slightly overestimate precipitation from May to September. On average over the Iberian Peninsula, the frequency of days with precipitation is overestimated by the two reanalyses. This overestimation is stronger in ERA-I, mostly during spring and summer and over Cantabrian and NE regions.
 In general, the better performance of ERA-I relatively to ERA-40 is confirmed by the squared correlations between IB02 and the two ECMWF reanalyses. In particular, both ERA-40 and ERA-I perform remarkably well in NW and SW regions. In the NE and SE, the reanalyses performance is far worse, reflecting the difficulty of NWP models to simulate the convective precipitation. In the Cantabrian region, contrarily to the other regions, ERA-40 outperforms ERA-I, possibly because of the tendency of ERA-I to strongly overestimate the number of wet days in this region. An important finding of this study is that, in general, the performance of the CRU data set is no better than the ERA-I reanalysis.
 Global precipitation data sets are commonly used in regional and global drought studies [e.g., Lloyd-Hughes and Saunders, 2002, and Vicente-Serrano et al., 2010]. While GPCC, CRU and IB02 are not updated frequently, ERA-I is produced in near-real time. This advantage of the reanalysis could be explored for drought monitoring purposes. This is supported by our findings that in terms of drought detection (indicated by SPI-12) all data sets clearly identify the major drought spells at the scale of Iberia.
 The performance of 5 days of accumulated precipitation from ERA-I and ERA-40 was also assessed using simple scores, as well as the correspondence ratio [Stensrud and Wandishin, 2000], which is defined as the ratio between the areas where both data sets depict precipitation and the area where at least one of the data sets exhibit precipitation.
 The squared correlation between the 5 days of precipitation from ERA-I and IB02 illustrate a seasonal cycle in the ERA-I performance, with the lowest skill in JJA and higher skill in DJF months. The exception is the SE region, where the poorest performance occurs in SON months. In addition, this skill score indicates that, on average, ERA-I outperforms ERA-40 since the beginning of the ERA-I period until 2001. Moreover, both reanalyses reveal a poor skill to predict the magnitude of precipitation for higher amounts (superior to 20 mm/5 d), mostly in summer. Nevertheless, ERA-40 and ERA-I demonstrate a greater ability to estimate correctly the peak locations, with ERA-I being superior to ERA-40.
 Given the large variety of precipitation regimes in the Iberian Peninsula, and a strong dependency of precipitation on orography, IB02, a combination of PT02 and Spain02 [Herrera et al., 2010] data sets, will be a valuable contribution to the validation of increasingly high resolution reanalyses and climate simulations, as well as, for hydrological applications and others. However, as mentioned by Herrera et al. , for purposes of regional trend analysis the users should be careful, because of the high temporal variability of the network density used to create this data set.
 The PT02 data set is freely available for noncommerical purposes (contact M. Belo-Pereira to obtain data access instructions).
 The priorities for future works are: to extend the PT02 data set to 2007, to improve the quality control procedures applying the methodology described by Štepánek et al. , to use the indicator kriging approach to improve the determination of the precipitation state (wet or dry) of the grid box and the application of multivariate versions of kriging. The multivariate algorithms allow to incorporate elevation into the gridding of precipitation, which could have a positive impact on the interpolation accuracy [Goovaerts, 2000], especially in topographical complex areas.
 The authors thank the Spanish Meteorology Agency (AEMET) and Santander Meteorology Group (UC) by Spain02 gridded precipitation data set used in this work. Raw precipitation data over Portugal were provided by the Portuguese Institute of Meteorology (IM) and by the Portuguese Institute of Water (INAG). The authors thank Fátima E. Santo and Álvaro Silva (IM, Portugal) for support in the development, quality control of the station data, and provision of ancillary data. The authors are also thankful to Nuno Moreira (IM, Portugal) and three anonymous reviewers for the constructive comments that helped to improve the manuscript. E. Dutra and P. Viterbo were supported by the Portuguese Foundation for Science and Technology (FCT) under project AMIC PTDC/AAC-CLI/109030/2008 cofinanced by the European Union under program FEDER.