Evaluation of Regional Climate Models (RCMs) Performance in Simulating Seasonal Precipitation over Mountainous Central Pindus (Greece)

During the last few years, there is a growing concern about climate change and its negative effects on water availability. This study aims to evaluate the performance of regional climate models (RCMs) in simulating seasonal precipitation over the mountainous range of Central Pindus (Greece). To this end, observed precipitation data from ground-based rain gauge stations were compared with RCMs grid point’s simulations for the baseline period 1974–2000. Statistical indexes such as root mean square error (RMSE), mean absolute error (MAE), Pearson correlation coefficient, and standard deviation (SD) were used in order to evaluate the model’s performance. The results demonstrated that RCMs fail to represent the temporal variability of precipitation time series with exception of REMO. Although, concerning the model’s prediction accuracy, it was found that better performance was achieved by the RegCM3 model in the study area. In addition, regarding a future projection (2074–2100), it was highlighted that precipitation will significantly decrease by the end of the 21st century, especially in spring (−30%). Therefore, adaption of mountainous catchment management to climate change is crucial to avoid water scarcity.


Introduction
During the last decades, increasing water demand (population growth, urbanization, and industrialization) and recent global warming are making water a precious, but not always available, asset [1,2]. Especially, the Mediterranean climate favors the development of water scarcity [3,4], as the precipitation regime presents spatial and temporal variability [5][6][7]. Moreover, the Intergovernmental Panel on Climate Change (IPCC) Fifth Assessment Report [8] emphasizes that the Mediterranean basin is expected to be drier by the end of the 21st century and much warmer than the global mean [9]. The aforementioned changes will have environmental and economic impacts as they affect flooding [10], soil erosion [11,12], drought phenomena [13,14], agriculture [15,16], wildfires [17,18], and tourism [19,20].
Precipitation is a key factor for the assessment of climate change impacts, due to its direct influence on water availability and natural hazards (floods, drought, landslides, etc.). Its spatial and temporal distribution is important to various scientific fields such as geography, hydrology, forest management, agriculture, ecology, and others [6,21,22]. Future projections are also essential for rational water resources management. Therefore, high spatial resolution spatial climate data are necessary for climate change impact assessment studies. For the Greek territory, located in the eastern Mediterranean, reducing monthly accumulated precipitations and more frequent extreme precipitation events, combined with increasing temperatures, are expected under future climate conditions [23][24][25].
Currently, general circulation models (GCMs) are the most advanced tools for the estimation of future climate changes in global scale. GCMs are numerical models representing physical processes in the atmosphere, ocean, cryosphere, and land surface, based on the principals of energy and water balance. They are tools for simulating climate responses in increasing concentrations of greenhouse gases and are able to provide assessments of climate change.
The spatial analysis of GCMs (200 to 300 km) gives satisfactory results for the global circulation across the planet. However, they have been proved to be ineffective in simulating hydrological variables at the catchment-scale due to their coarse spatial resolution. In addition, they cannot accurately simulate regional scale phenomena due to local conditions and particularities, such as the complex topography, lakes, and small islands [26].
In order to meet the demands of hydrological impact assessment studies for regional precipitation, several downscaling techniques have been developed to bridge the gap between the large-scale GCM estimation and local needs. The downscaling techniques are divided into statistical [27,28] and dynamic [29,30]. The statistical methods use the observed relationships between large-scale circulation and the local climate, whereas dynamic methods use physically based regional climate models (RCMs). The main advantages of a statistical downscaling approach are low computational requirements, whereas dynamic downscaling is appreciated by researchers for its superiority of embracing more systematic characteristics in relation to topography and climatic dynamical processes. In recent years, several studies used RCMs in order to assess climate change effect on hydrology [31][32][33][34][35]. The effectiveness of these models is mostly dependent on their inputs, especially the past climate data [36]. The comprehension of the hydrological cycle components response to the climate variability has become more and more fundamental. Understandings of the uncertainties in future climate projections are of major interest. Therefore, the ability of RCMs to represent climate conditions should be examined prior to their use in impact assessment studies [37][38][39].
In the frame of the European Union (EU) funding project ENSEMBLE, a set of multi-model RCM simulations to characterize climate change in Europe with high spatial resolution (25 × 25 km), were produced. There are many studies in Europe evaluating the ability of RCMs to represent precipitation [40][41][42][43][44]. It is noteworthy that most of these studies referred to plain areas and lowlands. Nevertheless, limited studies evaluate the performance of RCMs simulations in mountainous areas of the Mediterranean basin [45], while there is no such research for the mountainous ranges of Greece. Mountainous areas are of great interest for water resources management, since runoff generates and supplies (through catchments) lowlands with water.
This study aims to evaluate the performance of RCMs from the ENSEMBLE project in simulating the seasonal and annual precipitation over the mountainous range of Central Pindus (Greece) and quantify the effect of climate change on future precipitation.

Materials and Methods
Observations of monthly accumulations of precipitation data from 13 ground-based rain gauge stations over the mountainous range of Central Pindus (Greece) were used for the period 1974-2000. It is noteworthy that the study area has a dense network of rain gauge stations located in high elevation, in comparison with other areas within Greece. The data series are complete without missing values. The station data were checked for homogeneity using the Alexandersson [46] homogeneity test, on a monthly basis and for each station separately. The results verified that the data from all the stations are homogeneous. These stations are operated by the Ministry of Environment & Energy (Agiofylo, Agnanta, Chrysomilia, Elati, Katafyto and Malakasi, Megali Kerasia, Platanousa, and Theodoriana), the Public Power Corporation (Mesochora, Polyneri, and Stournareika) and the University Forest Administration and Management Fund (Pertouli). The characteristics of the above-mentioned stations are given in the next table (Table 1). Additionally, precipitation data were derived from five (5) RCMs. Simulations were developed for the European ENSEMBLE project (http://ensembles-eu.metoffice.com/), with high spatial resolution (25 × 25 km) under the A1B future emission scenario of the Special Report on Emissions Scenarios (SRES). The selected models were also driven by fifth-generation GCM ECHAM5-r3. The details of the RCMs used in this study are given in the following table (Table 2). In order to evaluate the RCMs performance, simulated annual and seasonal precipitation values from the RMCs were compared with the observed records of precipitation from the nearest ground-based rain gauge stations over the baseline period 1974-2000 ( Figure 1). The rain gauge within a 25 km radius and with an elevation difference of less than 200 m from a model grid point was considered as the nearest neighbor. With such an approach, the rain gauges used in this study lie on the same side of the Pindus' slopes and within comparable elevation as the RCMs grid points.
Initially, the statistical evaluation was presented with Taylor diagrams. These diagrams provide a way of graphically summarizing how closely a pattern (or a set of patterns) matches observations [52]. The similarity between two patterns was quantified in terms of their correlation and the normalized standard deviation. Furthermore, the root mean square error (RMSE) and mean absolute error (MAE) were used as evaluation indexes and presented in separate diagrams. The mathematical formulas of these indexes are given below:  Initially, the statistical evaluation was presented with Taylor diagrams. These diagrams provide a way of graphically summarizing how closely a pattern (or a set of patterns) matches observations [52]. The similarity between two patterns was quantified in terms of their correlation and the normalized standard deviation. Furthermore, the root mean square error (RMSE) and mean absolute error (MAE) were used as evaluation indexes and presented in separate diagrams. The mathematical formulas of these indexes are given below: where n is the number of observations, xi and yi are the observed and simulated values, respectively, for i = 1,2…n. The RMSE gives the weighted variations (residuals) in errors between the simulated and observed values, while MAE measures the weighted average magnitude of the errors. Since the errors in RMSE are squared before they are averaged, the RMSE gives a relatively high weight to large errors. This means that the RMSE should be more useful when large errors are particularly undesirable. MAE is the most natural and unambiguous measure of average error magnitude [53-

Results
Historical RCMs precipitation datasets were evaluated against the ground-based rain gauge stations for the period 1974-2000. An overview of the results for annual and seasonal precipitation is presented by Taylor diagrams (Figure 2).
The results showed low correlation for the majority of the RCMs for both annual and seasonal precipitation. It became clear that the better correlation coefficient was achieved by the REMO model followed by the RegCM3. Regarding the annual and autumn precipitation, the results better correlated with observations for the majority of the models, whereas the lowest coefficient correlation was found in summer precipitation. Concerning the variability of precipitation, the differences of the standard deviation values showed that the models did not manage to generate a low standard deviation sufficiently during the selected period. However, it should be mentioned that the REMO model presented better results, especially in summer, while the worst results were presented in autumn. To this end, it is obvious that RCMs lack the ability to realistically represent the temporal variability of precipitation time series.

Results
Historical RCMs precipitation datasets were evaluated against the ground-based rain gauge stations for the period 1974-2000. An overview of the results for annual and seasonal precipitation is presented by Taylor diagrams (Figure 2). The results showed low correlation for the majority of the RCMs for both annual and seasonal precipitation. It became clear that the better correlation coefficient was achieved by the REMO model followed by the RegCM3. Regarding the annual and autumn precipitation, the results better correlated with observations for the majority of the models, whereas the lowest coefficient correlation was found in summer precipitation. Concerning the variability of precipitation, the differences of the standard deviation values showed that the models did not manage to generate a low standard deviation sufficiently during the selected period. However, it should be mentioned that the REMO model presented better results, especially in summer, while the worst results were presented in autumn. To this end, it is obvious that RCMs lack the ability to realistically represent the temporal variability of precipitation time series.
Moreover, the RMSE and MAE were used as statistical indexes for evaluating the model's performance. In spite of their arithmetic similarity with the standard deviation, these two statistics are conceptually quite different: the RMSE and MAE are dispersions around a true value and are measures of accuracy that can be used to confirm or deny observation bias. The results of the %RMSE and %MAE can be seen in Figure 3. Moreover, the RMSE and MAE were used as statistical indexes for evaluating the model's performance. In spite of their arithmetic similarity with the standard deviation, these two statistics are conceptually quite different: the RMSE and MAE are dispersions around a true value and are measures of accuracy that can be used to confirm or deny observation bias. The results of the %RMSE and %MAE can be seen in Figure 3. Regarding the results of Figure 3, it is obvious that the best simulation was achieved from the RegCM3 model. In the case of winter, the RegCM3 model's effectiveness was found to be poor and the results were moderate and less accurate, especially in comparison with autumn and spring. On the other hand, the highest values of RMSE and MAE were found for the HIRHAM model. In addition, relatively high values of errors were also estimated for annual and autumn precipitation by the REMO model, for winter by the RCA model, and for spring and summer by the RACMO2 model.
To this end, the effect of future climate change on annual and seasonal precipitation was studied Regarding the results of Figure 3, it is obvious that the best simulation was achieved from the RegCM3 model. In the case of winter, the RegCM3 model's effectiveness was found to be poor and the Water 2020, 12, 2750 6 of 10 results were moderate and less accurate, especially in comparison with autumn and spring. On the other hand, the highest values of RMSE and MAE were found for the HIRHAM model. In addition, relatively high values of errors were also estimated for annual and autumn precipitation by the REMO model, for winter by the RCA model, and for spring and summer by the RACMO2 model.
To this end, the effect of future climate change on annual and seasonal precipitation was studied comparing the baseline period  with the future (2074-2100) precipitation simulation of the most accurate model (RegCM3).
The analysis of the future changes in the precipitation regime showed significant decrease of the amount of precipitation. The graphical representation of the results can be seen in the following figure  (Figure 4).  Regarding the results of Figure 3, it is obvious that the best simulation was achieved from the RegCM3 model. In the case of winter, the RegCM3 model's effectiveness was found to be poor and the results were moderate and less accurate, especially in comparison with autumn and spring. On the other hand, the highest values of RMSE and MAE were found for the HIRHAM model. In addition, relatively high values of errors were also estimated for annual and autumn precipitation by the REMO model, for winter by the RCA model, and for spring and summer by the RACMO2 model.
To this end, the effect of future climate change on annual and seasonal precipitation was studied comparing the baseline period  with the future (2074-2100) precipitation simulation of the most accurate model (RegCM3).
The analysis of the future changes in the precipitation regime showed significant decrease of the amount of precipitation. The graphical representation of the results can be seen in the following figure  (Figure 4). The highest percentage change, taking into account all grid points, was found in spring (30%), while annual winter and summer precipitation had almost the same percentage of decrease, around 15%. On the other hand, autumn precipitation seems to be stable until the end of the 21st century.  The highest percentage change, taking into account all grid points, was found in spring (30%), while annual winter and summer precipitation had almost the same percentage of decrease, around 15%. On the other hand, autumn precipitation seems to be stable until the end of the 21st century.

Discussion and Conclusions
Precipitation is one of the most important meteorological parameters with significant spatial and temporal variation. Nowadays, there is an urgent need for reliable future climate projections for rational catchment management and infrastructure projects scheduling. The most modern tool for simulating future climate conditions is the regional climate models (RCMs). However, before using climate simulation from dynamic downscaling, it is appropriate to evaluate their performance at different past spatial scales.
In the present study, the performance of five RCMs from the European project ENSEMBLE were evaluated in simulating annual and seasonal precipitation over the mountainous range of Pindus (Central Greece).
The results highlighted that RCMs lack the ability to represent temporal variability of precipitation time series, with the exception of the REMO model. In addition, concerning the model's prediction accuracy, it was found that better simulation was achieved for spring and autumn precipitation. The results were in agreement with similar studies in the Mediterranean basin [57,58]. Comparison between models showed that the RegCM3 model had the smaller errors. It was also suggested that finer spatial resolution RCMs of the EURO-CORDEX project better simulated precipitation over complex terrain areas [59].
Regarding the future precipitation regime, based on the RegCM3 model, a decrease of precipitation is expected both annually and seasonally until the end of the 21st century. Similar results were also reported in other studies in the Mediterranean basin [14,16,23,26]; however, these studies referred to lowland areas. On the contrary, in Central Europe, even though annual precipitation is projected to increase up to +10%, RCMs project a significant decrease of precipitation in summer [60].
Over the mountainous range of Central Pindus, the higher decrease of precipitation is expected in spring (30%), while in autumn, the accumulated precipitation will stay close to the current values. As for the annual precipitation in the summer and the winter, a future decrease of precipitation around 15% was found. These will favor the development of aridity phenomena in crucial mountainous ecosystems and probably effect biodiversity and water production. The findings of the current study contributed to the rational water resources management and infrastructure work scheduling based not only in current, but also future conditions. The decrease of future precipitation has to be considered by decision makers, and water reservoirs must be planed, especially in mountainous regions where there is excess water, especially in the winter and autumn months. It is suggested that similar research should be done in other mountainous Greek regions; the analysis should also include more factors of the hydrological cycle, such as temperature, humidity, wind, and snowfall, as well as a study of the associated changes in future atmospheric circulation [61]. Moreover, the examination of simulating finer timescale (e.g., daily, hourly) could raise significant conclusions on the effect of climate change on hydrometeorological hazards.