The mean 2-m air temperature is the first climatic parameter that is analysed and shown in Fig. 2. Higher observed (CRU) values are found in the southern part of the domain, which decrease northward, a pattern that comprises three zones of different temperature intervals (> 25 °C, 15–25 °C, < 25 °C). All six difference maps from the LSS experiments exhibit biases of > 10 °C over the eastern part of Turkey and Caucasus, western Iran and Morocco, and cold biases over the eastern part of the domain and Sudan. The large warm biases are uniform for all LSS and are due to the lower model elevation at 50 km resolution as they occur over the main mountain ranges of the domain [areas that are also known to be under-sampled by the CRU TS.3 dataset as shown in Fig. 3 of Harris et al. (2014)]. Hence, other smaller, but non-negligible differences in other areas are not easily seen with the applied colour scale. For example, the two NoahMP and the CLM are warmer by 2–3 °C from Noah at the northern African mainland (especially over 20 °E and 20 °N), as shown in Constantinidou et al. (2020).
The Taylor diagram produced for air temperature is presented in Fig. 3, where basic statistics for the different simulations are denoted with different colours for the LSS and symbols for the sub-domains. The results produced by all six schemes and for every region have very small differences from each other. The correlation coefficients for the six experiments are generally high, between 0.7 and 0.95 for all the sub-regions, while for the whole MENA domain it is highest under the WRF/Noah simulation (\(\sim\) 0.89). Regions E and F are closer to observations with a correlation of 0.95, normalized standard deviation (presented with blue arcs in Fig. 3) very close to 1.0 and root mean square error (RMSE) less than 0.5. The experiment using the Noah scheme is closer to observations when looking the results for region E and CLM for domain F.
The results of the more detailed spatial bias analysis in the seven MENA sub-domains is presented in Fig. 4 in the form of matrix-plots. The presented information includes the biases (colour bars) calculated for annual climatology, 5th and 95th percentiles, standard deviation and linear trend of the six simulations ranked (numbers) for each sub-region, their average and the whole MENA domain. From a sub-region inspection of the figure it is evident that box D, which includes Syria/Iraq area exhibits the largest biases for climatology and extremes consistently for all six LSS, while sub-domains A and B underestimate consistently the standard deviation and area G overestimates it. A comparison among the schemes reveals that CLM is warmer for climatology and extremes in most of the sub-regions, while for the other metrics none of the schemes performs in a consistent manner.
Land Surface Temperature
The satellite-based information for land surface temperature is presented in Fig. 5 (top map). Higher values are observed over the southern part of the MENA domain while lower temperatures are noted in the northern (European) part of the region. Over almost the whole study area, WRF simulates with all LSS colder conditions than observed. An exception is noted over the southwestern area of the model domain where all schemes simulate higher values than the observations.
The Taylor diagram in Fig. 6, shows that all schemes have correlation coefficient between 0.4 and 0.7 for several sub-domains, smaller overall than for air temperature. Exceptionally, box F (Saharan desert) has \(\sim\) 0.94. Normalized standard deviations for sub-domain F is close to 1.0 for all runs except the WRF/CLM (\(\sim\) 0.8) and for the rest of the investigated areas less than 0.8. RMSE obtained for all simulations and domains is less than 1.0, with the whole MENA \(\sim\) 0.9 and sub-region F a value of < 0.4.
The quantitative and sub-regional comparison included in Fig. 7 confirms the widespread cold model biases shown in Fig. 5. It can be seen that overall Noah is the best performing regarding annual climatology over the MENA and sub-regions. The 95th and 5th percentiles are mostly underestimated by the different LSS used for the simulations, expect Noah slightly overestimating warmest conditions and CLM overestimating the coldest conditions. The different performance of CLM in the 5th percentile, especially over the vegetated areas (including those with forests), could be due to the shading effects operating in this LSS (two-stream canopy radiation transfer scheme), where canopy and ground surface temperatures are separately computed. CLM also considers canopy gaps and calculates fractions of sunlit and shaded leaves together with the absorbed radiation which may lead to the land surface temperature overestimation noted in box C, where the Land Use index considered by the model includes grassland, shrubland, croplands and mixed forest.
In Fig. 8, it is obvious from the CRU observations that drier conditions (rainfall less than 25 mm/month annual average) prevail over most of the MENA domain (in northern African and Middle East), while the northern part of the domain (Europe, Anatolia, Caucasus) as well as the southern part of the Sahel region and the tropics are wetter. The six simulations strongly overestimate precipitation in the tropics (south of 15°N) but this is a region where different observational datasets tend to vary a lot (Tanarhte et al. 2012). Underestimation of precipitation is simulated by all LSS in large parts of Europe (except the Balkan Peninsula) and around the Mediterranean Sea.
All six simulations exhibit low correlation (0.0–0.4) with observations as presented in Fig. 9. Standard deviations of more than 1 are obtained from all different runs and regions, reaching values more than 4 for the simulations with both options of NoahMP and CLM over sub-domain F, which is also the case when focusing on the results for RSME. These large values (in contrast to air and land temperature which are not greater than 1) indicate, not surprisingly, the high month-to-month variability of precipitation.
Box C stands out in the detailed bias map of Fig. 10 as the area with the largest biases in most of the metrics and for all LSS. This figure also reveals that, while the Noah scheme achieves the least bias for several metrics (e.g. annual climatology and 95th percentile) in the whole MENA domain, for specific sub-regions (e.g. box G) its bias is the largest among the schemes (and of opposite sign)). The scheme suitability can be assessed for individual areas, for example, in sub-domain D (Levant and Mesopotamia), Fig. 10 demonstrates that the best performing LSS in simulating precipitation is RUC.
The annual climatology of net radiation observed by CERES is shown by the top map of Fig. 11. Most parts of the MENA domain measure net radiation < 100 W/m\(^2\), except from the coastal areas of northern Africa and the Arabian peninsula where it is greater than 100 W/m\(^2\). Looking at the comparison of the simulations with the observations, it is evident that all six WRF model options of LSS underestimate net radiation over the areas where observed values are \(>100\) W/m\(^2\) (Fig. 11). This distinct difference pattern is also recorded in the respective upward short-wave map for winter (not shown).
Figure 12 summarizes the statistical outcome of the six experiments and visualised in the form of Taylor diagram for net radiation compared to CERES satellite observations. The correlation of the different runs lies in the range of 0.4–0.7 and for the whole domain of interest is about 0.6. Root mean square error for all the options studied here takes values close to 1.0, while the standard deviation varies from 0.8 to 1.4. The compactness of these results suggest that different LSS in the six runs do not have a discernible effect, overall, on the model net radiation.
From the sub-regional analysis in Fig. 13, Noah appears to be the best performing scheme to simulate the net annual radiation climatology with relatively small annual climatology biases. In other metrics, the same scheme performs worse (relative to the rest), for example, in box G (Maghreb) for the 95th percentile and standard deviation, although all LSS have distinctly different biases here compared to the other sub-domains and metrics. The largest differences of the standard deviation (simulated minus observed) are calculated for region B (which includes the Balkans). The biases in linear trends of all sub-regions and simulations are positive with the largest obtained with the CLM run.
In Fig. 14 (top map) of the annual mean climatology of soil moisture from satellite observations (SMAP), it is evident that the northern part of the MENA domain is moister than the southern part. When comparing the six simulations with observations (Fig. 14), all schemes (except CLM) have relatively small biases (between – 0.05 and + 0.05 m\(^3\)/m\(^3\)) with an overall underestimation of soil moisture in the African continent and the Arabian peninsula and overestimation in the northern part of the domain. CLM exhibits larger biases both positive and negative.
In the Taylor diagram (Fig. 15), the normalized standard deviation and RMSE take large values values between 2 and 8 for all sub-domains, and correlation varies from 0.05 (box F with CLM) to 0.65 (MENA with Noah). These results appear more scattered in the diagram, implying a more variable performance which can be also seen in Fig. 16 where more sub-regional features are revealed for the additional metrics. An example is the opposite behaviour that CLM shows when looking at the upper and lower monthly distribution, where it takes the first and the last place in the ranking for the 95th and the 5th percentile respectively, over the whole MENA region. Overall, it seems that there is not much affinity in these biases with the respective ones for precipitation where, for example, the RUC scheme performs better in sub-domain D (and worse for the soil moisture).
A three-method intermediate ranking is applied and a further ranking, following the procedure described in Sect. 2.2, generates a final ranking. The results from the grand ranking are presented in Table 3 for each climatic variable considered, for the whole MENA domain (labelled “MENA”) and for the average of the seven sub-domains (labelled “all”). The latter distinction allows a scheme ranking based on the selected sub-regions of interest without considering the tropics (which are included in the whole “MENA” domain).
For most of the variables and schemes, the two ranking results (“all”, “MENA”) coincide. Generally, Noah ranks first for most variables, with the exception of air temperature. Both options of NoahMP (an augmented version of Noah) follow, succeeded by RUC (nine soil layers) and CLM, which also consider a more detailed scheme than Noah. The least performing LSS overall is RUC (six soil layers).
Excluding radiation, the final ranking results for the other four climatic parameters are further grouped into two different ways (“air” vs “land” and “thermal” vs “humid”) to provide another perspective to the assessment, as follows: air [mean 2m air temperature (Tmean) and precipitation (prcp)] vs land [soil moisture (smois) and land surface temperature (Tland)]; thermal [(mean 2-m air temperature (Tmean) and land surface temperature (Tland)] vs humid [(precipitation (prcp) and soil moisture (smois)]. The “air” variables are better simulated using RUC (nine soil layers), whereas Noah LSS performs best when considering “land” and “humid” variables. The group of air and land temperatures (“thermal”) is best simulated using the option of NoahMP with the dynamical vegetation option turned on.