Evaluation of the Performance of CFSR Reanalysis Data Set for Estimating Potential Evapotranspiration (PET) in Turkey

7 Evapotranspiration is an important parameter for hydrological, meteorological and agricultural studies. 8 However, the calculation of actual evapotranspiration is very challenging and costly. Therefore, 9 Potential Evapotranspiration (PET) is typically calculated using meteorological data to calculate actual 10 evapotranspiration. However, it is very difficult to get complete and accurate data from meteorology 11 stations in, rural and mountainous regions. This study examined the availability of the Climate 12 Forecast System Reanalysis (CFSR) reanalysis data set as an alternative to meteorological observation 13 stations in the computation of potential annual and seasonal evapotranspiration. The PET calculations 14 using the CFSR reanalysis dataset for the period 1987-2017 were compared to data observed at 259 15 weather stations observed in Turkey. As a result of the assessments, it was determined that the seasons 16 in which the CFSR reanalysis data set had the best prediction performance were the winter (C'= 0.76 17 and PBias = -3.77) and the autumn (C' = 0.75 and PBias = -12.10). The worst performance was 18 observed for the summer season. The performance of the annual prediction was determined as C'= 19 0.60 and PBias = -15.27. These findings indicate that the results of the PET calculation using the 20 CFSR reanalysis data set are relatively successful for the study area. However, the data should be 21 evaluated with observation data before being used especially in the summer models. different reanalysis datasets. They compared the estimates calculated using the with the results of the observation stations and reported that the CFSR dataset is a good et al. calculated the potential evapotranspiration required for hydrological modeling with the CFSR reanalysis data set in their study. They explained that the predictive performance of the CFSR dataset was good by evaluating the results obtained. These studies show that reanalysis datasets such as CFSR are of sufficient quality resolution to be used as inputs in basin modelling studies. In addition, this dataset can be an important for overcoming problems encountered in obtaining meteorological observation data. The of this study is to investigate the availability and use of the CFSR reanalysis dataset for the of using the FAO56-Penman method in


Introduction 24
The amount of water that evaporates from soil surfaces or open water and the transpiration of 25 plant leaves in the atmosphere is known as evapotranspiration (Tabari et al. 2013; Anderson et al. (1) 123 where; is the slope of the relationship between saturation vapor pressure and mean daily air 124 temperature (kPa °C -1 ), Rn is the net radiation at the crop surface (MJ m -2 day -1 ), G is the soil heat flux 125 density (MJ m -2 day -1 ), γ is the psychrometric constant which depends on the altitude of each location 126 (kPa °C -1 ), T is the mean daily air temperature (°C), u2 is the wind speed at 2 m height (m s -1 ); es is the 127 saturation vapor pressure (kPa); ea is the actual vapor pressure (kPa). 128

Inverse Distance Weighting (IDW) method 129
In this study, the IDW interpolation method was used to produce a spatial distribution of the where Zj is the unsampled location value, Zi is the known cell's value, β is the weight, and δ is 139 the parameter. The separation distance hijk is measured by a three-dimensional Euclidian distance.

Evaluation criteria 145
The four statistical methods were used to assess the PET estimates from the CFSA dataset 146 against the PET calculated using meteorological station data. These are coefficient of determination 147 (R 2 ), root-mean-square error (RMSE), PBias (percent bias), and the performance index (C'). ) 2

(4) 153
The value of RMSE should always be positive and it is desired to be close to zero. This 154 indicates that the lower the value, the better the model will perform. RMSE provides performance  PBias is determined by Eq.6. 162 The Willmott index of agreement (d) shows the degree of fit between observed and predicted 164 measurements between 0 and 1. The closer the result is to 1, the better the model performance is

Results and Discussion 176
Using meteorological data from the CFSR data set, the average seasonal and annual potential 177 evapotranspiration amounts for each observation station for 1987-2017 were estimated. The PET 178 estimates were compared with the PET computed using ground observation data. The accuracy and 179 usability of the CFSR reanalysis dataset were evaluated through statistical analysis. The results of that 180 analysis are presented in Table 2. In addition, maps were generated using IDW interpolation 181 techniques to show area distributions of PET results for the different seasons and the long-term annual 182 mean. 183

Results of PET estimation for the winter 184
The PET prediction results using the CFSR dataset for the winter season (December, January, 185 February) were compared with the PET results obtained from the observed data. The obtained values 186 for the stations were used to generate PET maps of Turkey.
9 Table 2. Results of the statistical analysis 188

189
PET values were classified into 6 categories between 20 and 200 mm, as shown in Figure 2. 190 When the map is examined, it is seen that CFSR has higher estimates in the southern and western 191 regions, but lower forecasts at eastern stations. It can be explained that the CFSR reanalysis dataset 192 has relatively high data on temperature and solar radiation in those regions, unlike the eastern region. the CFSR has a tendency to predict higher temperatures (>2°C) in the southwest in winter and colder 195 temperatures in the northeast. Station estimates were compared using the calculated PBias value (-196 3.77) and PET calculated using CFSR data was determined to be relatively high. However, according 197 to Table 2, these estimates are in acceptable ranges (very well <±10). 198

Evaluating PET estimates for the spring 215
The spatial distributions of the estimated and observed PET results over Turkey for the spring 216 season (March, April, May) are given in Figure 4. It was seen that there were relatively similar results 217 for the stations, especially in the inner regions. However, as in the winter forecasts, the CFSR has 218 shown overestimates in the southern and western regions, and underestimates at stations in the 219 northeast region. It is seen that the CFSR reanalysis data set tends to predict PET higher than the 220 observation data. This is likely due to the CFSR having overestimations for the stations having a 221 relatively higher temperature, solar radiation, and wind speed than others (Paredes et al. 2017). 222 The R 2 value was found 0.67 as seen in the scatter plot in Figure 5. This shows that the CFSR 223 re-analysis dataset has a good correlation with the observed data. The RMSE value was 33.85 mm 224 season -1 , and the C 'performance index, was 0.70. When the station estimates are compared, the 225 calculated PBias (-6.24) value indicates that the CFSR re-analysis made relatively overestimates, but 226 according to Table 2, it is in acceptable ranges (very good <±10). According to these performance 227 evaluations, it was concluded that estimations of PET using the CFSR data set are also good for spring 228 seasons. 229

Evaluating PET estimates for the summer season 234
When the predictions made by the CFSR for the summer season (June, July, August) are 235 compared with the observation data, the differences between the results are higher than in other 236 seasons as seen in Figure 6. The reason for this thought is that temperature and solar radiation increase 237 considerably in the summer months and the CFSR reanalysis data set cannot accurately predict these 238 changes. CFSR was generally overestimated from the observation data where differences were higher 239 between winter and summer seasons, especially in the southeastern and western regions. 240 PBias value was calculated -16.94 for the summer season. It shows that the CFSR re-analysis 241 made higher estimates in summer than winter and spring, but estimated PET for the summer is still in 242 acceptable (<± 25) ranges. 243 Although PET estimates are acceptable in terms of R 2 (0.67), the RMSE had the highest error 244 (RMSE=103.10 mm season -1 ), and the C' performance index is 0.57. According to these performance 245 evaluations, the C' value of the CFSR estimates is not acceptable (<0.60). The reason why the PET 246 prediction of the CFSR re-analysis dataset underperforms in the summer is due to the decrease in solar 247 radiation and temperature prediction capabilities. The reason can be explained that more convective 248 warming occurs in summer compared to other seasons. This type of convection may cause the 249 formation of different weather conditions on a small scale that CFSR cannot predict due to its coarse 250 resolution (Tian et al., 2014). Using the CFSR data set directly on models for the summer months will 251 result in unsuccessful simulation results. For this reason, preliminary procedures that will reduce this 252 dataset to a regional scale should be applied and re-evaluated before using it.

Evaluating PET estimates for the autumn 258
The predictions made by the CFSR for the autumn season (September, October, November) 259 PET estimated higher than observation data as seen in Figure 8. The PBias was found -12.10. This 260 shows that the CFSR re-analysis estimates PET is good (<± 15). The boxplot graph for the autumn season between the CFSR reanalysis and observation data 266 sets is given in Figure 9. When comparing the situation between quarters, it is seen that the higher 267 estimates of CFSR for the autumn season are more intense. Because of the R 2 value found 0.79, PET 268 estimates of CFSR data are good for the autumn season. This shows that the CFSR re-analysis dataset 269 has a good correlation with the observed data. The RMSE value and C' performance index were 270 calculated 32.47 mm season -1 , and 0.75 respectively. According to the performance evaluation, the C 271 'value of the CFSR estimates is quite good (>0.75). These results show that the CFSR estimates of 272 PET for the autumn season can be used safely. 273

Evaluating long-term average annual pet estimates 274
The long-term annual average PET estimations using the CFSR data set and observed data for 275 the years 1987-2017 are shown in Figure 10. PET was estimated between 1300-1900 mm/year for 276 Southern and Western regions in Turkey. In this region, PET was calculated between 1100-1700 mm 277 using data from meteorological observation stations. In contrast, estimated and observed PET were 278 found lower in the northern and eastern regions of Turkey. Estimated PET using CFSR was between 279 700-1100 mm year -1 in the northern and eastern regions and 900-1300 mm in inland regions. These 280 estimates are very close to the observation data as can be seen in Figure 10. In the study conducted in 281 China, while PET estimates for the south and west regions were high, it was observed that the PET 282 estimates for other regions were similar to the station results (Tian et al. 2018). Calculated and estimated PET distributions using the average long annual data are shown in 289 Figure 11 with the boxplot. When comparing the situation between quarters, it is seen that the higher 290 estimates are more intense. It has been observed that the minimum values are close to each other, but 291 in the difference between the maximum values, it is seen that the CFSR tends to overestimate PET 292 annually from the observation data. The R 2 value (0.68) shows that the annual PET estimates are 293 acceptable and have a good correlation with observation data. The RMSE value showing the amount 294 of error in the data set was calculated as 208.37 mm year -1 . The C 'performance index, which shows 295 the success of the predictions, was obtained as 0.60. According to the performance evaluation, it has 296 been observed that the C' value of the CFSR estimates is acceptable (> 0.60). Alfaro et al. (2020) 297 calculated the C 'performance index of the CFSR reanalysis data set as 0.72 in their study conducted in 298 Brazil and explained that the prediction performance was acceptable similar to our study results. 299 Figure 11. Boxplot and scatter plot of long-term average annual CFSR re-analysis data set 301

Conclusion 302
PET is a very important parameter for hydrological, meteorological, and agricultural studies. 303 However, it is very difficult to obtain the meteorological data for calculation or estimation of this 304 parameter in developing countries for the required regions. In this study, PET was estimated by the 305 FAO56-PM method using observed and CFSR data set for Turkey. Accuracy of seasonal and annual 306 estimations was statistically evaluated by comparing calculated PET. Data from 259 stations covers 307 the period from 1987 to 2017 used to calculate PET. 308 As a result of the evaluations, the periods in which the prediction performance of the CFSR 309 reanalysis data set was the highest were determined as Winter (C'= 0.76 and PBias = -3.77) and 310 Autumn (C' = 0.75 and PBias = -12.10) seasons. Also, the lowest RMSE values were calculated (22.27 311 and 32.47) in these two seasons. The worst performance was seen for the Summer season (C'= 0.57 312 and PBias = -16.94). The reason for this, the increase in solar radiation and temperature values during 313 the summer months cannot be estimated by the CFSR accurately as mentioned by Tian et al. (2014). In 314 terms of annual performance, it has been calculated as C'= 0.60 and PBias = -15.27. These results 315 show that the PET prediction ability of the CFSR re-analysis dataset is relatively good for the study 316 area. 317 PBias value was calculated as negative in annual and seasonal evaluations. Especially in the 318 southern and western regions, it has been observed that CFSR tends to overestimate the observation 319 Therefore, when the CFSR reanalysis data set is evaluated in general, it can be seen as a good 322 potential data source. However, it is recommended to evaluate the data with observation data before 323 being used especially in summer seasons and to be used after regionalization with downscaling 324 methods before being used in models. 325