Regional Climate Simulation Ensembles within CORDEX-EA Framework over the Loess Plateau: Evaluation and Future Projections

Abstract

As a semi-arid to semi-humid transitional zone, the Loess Plateau is sensitive to climate change due to its fragile ecological environment and geographic features. This study assesses the performance of six historical experiments from the Coordinated Regional Climate Downscaling Experiment (CORDEX) in this region during 1980–2005. In addition, projected future changes in surface air temperature and precipitation are investigated under the representative concentration pathways (RCP) 2.6 and 8.5 during three periods in the 21st century: the early future (2011–2040), middle future (2041–2070), and late future (2071–2099). Results show that experiments reasonably reproduce the spatial pattern of 2m temperature and precipitation for all seasons, yet with a slight warm bias and prominent wet bias. In the future, the area-averaged magnitude of change will be 1.1 °C, 1.4 °C, and 1.4 °C under RCP2.6 and 1.3 °C, 2.7 °C, and 4.5 °C under RCP8.5 for the early, middle, and late periods, respectively. The warming effect is greater in elevated areas. Precipitation change in future periods is more complex, with both increasing and decreasing trends, depending on the season, location, and scenario. The results are expected to provide regional climate information for decision makers and benefit applications such as agriculture, ecological environment protection, and water resource management.

1. Introduction

Climate change information is required for an impact on the social economy, vulnerability assessments, and the implementation of mitigation and adaptation measures. Climate modeling is the primary tool employed by the IPCC (Intergovernmental Panel on Climate Change) to assess past, present, and future climate conditions. However, due to its coarse horizontal resolution (approximately 100–200 km), general circulation models (GCMs) are limited in representing orography or sub-grid processes at regional or local scales [1,2].

Regional climate models (RCMs) are widely used to dynamically downscale the synoptic fields from reanalysis/GCMs. They can realistically represent local climates with refined atmospheric and land surface processes at spatial scales beyond current general circulation model capabilities [1,3,4,5,6]. To better understand relevant regional/local climate phenomena and their variability and changes, the World Climate Research Program (WCRP) Coordinated Regional Climate Downscaling Experiment (CORDEX) was established. The project aims to design and conduct high-resolution experiments over prescribed spatial domains across the globe [5], such as Africa [7,8], Europe [9,10], the Mediterranean [11], North America [12], East Asia [13], and the Arctic [14]. The first and second plan of the East Asian branch within the CORDEX framework (CORDEX-EA) provide regional climate projections for East Asia with the horizontal resolution at 0.44° and 0.22°, respectively [13,15,16].

Due to its fragile ecological environment and geographic features, the Loess Plateau is sensitive to climate change as a semi-arid to semi-humid transitional zone [17,18]. Future climate change may have greater effects on soil erosion, restored vegetation, limited water resources, and crop production [19,20,21]. Therefore, climate projection information over this region is essential for climate adaptation, environmental protection, and ecological development.

Numerous previous studies have investigated the historical and projected future changes of temperature and precipitation over large areas such as East Asia [13,22,23,24,25,26,27,28,29,30] and China [31,32,33,34]. These studies utilized coarse resolution datasets (i.e., 100–200 km from GCM simulations) or early-phase CORDEX-EA results (0.44° horizontal grid spacing). However, climate information produced at such large spatial extents and relatively low resolution is not adequate for exploring the characteristics of a specific region such as the Loess Plateau. In addition, it is generally accepted that the model uncertainties may be reduced and the model credibility can be improved by employing multi-model ensemble [9,10,35,36], while many previous studies are based on single model. Therefore, it is necessary to investigate the impacts for a specific region such as the Loess Plateau to develop adaptation and mitigation strategies utilizing higher resolution simulation ensemble.

In this study, the performance of 0.22° historical simulations within the framework of CORDEX-EA is evaluated relative to the 0.1° China Meteorological Forcing Dataset (CMFD) over the Loess Plateau in terms of surface air temperature and precipitation. In addition, future projections of temperature and precipitation change in the study area under two greenhouse gas concentration emission scenarios (RCP2.6 and RCP8.5) are analyzed. The results are expected to benefit applications such as agriculture, ecological environment protection, and water resource management.

2. Data and Methods

2.1. Study Area

The study area is the Loess Plateau, which has geographical boundaries of 33°41′ N~41°16′ N and 100°52′ E~114°33′ E. The elevation data from ETOPO1 are used to present the topography of the domain (Figure 1).

The Loess Plateau, located in central northern China among the middle reach of the Yellow River, occupies an area of over 640,000 km² land surface. It crosses arid, semi-arid, and semi-humid climate zones, with an annual average temperature range of 4.3 °C–14.3 °C and an average annual precipitation range of 200–750 mm from northwest to southeast.

2.2. RCP Scenarios

To facilitate future assessments of climate change, representative concentration pathways (RCPs) were developed for the climate modeling community as a basis for long-term and near-term modeling experiments. In contrast to the IPCC’s Special Report on Emissions (SRES) scenarios, the radiative forcing trajectories of RCPs are not associated with predefined storylines and can reflect various possible combinations of economic, technological, demographic, and policy developments [37].

Two future scenarios, leading to a very low forcing level (RCP2.6) and a very high forcing level (RCP8.5), are utilized in the present study. The RCP2.6 scenario is designed to meet the 2 °C global average warming target compared to pre-industrial conditions [38], which peak in the radiative forcing at approximately 3 W/m² before 2100 and then decline to 2.6 W/m² by the end of the 21st century. RCP8.5 assumes a high rate of radiative forcing increase, peaking at 8.5 W/m² in year 2100 [39].

2.3. CORDEX-EA

CORDEX-EA simulations at 0.22° (approximately 25 km) resolution are utilized in the present study, which are available at the Earth System Grid Federation (ESGF) nodes for CORDEX (https://cordex.org/data-access/cordex-data-on-esgf/ (accessed on 12 March 2023)). The monthly aggregated variables daily average 2m temperature (tas) and daily precipitation (pr) were selected for historical and future scenario experiments, and only realization r1p1i1 is utilized in the present analysis.

Six combinations of driving GCMs and RCMs were identified, whose detailed information is summarized in

Table 1. List of RCMs and their driving models used in the present study.

Ensemble Member	Institute	GCM	RCM	Experiments
M01	GERICS	NCC-NorESM1-M	REMO2015	historical, rcp2.6, rcp8.5
M02	GERICS	MOHC-HadGEM2-ES	REMO2015	historical, rcp2.6, rcp8.5
M03	GERICS	MPI-M-MPI-ESM-LR	REMO2015	historical, rcp2.6, rcp8.5
M04	ICTP	NCC-NorESM1-M	RegCM4-7	historical, rcp2.6, rcp8.5
M05	ICTP	MOHC-HadGEM2-ES	RegCM4-7	historical, rcp2.6, rcp8.5
M06	ICTP	MPI-M-MPI-ESM-MR	RegCM4-7	historical, rcp2.6, rcp8.5

. The regional simulations were driven by three GCMs, namely the Max Planck Institute (MPI) Earth System Model (MPI-ESM-LR), the Norwegian Climate Centre (NCC) Earth System Model Version 1 (NorESM1-M), and the Meteorological Office Hadley Centre (MOHC) Global Environmental Model Version 2 with Earth System Configuration (HadGEM2-ES). The GCM NorESM is characterized by low equilibrium climate sensitivities (ECS), the GCM MPI-ESM is characterized by a medium global ECS, and HADGEM2ES is characterized by a comparatively high ECS [40]. Two RCMs were utilized for dynamical downscaling driven by each GCM, including REMO2015 and RegCM4.7. REMO2015 is the latest hydrostatic version of the REMO regional climate model developed by the German Climate Services Center (GERICS) [41]. RegCM4.7 is the latest version of the RegCM regional climate model and is developed and maintained by the Abdus Salam International Center for Theoretical Physics (ICTP) in Trieste, Italy [42].

The multi-model ensemble (MME) approach, which allows the improvement of the quality and credibility of climate change information, has become a common practice in climate research [10]. There are various ensemble approaches developed for regional climate applications; for example, the simple ensemble approach [43], weighting average approach [44,45], and Bayesian and Power method [46,47] are widely used. In this study, the ensemble mean (ENS) is calculated as the arithmetic average of the outputs from six downscaling experiments.

2.4. CMFD Gridded Dataset

To evaluate the performance of CORDEX-EA historical experiments, the China Meteorological Forcing Dataset (CMFD) provided by the National Tibetan Plateau Data Center (http://data.tpdc.ac.cn) is used as a reference observation. As a newly compiled gridded daily meteorological dataset, it is developed at a 0.1° spatial resolution covering the period from 1979 to 2018 [48]. It was compiled based on ground measurements obtained from the China Meteorological Administration and other datasets, such as satellite precipitation data and Global Land Data Assimilation System data. The high-resolution elevation data were introduced in the observed air temperature interpolation.

The gridded observation datasets described above were re-projected and aggregated to 0.22° rotated lat/lon projection to be consistent with CORDEX RCM model outputs. The variables daily maximum, minimum, and average air temperature and daily precipitation are investigated, which are further aggregated into monthly and seasonal averaged values. The entire year is separated into four different seasons: winter (December–February DJF), spring (March–May MAM), summer (June–August JJA), and fall (September–November SON).

2.5. Evaluation Metrics

The performances of climate models need to be evaluated before future projections, which are important to assess model reliability. Three metrics are calculated, namely bias (BIAS), root mean squared error (RMSE), and Pearson’s correlation coefficient (R), in order to evaluate the performance of regional climate simulations against gridded observations. These metrics are listed as follows:

$$\text{BIAS}=\frac{{\displaystyle {\sum }_{\mathrm{i}=1}^{\mathrm{n}}({\mathrm{y}}_{\mathrm{i}}-{\mathrm{x}}_{\mathrm{i}})}}{\mathrm{n}}$$

(1)

$$\text{RMSE}=\sqrt{\frac{{\displaystyle {\sum }_{\mathrm{i}=1}^{\mathrm{n}}{({\mathrm{y}}_{\mathrm{i}}-{\mathrm{x}}_{\mathrm{i}})}^2}}{\mathrm{n}}}$$

(2)

$$\mathrm{R}=\frac{{\displaystyle \sum _{\mathrm{i}=1}^{\mathrm{n}}({\mathrm{x}}_{\mathrm{i}}-\overline{\mathrm{x}})({\mathrm{y}}_{\mathrm{i}}-\overline{\mathrm{y}})}}{\sqrt{{\displaystyle \sum _{\mathrm{i}=1}^{\mathrm{n}}{({\mathrm{x}}_{\mathrm{i}}-\overline{\mathrm{x}})}^2}}\sqrt{{\displaystyle \sum _{\mathrm{i}=1}^{\mathrm{n}}{({\mathrm{y}}_{\mathrm{i}}-\overline{\mathrm{y}})}^2}}}$$

(3)

where ${\mathrm{y}}_{\mathrm{i}}$ is the modeled value; ${\mathrm{x}}_{\mathrm{i}}$ is the observation-derived value; $\mathrm{n}$ is the number of time steps. $\overline{\mathrm{x}}$ and $\overline{\mathrm{y}}$ are the time means.

The evaluation process is conducted for 26 years from 1980 to 2015, the period when the two datasets overlapped. The model quality assessment focuses on the capacity of the RCM historical runs to correctly represent the spatial pattern and temporal distribution of 2m temperature and precipitation distributions. Therefore, the metrics are calculated for each grid box and accumulation period on a monthly scale for the current study against the corresponding regular gridded datasets. The significance of the projected change is evaluated using a two-tailed Student’s t test.

3. Results

3.1. Evaluation of Historical Experiments

The quality assessment of RCM historical simulations is a key step to understand the reliability of model information for characterizing the present climate and to estimate the climate change signal on the temperature and rainfall properties over a region.

3.1.1. Temperature

Figure 2 provides an overview of the spatial distribution of the 26-year mean seasonal values and biases of the CORDEX-EA 0.22° ensemble for average 2m temperature. The observed surface air temperature shows a clear seasonal cycle, with the highest values in JJA (20 °C to 28 °C) and the lowest in DJF (−10 °C to 6 °C). In general, the RCM simulations reproduce well the spatial variability of 2m temperature for each season, including the south-north gradient and elevation effects (lower values in the western high-elevation area, and relatively higher values in the Guanzhong plain located in the southern part). Nevertheless, biases can be discerned by comparing the ensemble mean with the gridded observation. In DJF, warm biases can be identified in most areas, particularly the northern border of SX province along with NM province, the western part of S’X province, and the northwestern part of NX province. In JJA, 2m temperature is underestimated over almost the whole plateau.

As summarized in Table S1, the spatial correlations of annual and seasonal mean 2m temperature for all experiments are generally above 0.90, which indicates that RCMs reproduce well the observed spatial variation. The RMSE of annual and seasonal mean 2m temperature for all experiments ranges from 0.09–0.27, which is relatively small. Generally, the REMO series (M01~M03) simulates a higher temperature than the RegCM series (M04~M06), which is particularly prominent in DJF (Figure S1). The ensemble mean (ENS) presents a slight negative bias as an annual mean, which is closer to the observation than the individual experiments (0.3 °C, 0.4 °C, and 0.6 °C for M01~M03, and −0.4 °C, −0.7 °C, and −1.2 °C for M04~M06). The ENS has the lowest RMSE as an annual mean compared with the single model simulation.

The annual cycle of the 26-year averaged CORDEX-EAS 0.22° surface air temperature is calculated for the ensemble mean and for each experiment, as shown in Figure 3. Overall, RCM simulations reasonably capture the monthly variation of observed air temperature. The observation value is within the range of model spread from March to November, while it is systematically lower than all model simulations from December to February. It yields a warm bias around 0.9 °C for the ensemble mean in DJF.

Statistics of area-averaged 2m temperature in each season are summarized using the Taylor diagram [49] shown in Figure 4, and detailed statistical results are included in the Supplementary Material as Table S1. The spatial correlation coefficients are generally above 0.90 for almost all experiments in all seasons, with the highest value in JJA. The normalized standardized deviations (NSD) are around 1, within a range of 0.5–1.5 for each experiment, with the least spread in DJF, which indicates that models can reasonably simulate the spatial variability of surface air temperature, although the simulated spatial variations of air temperature are smaller or greater than those of the observations. Generally, the NSDs of the REMO series M01~M03 are larger than the RegCM series M04~M06. RMSDs are within the range of 0.50 °C to 0.75 °C for all experiments in all seasons. Considering the three statistical metrics for the four seasons, M02 seems to outperform the others. The ENS presents a high spatial correlation coefficient (larger than most models), indicating that it better captures the spatial distribution than most individual experiments. Moreover, the ENS generates an NSD value between the REMO series M01~M03 (more than one) and the RegCM series M04~M06 (less than one), closer to one.

3.1.2. Precipitation

Figure 5 shows the spatial distribution of annual and seasonal mean precipitation from the model simulation ensemble mean and gridded observation. Generally, RCMs reasonably reproduce the spatial pattern of precipitation. The gradient from southeast to northwest can be identified for four seasons, particularly in summer. In all seasons, precipitation is overestimated over almost the whole LP, except for a small part along the southern edge. In summer, the wet bias is the largest, with a magnitude exceeding 2.0 mm/d in most part of SX, the southern part of NX, the eastern part of GS, and a part of NM along the border with SX.

As shown in Table S2, the spatial correlations of annual mean precipitation for all experiments are generally within the range of 0.76 to 0.94, which indicates that RCMs are capable of reasonably capturing the observed spatial variation. Nevertheless, the spatial correlation coefficient is the lowest in DJF, which can be 0.2 for M01. The RMSE of annual and seasonal mean precipitation for all experiments ranges from 0.0 mm/d to 0.3 mm/d. Generally, the REMO series (M01~M03) simulates lower precipitation than the RegCM series (M04~M06), which is particularly prominent in JJA (Figure S2). All the experiments exhibit a positive bias for precipitation. Compared with single experiments, the annual average of the ENS exhibits the highest correlation coefficient of 0.94, an RMSE of 0.10 mm/d, and a positive bias of 0.8 mm/d, which is in-between the bias range of the two series.

Figure 6 compares the annual cycles of area-averaged precipitation simulated by RCMs and their ensemble means with gridded observation. The accuracy of precipitation is not simulated as well as temperature. Although the timing of rainfall rise (May–July), peak-reaching (July), and fall (July–October) is correctly captured by most experiments, there is still prominent bias and large inter-model variance. Generally, RCMs overestimate precipitation amounts in most periods during the year. In JJA, the M01 and M04 simulations are relatively closer to the observation, while the other simulations are much higher. In DJF, observed precipitation is very low, while RCM simulations are generally above 0.2 mm/d. The ensemble mean generally follows the annual cycle of observation, with a greater amount of magnitude (over 2 mm/d in July and August).

Taylor diagrams are constructed to synthetically evaluate the performance of simulated seasonal mean precipitation (Figure 7). Overall, the spatial correlation is around 0.7 for JJA, between 0.5–0.8 for SON and DJF, and within the range of 0.65–0.85 in MAM. It is worth noting that the normalized standardized deviations (SDs) are generally larger than one for all seasons (with M02 in SON as an exception), and prominently larger than three for DJF and MAM. In DJF, the normalized SD of M02 is around eight, while the correspondence of M01 and M03 are far beyond eight and therefore out of the maximum axis limit (Figure 7a). This is due to the overestimation of model simulations compared with the very small value of observed precipitation in winter (Figure 6). As summarized in Table S1 and shown in Figure S2, the ensemble mean generally has a spatial correlation above 0.78 for all seasons except DJF (0.47). Overall, M01 and M03 have a smaller spatial correlation than other experiments for all seasons, indicating that they are less accurate in capturing the spatial variation of precipitation.

3.2. Projections of Future Changes

Projected changes under the low-emission scenario RCP2.6 and the high-emission scenario RCP8.5 during three future periods in the 21st century (early period (2011–2040), middle period (2041–2070), and late period (2071–2099)) are calculated, taking the historical period (1980–2005) as a reference.

3.2.1. Temperature

Figure 8 illustrates the projected change of 2m temperature over the whole domain in three future periods under the RCP2.6 scenario. The results point to a general increase of the annual mean temperature, with an area-averaged magnitude of 1.1 °C, 1.4 °C, and 1.4 °C for the early, middle, and late periods, respectively. The augment magnitude of JJA, SON, and DJF are slightly greater than MAM for all three periods. Moreover, the temperature rise magnitude is not evenly distributed over the study area, which is more prominent in elevated areas, particularly in the west corner located in QH province. The significance of the climate change signals is present across all domains for the annual mean, while it is present in the southwest part for the seasonal mean during all three future time periods.

For each future period, the magnitude of seasonal averaged 2m temperature change and its spatial distribution are different for individual experiments (Figure S3). Overall, for the annual and all seasonal mean 2m temperature during all three periods, M02 projects the largest increase, while M04 exhibits the smallest increase (Table S3). For a single season during the three periods, the largest increase occurs during DJF and SON in the middle period, as simulated by M02, and JJA and SON in the middle and late periods, as simulated by M05.

Under the more severe scenario (RCP8.5), the temperature rise magnitude is generally larger than under RCP2.6, with area-averaged annual mean values of 1.3 °C, 2.7 °C, and 4.5 °C for the early, middle, and late periods, respectively (Figure 9). Similarly to RCP2.6, the temperature rise is more prominent in elevated areas. The augment magnitudes of JJA and SON are slightly greater than DJF and MAM. The significance of the climate change signals is present across all domains for the annual mean during all three future time periods. It is present in the southwest part for DJF and MAM, across almost all domains in JJA, and across all domains in SON during all three future time periods.

The simulated 2m temperature change by individual experiments under RCP8.5 generally exhibits a higher increase than RCP2.6, but with a different magnitude and spatial distribution (Figure S4). Overall, for annual and all seasonal mean 2m temperature during all three periods, M02 and M05 projected greater increases in magnitude than other experiments (Table S3). For a single season during the three periods, the largest increase (over 5 °C) occurs during all four seasons in the late period, as simulated by M02, and JJA and SON in the late period, as simulated by M05.

Figure 10 provides an overview of the magnitude of monthly 2m temperature changes averaged over the whole domain during the three future periods for both RCPs. Overall, for each of the three periods, the temperature augmentation magnitude of RCP8.5 is larger than the corresponding one of RCP2.6. Under RCP2.6, in most months (except June and July), the magnitude rises from the early to the middle future, reaches its maximum, and then falls in the late future. The case is different for RCP8.5, in which the magnitude augments monotonously from the early to middle future until the late future.

3.2.2. Precipitation

Figure 11 shows the projected precipitation change under the RCP2.6 scenario. The annual mean area-averaged precipitation increases at the magnitudes of 4.2%, 6.3%, and 2.1% for the early, middle, and late periods, respectively. However, the change magnitude is not evenly distributed seasonally and spatially, with both positive and negative directions. In general, a rainfall decrease can be found in part of the domain in summer and autumn in the middle and late periods, while an increase prevails in the early period and for most areas during winter and spring in the middle and late periods. The significance of the climate change signals is present across all domains for the annual mean during all three future time periods. It is present over the central-east part for DJF, in part of the east corner in MAM, in a northern slice in JJA, and in the south-east part for SON during all three future time periods.

The ensemble means of the model simulations exhibit relatively small fluctuations compared to the individual simulations (Table S4). The simulated precipitation change from the individual experiments shows a quite different magnitude and spatial distribution (Figure S5). Overall, M05 and M06 projected the largest increase in annual mean precipitation for all three periods, while M01-M04 exhibited a decrease or smaller increase in magnitude.

Under the RCP8.5 scenario, the annual mean area-averaged precipitation increases at the magnitudes of 0.8%, 6.0%, and 9.5% for the early, middle, and late periods, respectively (Figure 12). In general, a rainfall decrease can be found in part of the domain in summer and autumn, while an increase prevails over almost the entire plateau in winter and spring. The significance of the climate change signals is present across almost all domains except in the southern corner for the annual mean during all three future time periods. It is present over a small part of the domain during all three future time periods.

Simulated precipitation changes from individual experiments show quite different magnitude and spatial distributions (Figure S6). Similar to RCP2.6, the ensemble means of the model simulations exhibit relatively small fluctuations compared to the individual simulations under RCP8.5 (Table S4). Overall, M05 projected the largest increase in annual mean precipitation for all three periods, while other experiment exhibited smaller increases in magnitude or even decreases.

Figure 13 provides an overview of monthly area-averaged precipitation changes of the ensemble means in the three future periods under both RCPs. Overall, the wetting trend dominates during the year for the three future periods and both RCPs. The magnitude is much larger in winter than in summer. In JJA, the wetting or drying trend is not prominent, which can be negative in the middle and future periods under RCP8.5.

4. Discussion

The purpose of the current study is neither to evaluate the added value of high-resolution RCM simulations compared to their diving GCMs, nor to evaluate the benefit of the simulations in 0.22° grid spacing compared to 0.44°. Instead, we attempt to evaluate the performance of regional simulations in the historical period for reliable projections for future periods.

Compared to many previous studies, the current paper focuses on the Loess Plateau instead of the entirety of China [32,50,51] or East Asia [16,30,52]. Another major difference is the usage of the highly resolved observation dataset in this study. It is worth noting that the recently available observation dataset CMFD (0.1° × 0.1°) was chosen as a reference in our study. It has a higher spatial resolution compared to the two other widely used datasets: viz. CN05.1 (0.25° × 0.25°) [53], which covers China, and APHRODITE (0.5° × 0.5°) [54], which covers East Asia.

In the present study, 2m temperatures generated by CORDEX-EA experiments generally show a cold bias in summer and a warm bias in winter over the Loess Plateau, which could be partly attributed to the RCM’s atmospheric forcing. Previous studies show that the NCC and MOHC GCMs tend to produce cold biases [55] and a negative surface albedo-caused bias related to overestimated snow cover in high elevations [56,57]. The wet bias simulated by the regional model REMO in almost all seasons on the Loess Plateau can be partially ascribed to the strong wet bias in the GCMs [55].

The projected magnitudes of future changes in terms of surface air temperature and precipitation over the Loess Plateau are comparable with the results from several existing studies at regional or country levels. Yu et al. [50] analyzed the future climate change over China using CMIP5, CMIP6, and CORDEX datasets. Their results project an area-averaged temperature rise of 1.33 °C, 1.27 °C under RCP2.6 and 2.31 °C, 4.38 °C under RCP8.5 for the mid-future and far-future; area-averaged precipitation changes of 3.83% and 1.01% under RCP2.6; and 3.97%, 7.02% under RCP8.5 for the mid-future and far-future over the Loess Plateau sub-region. These findings concur with our results, although the precipitation change is slightly lower. Another study investigated climate change in the 21st century over China under RCP4.5 and RCP8.5 using RegCM4 ensembles [51]. Their results show an increase of area-averaged precipitation of 3/6/15% for the annual mean, 13/23/49% in winter, and 2/4/7% in summer over the Loess Plateau sub-region under RCP8.5 for the near-/mid-/far-future. They also show a rise of 1.2/2.1/4.3 °C for the daily maximum temperature and a rise of 1.3/2.6/5.1% for the daily minimum temperature under RCP8.5 during the near-/mid-/far-future. The precipitation change magnitudes of the annual mean and in winter from our study are comparable with their results, while our simulated change of precipitation in winter is negative in summer, compared to a slight positive change from their results. Our simulated 2m average temperature change magnitude is generally comparable with the maximum and minimum temperature rise ranges indicated by their results. A similar study using REMO driven by three GCMs was conducted for three future periods under RCP2.6 and RCP8.5 over the upper and middle reaches of the Yellow River basin [55]. This area overlaps in large part with our study domain of the Loess Plateau. Their results indicate that the increases in mean temperature are strongest for the far-term in winter, with an average increase of 5.6 °C under RCP 8.5, and show a substantial increase in precipitation (34%) occurring in winter under RCP 8.5 for the far-term. This is also broadly consistent with our results. Nevertheless, it is noteworthy that the comparisons should be considered with caution as these simulations use different driving GCMs, different RCM ensembles, and different future periods. Moreover, the domains analyzed to generate area-averaged values are not exactly the same.

There are limitations in the present study. Firstly, only the mean climate state has been taken into account, while some studies argue that the impacts of extreme events are more obvious and direct than climate averages [58]. To avoid severe risks resulting from extreme climate events and to assess their impacts on the economy and natural ecosystems, it is important to quantify their magnitude and frequency change. In future work, extreme climate indices will be calculated from daily temperatures and precipitation for historical and future periods.

In addition, the bias-correction method is not applied to the outputs of RCMs, which is beyond the scope of the current study. In this study, the common assumption which accepts that biases do not depend on the climate state is followed. However, recent studies showed that biases depend on the model state and need to be extrapolated from the present to the future. In a future study, we will attempt to conduct bias-correction and take the temperature-dependent bias into consideration.

Moreover, in the current study, a six-member ensemble was employed made by two regional climate models downscaling three global climate models. These are all the experiments within the CORDEX-CORE framework which are dispatched from the ESGF node at the moment [40]. With the development of modeling and availability of computation resources, it is foreseen that more high-resolution dynamical downscaling datasets will be generated and accessible. Therefore, a greater number of experiments will be employed in the future study.

5. Conclusions

In this study, the performances of six GCM-driven RCM ensembles of the CORDEX-EA are evaluated compared to fine-resolution CMFD gridded observations over the Loess Plateau during 1980–2005. In addition, projected future changes in surface air temperature and precipitation are investigated under two scenarios (RCP2.6, RCP8.5) during three periods in the 21st century: the early future (2011–2040), middle future (2041–2070), and late future (2071–2099). The main conclusions include the following:

(1)

CORDEX-EA experiments reproduce well the spatial distribution of 2m temperature for each season. Warm biases can be identified in most areas in DJF, while cold biases exist over almost the whole plateau in JJA. In addition, these experiments generate a good reproduction of observed monthly variation. The observation is within the range of model spread from March to November, while it is lower than all model simulations from December to February, which yield a warm bias of 0.93 °C for the ensemble mean in DJF.

(2)

RCMs generally reasonably reproduce the spatial pattern of precipitation. The gradient from southeast to northwest can be identified for the four seasons, particularly in summer. In all seasons, precipitation is overestimated over almost the whole plateau except for a small part along the southern edge. Although the timing of rainfall rise (May-July), peak-reaching (July), and fall (July–October) is correctly captured by most experiments, there is still prominent bias and large inter-model variance.

(3)

In the future, for both RCP2.6 and RCP8.5, the temperature rise is unevenly distributed and more prominent in elevated areas. The area-averaged magnitude of change is 1.3 °C, 2.7 °C, and 4.5 °C under RCP8.5 compared with 1.1 °C, 1.4 °C, and 1.4 °C under RCP2.6 for the early, middle, and late periods, respectively. Overall, for each of the three periods, the temperature augmentation magnitude of RCP8.5 is larger than the corresponding one of RCP2.6. Under RCP2.6, in most months (except June and July) the magnitude rises from the early to middle future, reaches its maximum, and then falls in the late future. The case is different for RCP8.5, in which the magnitude augments monotonously from the early to middle future until the late future.

(4)

Annual mean area-averaged precipitation increases at the magnitudes of 4.2%, 6.3%, and 2.1% under RCP2.6 and 0.8%, 6.0%, and 9.5% under RCP8.5 for the early, middle, and late periods, respectively. For both RCPs, rainfall augments for most areas during winter and spring during the three future periods. Rainfall decrease can be found in part of the domain in summer and autumn during the middle and late periods for RCP2.6 and during all three periods for RCP8.5.

The results in the current study may provide policy makers with the regional climate information required for mitigation and adaptation measures. They are expected to benefit applications such as agriculture, ecological environment protection, and water resource management over the Loess Plateau.