Precipitation–temperature relationships over Europe in CORDEX regional climate models
Funding information: Grantová Agentura České Republiky, Grant/Award Number: 20-28560S; Ministerstvo Školství, Mládeže a Tělovýchovy, Grant/Award Numbers: CZ.02.1.01/0.0/0.0/16_019/0000797, LTC19044
Abstract
We studied spatial and temporal patterns of precipitation–temperature (P–T) relationships through correlations between monthly standardized precipitation index (SPI) and monthly temperature anomalies in individual climatic seasons over Europe. In the observed data (represented by E-OBS), positive correlations (wet–warm/dry–cold relationships) prevail during winter over most of Europe, while negative values (dry–warm/wet–cold) are dominant in summer. In the next step, an ensemble of seven regional climate models (RCMs) from the CORDEX project driven by the ERA-Interim reanalysis were examined as to their reproduction of the regional patterns of the P–T correlations. In winter, the RCMs yielded overly strong positive P–T correlations over northern Europe, while the correlations were too weak in the south compared to observed data. During summer, the biases were generally larger; the RCMs were able to capture the overall negative P–T correlations but these tended to be too weak over northern Europe. This deficiency was found to be linked to simulated differences in shortwave radiation (a proxy for cloud cover) between dry and wet months. In western, central, and southeastern Europe, by contrast, most RCMs yielded too strong negative correlations in summer, and overly large decreases of relative humidity during dry months probably contributed to these errors. The results pointed up issues that should be addressed as the reported RCMs' deficiencies may lower credibility of projected compound dry–hot events in climate change scenarios.
1 INTRODUCTION
Precipitation is linked to air temperature by various physical processes. The maximum amount of wintertime daily precipitation is often determined by atmospheric moisture-holding capacity according to the Clausius–Clapeyron relation, while in summer, the availability of moisture is usually a limiting factor in middle latitudes (Berg et al., 2009). Correlation between precipitation and temperature (P–T) therefore depends on climatic season and region analysed. On the global scale, Wu et al. (2013) identified positive P–T correlations (wet–warm/dry–cold relationships) over equatorial Pacific, Atlantic, and Indian oceans during the whole year. In boreal summer, negative P–T correlations (dry–warm and wet–cold) dominated in Northern Hemisphere midlatitudes, while indistinct patterns were observed in the Southern Hemisphere. During boreal winter, by contrast, positive P–T correlations were found over large parts of Eurasia and North America, while negative values were located over South America, Australia, and southern parts of Africa (Wu et al., 2013). The spatial resolution of data in their study was rather coarse (2.5 × 2.5°), however, and captured regional P–T features with difficulties due to windward/leeward effects, land–ocean interactions, and so on.
Temperature, precipitation, and their relationships are the main characteristics of a regional climate (e.g., in Köppen–Geiger climate classification; Kottek et al., 2006) and key meteorological variables in climate change adaptation planning. Although changes in a single variable may have considerable impacts on society and ecosystems, these impacts tend to be larger when effects of two or more drivers combine, resulting in hazardous compound events (Zscheischler et al., 2020). For instance, summertime drought-and-heat events are associated with large negative impacts on food production (Zampieri et al., 2017), ecosystems (Sippel et al., 2018), water resources, and electricity production (van Vliet et al., 2016). Another example consists in wintertime rain-on-snow events (Cohen et al., 2015). Combined with a rapid increment of temperature, these compound events often trigger hazardous and hardly predictable flooding (Rössler et al., 2014).
Credible scenarios of future changes in compound events require a high-quality ensemble of climate models capable of resolving smaller-scale processes; simulations of regional climate models (RCMs) from the CORDEX project (Giorgi and Gutowski, 2015) represent such ensembles. Kotlarski et al. (2014) evaluated the CORDEX RCMs in terms of temperature and precipitation. When driven by the ERA-Interim reanalysis (Dee et al., 2011), they tended to overestimate mean summertime temperature in the Mediterranean and southeastern Europe, while negative temperature biases were found over northern parts of the domain. In winter, the biases were rather negative, however, a relatively large intermodel variability was found. Lhotka and Kyselý (2018) associated the wintertime temperature biases to simulation of atmospheric circulation and found too cold westerly and northerly advection in the CORDEX RCMs. Regarding precipitation, Kotlarski et al. (2014) concluded that a majority of the RCMs had positive precipitation biases in both summer and winter (averaged across Europe). These positive biases were more pronounced in northern parts of the European domain and tended to be larger in those RCMs running in finer (~12.5 km) resolution.
Precipitation is one of the key factors modulating severity of heat waves. Vautard et al. (2013) found too severe and persistent heat waves in the CORDEX RCMs and linked this deficiency to possibly exaggerated land–atmosphere feedbacks. A detailed analysis for central Europe (Lhotka et al., 2018) revealed suspicious driving mechanisms of major heat waves in the CORDEX RCMs within this region. These errors were related to simulation of large-scale circulation and land–atmosphere interactions. Those RCMs that substantially overestimated the severity of heat waves had the highest Bowen ratio (the ratio between sensible and latent heat flux) and one of the largest monthly sums of downwelling shortwave radiation within the analysed ensemble (Lhotka et al., 2018). The essential role of precipitation and land–atmosphere feedbacks in general on summertime temperature maxima was also shown by other modelling studies (e.g., Jaeger and Seneviratne, 2011; Davin et al., 2016).
In as much as capability of RCMs to reproduce P–T relationships is crucial for credible simulation of events such as heat waves or droughts in a possible future climate, we evaluated an ensemble of CORDEX RCMs, which commonly have been used for constructing climate change scenarios over Europe (e.g., Coppola et al., 2021). The role of RCMs' spatial resolution (~12.5 and 50 km) was also studied. The modelled summertime P–T correlations were then linked to shortwave radiation and relative humidity and their differences between dry and wet months, in order to study possible physical mechanisms contributing to RCMs' deficiencies.
2 DATA AND METHODS
2.1 Observed data and regional climate models
Observed daily precipitation sums (P) and mean daily temperatures (T) were taken from the E-OBS 22.0e dataset (0.1° grid; Cornes et al., 2018). Although the data span from 1950 to 2019, we used the 1989–2008 period only, in order to allow a direct comparison with evaluation (reanalysis-driven) runs of the CORDEX RCMs (having the same common length). During this period, E-OBS data covered the entirety of Europe except for Iceland, eastern parts of Ukraine, Sicily, and Greece. In addition, daily mean downwelling shortwave radiation flux (SR) and mean relative humidity (RH) were taken from the same dataset for analysing possible physical mechanisms contributing to P–T correlation errors in summer.
We analysed seven individual RCMs from the CORDEX project driven by the ERA-Interim reanalysis (Table 1). All simulations were provided in both higher (~12.5 km; EUR-11 domain) and lower (~50 km; EUR-44 domain) spatial resolution. Various schemes of radiation, convection, microphysics, land-surface, and boundary layer were implemented in the RCMs. A detailed list including references is available in Vautard et al. (2013).
| Institution | RCM | Acronym |
|---|---|---|
| CLM | CCLM4-8-17 | CLM-CCLM |
| CNRM | ALADIN53 | CNRM-ALADIN |
| DMI | HIRHAM5 | DMI-HIRHAM |
| IPSL | WRF331F | IPSL-WRF |
| KNMI | RACMO22E | KNMI-RACMO |
| MPI | REMO2009 | MPI-REMO |
| SMHI | RCA4 | SMHI-RCA |
2.2 Remapping to common grid
While the E-OBS data are provided in a latitude–longitude grid with 0.1° spacing, six out of the seven RCMs are mapped into a rotated pole grid. In order to evaluate spatial patterns of P–T relationships against E-OBS and make all data comparable, the E-OBS data were remapped into the ~50 km rotated pole grid (the EUR-44 domain).
RCM simulations in higher resolution were remapped into the lower resolution grid, too. This was performed through averaging P and T values in 16 nearest grid points from the respective grid point in the EUR-44 domain. After this remapping, the EUR-11 RCMs still retain P and T features obtained through resolved-scale dynamics (e.g., large-scale convection, windward effects; Prein et al., 2016). CNRM-ALADIN, which uses the Lambert conformal projection, was additionally remapped into the rotated pole grid by the nearest neighbour method (P and T values were assigned to the nearest grid point in the rotated pole grid).
The simple remapping technique cannot be applied to E-OBS, because in this dataset, an area represented by a grid box depends on latitude and therefore it decreases towards the North Pole in the regular latitude–longitude grid (distances between meridians decrease poleward). Thus, P and T values assigned to the individual EUR-44 grid points were calculated as averages of 25 (5 × 5), 35 (7 × 5), 45 (9 × 5), or 55 (11 × 5) grid points of the original E-OBS latitude–longitude grid. This procedure was similar to that used in Kotlarski et al. (2014), who also calculated the value of a target grid point by an area-weighted average of all overlapping grid points from the original (finer) grid. Analogous procedures were performed for SR and RH variables.
2.3 Calculation of the standardized precipitation index and its link to temperature
Daily precipitation values were aggregated into monthly sums, in order to compute SPI (Guttman, 1998)—a drought index that does not require temperature data (to avoid an additional statistical dependency in this study). SPI was calculated without incorporating precipitation from previous months (SPI-1) and was calibrated in the 1989–2008 period. The function standardizes precipitation sums in individual months following a log-logistic distribution (i.e., all Januaries are treated separately). In short, a negative SPI value indicates drier conditions during a given month compared to the average for the same month in the historical period, and vice versa.
Anomalies of monthly temperatures against the 1989–2008 climatology were correlated with monthly SPI values using the Pearson correlation coefficient (r). Positive P–T correlations indicate dry–cold and wet–warm relationships, while negative r represents dry–hot and wet–cold links. Regions with negative r are prone to droughts especially in summer, during which high temperatures amplify evapotranspiration. On the seasonal scale, monthly SPI and temperature anomalies were averaged for a respective season and these seasonal values were thereafter correlated. We examined also the use of SPI-3 (Guttman, 1998) for analyses on the seasonal scale, with analogous results. The significance of correlations was tested through a parametric correlation test (Best and Roberts, 1975) and evaluated at the 5% significance level.
2.4 Precipitation–temperature relationships in European regions
In order to analyse RCMs' differences in P–T relationships in more detail, the European domain was divided into five regions (Scandinavia – SC, eastern Europe – EE, western Europe – WE, central Europe – CE, and the Mediterranean – MD; Figure 1) that represent a simplification of Köppen–Geiger climate classification (Kottek et al., 2006). SC is characterized by “boreal fully humid climate with cool summers” (type Dfc) with patches of Arctic Tundra. EE is dominated by “fully humid warm summer continental climate” (type Dfb). WE and CE regions are mainly represented by “fully humid warm temperate climate” (type Cfb) but they were analysed separately because the CE climate results from interaction of both maritime and continental air masses. Finally, MD is mainly characterized by “warm temperate climate with dry and hot summers” (Csa). The number of continental grid points (common for E-OBS and remapped RCMs) in the individual regions ranged from 283 to 776 and the represented land area from 0.69 to 1.81 mil km2 (Table 2).
| Name | Abbr. | Xmin (°E) | Xmax (°E) | Ymin (°N) | Ymax (°N) | Number of grid points | Area (mil. km2) |
|---|---|---|---|---|---|---|---|
| Scandinavia | SC | 4 | 36 | 55 | 71.5 | 776 | 1.81 |
| Eastern Europe | EE | 28 | 45 | 43 | 60.5 | 451 | 1.08 |
| Western Europe | WE | −13 | 6 | 45 | 60 | 283 | 0.69 |
| Central Europe | CE | 6 | 28 | 45 | 55 | 700 | 1.67 |
| Mediterranean | MD | –13 | 35 | 34 | 45 | 633 | 1.49 |
- Note: Xmin, Xmax, Ymin, and Ymax represent coordinates of a respective region's corner.
2.5 Role of shortwave radiation and relative humidity
Differences in P–T relationships between the RCMs and observed data in summer (months June, July, and August) were linked to simulation of SR and RH. These variables were obtained from the E-OBS dataset (Cornes et al., 2018) and serve as proxies for land–atmosphere characteristics, which can substantially alter regional climate in summer (Jaeger and Seneviratne, 2011).
Besides studying summertime SR and RH climatology in the RCMs and observed data, differences in mean SR and RH between dry and wet months were assessed. Dry months were determined separately for each grid point and dataset, that is, the precipitation threshold below which an SPI value is considered “dry” varies across Europe and in individual RCMs or E-OBS. In order to obtain data samples of the same length from all grid points and datasets, 20 lowest SPI values (out of 60; 20 years in 1989–2008 × 3 summer months) were extracted from individual grid points. Analogously, 20 wet months were marked based on the highest SPI values.
It should be noted that the definition based on the ranked SPI values and fixed number of dry and wet months may result in relatively small differences in actual precipitation rates between dry and wet months, especially in those RCMs with low month-to-month variability compared to E-OBS. Although this drawback might affect simulated differences in SR between dry and wet months (used as a proxy for cloud cover variations), the employed methodology based on the ranked SPI values overcomes issues with various European climate zones (with very different precipitation rates) and biases in the RCMs, which would result in lack of dry months in some regions/RCMs and too many dry months in others.
Another factor that possibly affects the P–T relationships is the soil moisture content and associated ratio between sensible and latent heat fluxes (the Bowen ratio). Because this characteristic cannot be obtained from E-OBS, RH was used. A substantial decrease of RH may suggest a depletion of soil moisture and consequently increase of the Bowen ratio. RH was available directly for five out of the seven RCMs; for CLM-CCLM-ERA and MPI-REMO-ERA it was calculated via cranR package “humidity” from mean daily specific humidity and temperature. Finally, for each grid point, SR and RH values were averaged separately for dry and wet months and differences between these two samples were plotted.
3 PRECIPITATION–TEMPERATURE (P–T) RELATIONSHIPS IN OBSERVED DATA
A distinct annual cycle of the P–T relationships was found in the E-OBS data (Figure 2). January was characterized by widespread positive P–T correlations over Europe, except for Italy, the Alps and southeastern Europe. A similar spatial pattern was observed also for February, however, in contrast to January, the correlations turned negative in the Iberian Peninsula. In March, a shift to negative P–T correlations was found in regions south of 50°N and in the British Isles. The area affected by positive P–T correlations was gradually reduced in the following months and negative P–T correlations dominated in June over the whole of Europe. This regime persisted until September, when positive P–T correlations started forming over Scandinavia and the British Isles and these propagated southwards in the following months. In December, the annual cycle returned to positive P–T correlations across Europe.
Statistical significance of the correlation coefficient was tested on the seasonal scale. In winter, significant (p = .05) positive P–T correlations were found in 37% of the analysed area (primarily over the northern half of the domain), while significant negative r values were limited to small patches around 45°N and 25°E (Figure 3). The mean r over Europe was 0.30 in winter season (interquartile range 0.08–0.56). In spring, by contrast, significant positive P–T correlations were reduced to several grid points but significant negative r values started forming over southern Europe. Turning to summer, approximately half of Europe was covered by significant negative P–T correlations that contributed to the mean r value of −0.42, with interquartile range from −0.54 to −0.32. Finally, in autumn, a dipole pattern of significant positive P–T correlations over northeastern Europe and significant negative P–T correlations in southeastern Europe was found. In general, the P–T correlations were significant mainly during summer (negative) and winter (positive), while their significance was lower in transition seasons.
4 EVALUATION OF REGIONAL CLIMATE MODELS
In this section, the capability of the RCMs to reproduce the P–T relationships over Europe was evaluated. In winter (Figure 4a), the RCMs tended to yield larger r over SC and EE, which resulted in overly strong positive P–T correlations on average over those regions compared to E-OBS. By contrast, all RCMs (except for SMHI-RCA) had smaller r in MD compared to E-OBS, with mean r close to 0 or even negative values. In WE and CE, r was reproduced reasonably well considering the model mean but relatively large variance among the RCMs was found. For example, the mean r for CE was 0.0 for DMI-HIRHAM, while SMHI-RCA yielded 0.4, considering EUR-11 models. No substantial differences between model means from the EUR-11 and EUR-44 RCMs were found and the mean value of r was similar in both resolutions (Figure 4a).
During spring, negative r formed across Europe in E-OBS with north–south gradient (strength of the negative P–T correlations increased towards south; Figure 3). The majority of the RCMs were able to reproduce the mean r in the SC and EE regions but negative P–T correlations tended to be too strong in WE, CE, and MD (Figure 4b). This error was especially pronounced in WE, where all RCMs (regardless their spatial resolution) had smaller r compared to E-OBS. Variance in the mean r between the individual RCMs was larger than in winter mainly due to SMHI-RCA, which yielded overly strong positive P–T correlations in SC, EE, and CE regions.
In summer, dry–hot/wet–cold regime (negative r) is developed through Europe in E-OBS (Figure 3). The RCMs were able to reproduce this general pattern, but the negative P–T correlations were too weak over SC and EE in the majority of RCMs compared to E-OBS (Figure 4c). By contrast, all EUR-11 RCMs yielded overly strong negative P–T correlations in WE (ranging from −0.34 to −0.58, compared to −0.33 in E-OBS) and in CE (−0.47 to −0.75 vs. –0.41 in E-OBS). Considering the spatial pattern of r deviations (model mean; Figure 5), their negative values, which indicated more pronounced dry–hot/wet–cold regime in this season, were found approximately south of 55°N, while positive r deviations were located further north. The only exceptions were negative r values over the Kola Peninsula in the northeast and positive r over the Iberian Peninsula in the southwest. The latter anomaly was related to seemingly good reproduction of r over the MD region (Figure 4c), which was caused by a compensatory effect of positive r deviations in the west and negative ones over the eastern parts of the region (Figure 5).
Five out of seven individual RCMs shared the distinct latitudinal pattern of r deviations in summer, which was especially pronounced in CNRM-ALADIN and SMHI-RCA (Figure 5). CNRM-ALADIN simulated the most pronounced dry–hot/wet–cold regime across WE and CE, while SMHI-RCA yielded mean r close to 0 over the SC region. The latitudinal pattern was less distinct in DMI-HIRHAM and not present in CLM-CCLM, which produced too strong negative P–T correlation all over Europe (Figure 5), amplifying summertime dry–hot/wet–cold regime. No systematic differences between EUR-11 RCMs and their lower-resolution counterparts were found and the patterns of the P–T correlation biases were similar to those simulated by the EUR-44 RCMs (Figure S1, Supporting Information).
During autumn, the overall differences between the model mean and observed data were smallest (Figure 4d). SMHI-RCA yielded highest r in all five regions analysed (considering its EUR-11 version), a feature present also during winter and spring seasons. By contrast, MPI-REMO had one of the lowest mean r values among the RCMs in winter, spring, and autumn (Figure 4). Similar RCMs' behaviour was not found in summer, probably due to growing importance of smaller-scale processes (convection, land–atmosphere interactions, boundary layer properties), which possibly led to the largest r biases in this season (Figure 4c).
5 POSSIBLE PHYSICAL MECHANISMS BEYOND BIASES IN THE SUMMERTIME P–T RELATIONSHIPS
The biases in the summertime P–T relationships may considerably alter temperature characteristics of simulated dry episodes. The too weak negative P–T correlation in majority of the RCMs over SC and EE is manifested in lower temperature anomalies during dry months (see section 2.5) in these regions (Table 3). The RCMs with too low temperature in dry months (compared to E-OBS) analogously simulated too high temperature in wet months and vice versa. By contrast, the overly strong negative P–T correlations in WE and CE were associated with too high temperature anomalies in dry months, as simulated by all RCMs (except for IPSL-WRF in CE). This temperature anomaly was especially high in CNRM-ALADIN compared to E-OBS, that is, in the RCM with the lowest mean r in those regions (cf., 0.99 vs. 0.36°C in WE and 1.45 vs. 0.71°C in CE; Table 3).
| Dry months | Wet months | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| SC | EE | WE | CE | MD | SC | EE | WE | CE | MD | |
| CLM-CCLM | 0.90 | 0.96 | 0.77 | 0.97 | 0.93 | −0.77 | −0.89 | −0.64 | −0.82 | −0.96 |
| CNRM-ALADIN | 0.35 | 0.85 | 0.99 | 1.45 | 0.71 | −0.33 | −0.83 | −0.96 | −1.30 | −0.76 |
| DMI-HIRHAM | 0.50 | 0.45 | 0.55 | 0.79 | 0.73 | −0.46 | −0.38 | −0.53 | −0.74 | −0.73 |
| IPSL-WRF | 0.31 | 0.22 | 0.55 | 0.67 | 0.41 | −0.30 | −0.15 | −0.32 | −0.55 | −0.38 |
| KNMI-RACMO | 0.57 | 0.55 | 0.67 | 0.94 | 0.63 | −0.48 | −0.40 | −0.61 | −0.87 | −0.66 |
| MPI-REMO | 0.51 | 0.52 | 0.48 | 0.88 | 0.75 | −0.38 | −0.38 | −0.54 | −0.78 | −0.74 |
| SMHI-RCA | 0.07 | 0.36 | 0.64 | 0.81 | 0.47 | −0.06 | −0.28 | −0.60 | −0.66 | −0.46 |
| MODEL MEAN | 0.46 | 0.56 | 0.66 | 0.93 | 0.66 | −0.40 | −0.47 | −0.60 | −0.82 | −0.67 |
| E-OBS | 0.65 | 0.61 | 0.36 | 0.71 | 0.54 | −0.59 | −0.54 | −0.39 | −0.59 | −0.62 |
The realistic reproduction of P–T relationships is therefore crucial for a proper simulation of hazardous dry and hot compound events (Zscheischler et al., 2020). In this section, we examine the RCMs' biases in relation to SR and RH, that is, meteorological variables linked to dry and hot compound events. Although CLM-CCLM, DMI-HIRHAM, and MPI-REMO were able to reproduce the spatial pattern of mean SR in summer, the rest of the RCMs substantially overestimated mean SR in some regions at least (Figure 6). CNRM-ALADIN and IPSL-WRF considerably overestimated the mean summertime SR across Europe, and SMHI-RCA in the MD region.
The amount of SR differed between dry and wet months in E-OBS and this difference was largest (up to 50 W·m−2) over northeastern Europe (roughly 55°–60°N and 10°–35°E; hereafter referred as the Baltic region; Figure 7). The majority of RCMs yielded too low differences between dry and wet months over this area, associated with weaker negative P–T correlation in SC and EE. These weaker correlations were not present in CLM-CCLM, which simulated larger differences in mean SR between dry and wet months over northern Europe (Figure 7). We also note that the two RCMs with the largest overestimation of SR across Europe (Figure 6) had by far the smallest (and mostly negligible) differences in SR between dry and wet months. This is probably related to large deficiencies in simulated cloud cover characteristics in CNRM-ALADIN and IPSL-WRF in summer.
Over the aforementioned Baltic region, CNRM-ALADIN yielded too strong negative P–T correlations despite the fact that this RCM had smaller differences in mean SR between dry and wet months compared to E-OBS (Figure 7). This anomaly was associated with the reproduction of RH. In general, the RCMs simulated too pronounced north–south gradient of RH, that is, they tended to have overly high RH over the northern part of the domain, while too low RH in the south (Figure 8). CNRM-ALADIN yielded the lowest RH in the Baltic region, which probably contributed to overly strong negative P–T correlations in this area (Figure 5).
In addition, the too low r values were also linked to large differences in RH between dry and wet months. Although magnitude and spatial pattern of these differences were captured realistically in model mean (Figure 9), CNRM-ALADIN simulated RH variation between dry and wet months as too large over central parts of the domain. This feature spatially coincided with the most pronounced negative P–T correlations (Figure 5). Even larger differences in RH between dry and wet months (in terms of magnitude and spatial extent; Figure 9) were simulated by CLM-CCLM and this feature was associated with too strong negative P–T correlations across Europe (Figure 5). By contrast, IPSL-WRF yielded very small variations in RH (and SR) between dry and wet months over SC, EE, and MD regions (Figure 9), which was related to underestimated strength of the negative P–T correlations in these regions. Similar differences/links were not found in those RCMs with SR and RH properties close to observed data.
6 DISCUSSION
6.1 Possible mechanisms behind the RCMs biases
In the ensemble of seven CORDEX RCMs driven by the ERA-Interim reanalysis, we demonstrated a tendency to underestimate the strength of summertime P–T correlations over northern parts of Europe and the Iberian Peninsula, while too strong negative correlations were found over western and central Europe compared to observed data. The too weak P–T correlations in northern Europe were linked to simulation of SR variations between dry and wet months. Shortwave radiation characteristics suffer from relatively large biases in CORDEX RCMs (Bartók et al., 2017) and largely depend on convection and radiation schemes used (Katragkou et al., 2015). In addition, the radiation biases in northern Europe were probably also linked to representation of blocking anticyclones in the RCMs. Jury et al. (2019) reported fewer blocks in the CORDEX RCMs (despite driven by the ERA-Interim reanalysis) compared to their driving data, suggesting less frequent anticyclonic weather that contributes to higher values of P–T correlations in summer. These RCMs drawbacks probably resulted in the largest intermodel spread in P–T correlations in Scandinavia compared to other European regions.
The overly strong negative P–T correlations in western and central Europe in summer were indirectly linked (via differences in relative humidity between dry and wet months) to a depletion of soil moisture in some RCMs. Knist et al. (2017) showed that CORDEX RCMs underestimate the strength of positive correlations between air temperature and latent heat flux. Thus, higher temperatures are accompanied by lower latent heat flux compared to observations, suggesting a stronger dry-hot relationship in RCMs. In addition, Vautard et al. (2013) reported that several reanalysis-driven CORDEX RCMs share a summertime dry bias associated with too high temperatures, especially the CLM-CCLM and CNRM-ALADIN RCMs (considering the whole European domain), which is in accordance with our study.
Despite the relatively fine resolution of the examined CORDEX RCMs (~12.5 km in their higher-resolution versions), atmospheric convection has to be parameterized. This subgrid process is regarded as a major source of model errors (Stegehuis et al., 2015; Prein et al., 2016) demonstrated that the choice of a convection scheme has a substantial impact on heat waves in the WRF RCMs. For example, the CNRM-ALADIN RCM (its previous version from the ENSEMBLES project; Lorenz and Jacob, 2010) simulated an erroneously large share (roughly 90%) of convective precipitation in summer over central Europe (among 12 ENSEMBLES RCMs), while the observed value was approximately 50% (Kyselý et al., 2016). This error may propagate into the newer version of the model (used in our study) and may be linked to the largest errors in the P–T correlation over western and central Europe in this RCM. The amount of convective precipitation is linked to temperature (Berg et al., 2013), and characteristics of temperature–precipitation compound events are also related to atmospheric circulation (Rulfová et al., 2021). Thus, biases in the amount and/or share of convective precipitation may also be linked to atmospheric circulation.
Although we showed possible links between simulated P–T relationships, shortwave radiation and humidity, the overall bias pattern in summertime P–T correlations, that is, too low r in the north and the Iberian Peninsula and overly high r mostly over western and central Europe was present in the majority of the RCMs. For example, IPSL-WRF and SMHI-RCA yielded similar P–T correlations in summer, despite the fact that SR and RH characteristics differed in those RCMs. This suggests a possible role of other contributors, such as atmospheric circulation. Simulated P–T relationships may also depend on summer precipitation amounts and their variability, which may substantially vary among RCMs (Kotlarski et al., 2014). Because we characterized precipitation by the dimensionless SPI index (Guttman, 1998) that was fitted separately into each RCM and observed data, this cannot be analysed within the present study.
6.2 Role of models' spatial resolution, driving data, and temporal preconditioning
Only minor differences were found between RCMs' simulations with higher (~12.5 km) and lower (~50 km) resolution. Added value of the higher resolution originates mainly from better representation of orographic features and is manifested primarily in (sub)daily characteristics of precipitation and their extremes especially over mountainous regions (Prein et al., 2016). In terms of extreme temperatures, Vautard et al. (2013) found a slight positive effect of higher resolution of CORDEX RCMs on heat waves' persistency. Our study, however, used variables focused on a considerably longer (monthly) time scale and large-scale spatial patterns over Europe. The results obtained from the present study are comparable to those of Kotlarski et al. (2014) who evaluated seasonal temperature and precipitation characteristics in CORDEX RCMs and found no clear benefit of an increased spatial resolution from 50 to 12.5 km.
In our study, the RCMs were driven by the ERA-Interim reanalysis only. Crhová and Holtanová (2018) analysed the P–T relationship in two RCMs driven by various global climate models (GCMs) and concluded that the choice of driving data did not substantially affect the simulated P–T correlations in summer. The effect of driving data would be more pronounced in winter, however, during which precipitation and temperature are more strongly linked to atmospheric circulation (Cattiaux et al., 2012). Atmospheric circulation in RCMs is largely preconditioned by the driving GCM (e.g., Li et al., 2021), and therefore the choice of a GCM may affect the resulting P–T correlations in winter. In summer, by contrast, precipitation and temperature characteristics are strongly related to interactions between land surface and atmosphere (e.g., Davin et al., 2016), that is, processes that are simulated (or parameterized) inside an RCM's domain. Summertime temperature is linked also to precipitation in previous months. Della-Marta et al. (2007) and Russo et al. (2019) showed that hot summers in the Mediterranean tend to be preconditioned by dry springs and, in addition, dry springs over the Mediterranean also contribute to the development of extreme heat in continental Europe (Zampieri et al., 2009). Bastos et al. (2021) concluded that amplified evapotranspiration during the hot spring of 2018 in Europe was related to dry and hot conditions in summer that year. These lagged relationships and their simulation in the RCMs are beyond the scope of the present study but deserve further investigation.
7 SUMMARY AND CONCLUSIONS
- Distinct seasonal patterns of the P–T relationships were found in observed data. Positive (negative) P–T correlations were statistically significant mainly in winter (summer) over large European areas. During spring and autumn, negative P–T correlations were located predominantly in southern Europe, while positive correlations were found over northern parts of the domain.
- The biases in the P–T relationships were largest in summer. Most RCMs yielded too weak negative correlations over northern Europe and the Iberian Peninsula while the strength of the negative correlations was overestimated in western, central, and southeastern Europe, suggesting more pronounced dry–hot/wet–cold regime here compared to observed data. As this is area where heat waves and droughts often develop and have been increasing recently, their simulation as compound events in RCMs may suffer from relatively large biases. This might hold also for their climate change scenarios.
- Erroneously simulated characteristics of shortwave radiation and relative humidity contributed to the P–T relationship biases in summer but the spatial pattern of the bias (north–positive, south–negative) was present in most RCMs. The only RCM in which this pattern was not found (CLM-CCLM) substantially overestimated differences in shortwave radiation and relative humidity between dry and wet months, which were manifested in too strong P–T relationships across Europe.
- In winter, the P–T relationships were too strong over northern Europe while too weak in the south in most RCMs compared to observed data. The RCMs' biases in the transition seasons were closer to the winter than summer pattern, and smallest in autumn.
- The RCMs tended to share their strengths and weaknesses in both higher- (EUR-11) and lower- (EUR-44) resolution versions.
The results of the study revealed rather general RCMs' biases in the P–T relationships over some European regions. These may considerably affect the simulation of precipitation- and temperature-related extremes, such as droughts, heat waves, or rain-on-snow events. We also found potential links between the P–T correlation biases and shortwave radiation (related to cloud cover) and relative humidity. These findings may be useful for further improvements of climate models and, ultimately, for obtaining more credible climate change scenarios.
ACKNOWLEDGEMENTS
This study was supported by the Czech Science Foundation, project 20-28560S. OL contribution was carried out also within the INTER-COST project funded by the Ministry of Education, Youth and Sports of the Czech Republic (project no. LTC19044) and project SustES – Adaptation strategies for sustainable ecosystem services and food security under adverse environmental conditions (CZ.02.1.01/0.0/0.0/16_019/0000797). We acknowledge the World Climate Research Programme's Working Group on Regional Climate and the Working Group on Coupled Modelling, former coordinating body of CORDEX and responsible panel for CMIP5 and the E-OBS dataset from the EU-FP6 project UERRA (http://www.uerra.eu) and the Copernicus Climate Change Service, and the data providers in the ECA&D project (https://www.ecad.eu).
AUTHOR CONTRIBUTIONS
Jan Kyselý: Conceptualization; funding acquisition; resources; supervision; validation. Ondřej Lhotka: Data curation; formal analysis; investigation; methodology; resources; software; visualization; writing – original draft.
