3.1.2. Solar Radiation

Table 3 shows the statistical indicators of solar radiation (*Rs*) in the CLDAS data for the seven climate zones. Across climate zones, RMSE ranged from 5.18 to 6.21 MJ m−<sup>2</sup> d<sup>−</sup>1, and MAE ranged from 3.83 to 4.54 MJ m−<sup>2</sup> d<sup>−</sup>1. The differences in *Rs* errors among different climate regions were not as apparent as those in air temperature. However, the *R*<sup>2</sup> of climate zone 7 was significantly lower than that of other climatic regions. These results were similar to the results reported by Liu et al. (2009) [38]. However, their values were generally higher than that of the radiation model based on temperature, where the median RMSE was 3.3 MJ m−<sup>2</sup> d−<sup>1</sup> in humid regions of China (Fan et al., 2019) [39]. The above phenomenon indicated that the radiation data in the CLDAS data set did not perform well.

Figures 2–4 show the spatial distribution of *Rs* errors. Overall, the error of *Rs* in climate zones 1–3 was better than in other climate zones. The RMSE of most stations was more significant than 6 MJ m−<sup>2</sup> d−1. This might be due to the severe air pollution in the above areas [40], which would pose particular challenges to accurate simulation.


**Table 3.** Statistical indicators of solar radiation, relative humidity, and wind speed in different climate zones of China.

#### 3.1.3. Relative Humidity

Statistical indicators of CLDAS RH are shown in Table 3. Climate zone 7 had the most significant error among all regions, with RMSE and MAE reaching 31.29% and 27.62%, respectively, close to a random distribution. The consistency between CLDAS and site data in other climatic regions was also not high, with *R*<sup>2</sup> ranging from 0.39 to 0.59. However, the values of RMSE and MAE indicated that they were still within acceptable limits. Compared climate zone 1 with climate zone 6, the consistency of climate zone 1 was higher, but the RMSE and MAE of climate zone 6 are lower. This result could be attributed to the fact that climate zone 6 is within the humid region with high annual average relative humidity, while climate zone 1 is in the arid area where the relative humidity changes more sharply. Figures 2–4 showed that the overall error of this data set was larger in climate zone 7 than in other climate zones. In addition, compared with the northeast part of climate region 5, the error of RH in the western part of the same climate zone (i.e., areas bordering climate zone 7) was significantly larger. Although there was a relatively large error for RH in some regions, the estimation of *ET*<sup>0</sup> would unlikely be affected, as previous studies have found that RH would have a low contribution to *ET*<sup>0</sup> in most regions of China [41].

### 3.1.4. Wind Speed

Statistical indicators of CLDAS U are shown in Table 3. From the perspective of *R*2, the consistency between CLDAS near-surface wind speed and station data was poor in all climate zones, while from the perspective of RMSE and MAE, their accuracies were acceptable. In addition, the mean difference between climate zones was within 30%. However, according to the spatial distribution of the error (Figures 2–4), the RMSE of some stations was more than3ms<sup>−</sup>1, of which the errors of most stations in climate zone 7 were significant. Compared with the inland stations, the *R*<sup>2</sup> of the coastal stations was higher. However, RMSE and MAE were also higher, indicating a problem of overestimation or underestimation. The worldwide modeling for wind speed is challenging and often inaccurate. Similar results were obtained for ERA5 [22], NCEP/NCAR [27], and GLDAS [42]. This is mainly due to the complex terrain changes on the ground, and the wind speed is greatly affected by the roughness of the underlying surface. In addition, it is not easy to simulate the movement direction of winds accurately.

#### *3.2. Reference Evapotranspiration*

The statistical indicators of calculated *ET*<sup>0</sup> based on the CLDAS dataset are shown in Table 4. Among all climate zones, climate zone 1 had the best consistency (*R*<sup>2</sup> = 0.84) between CLDAS data and station data, while climate zone 3 showed the lowest errors (RMSE = 0.87 mm d−<sup>1</sup> and MAE = 0.58 mm d−1). For climate zone 7, the values of RMSE (1.37 mm d<sup>−</sup>1) and MAE (1.19 mm d−1) were higher than the corresponding values in any of the other climate zones. Figure 5 shows the spatial distribution of statistical indicators. Across climate zones, *R*<sup>2</sup> overall showed a decreasing trend from the north to the south,

and the southernmost region (i.e., climate zone 6) had the lowest value of *R*2. However, the spatial distributing patterns of RMSE and MAE were different from *R*2. The stations with significant errors are mainly distributed west of climate zone 1, the coastal areas, and the boundary areas between climate zone 7 and other climate zones. This is mainly due to the more complex climate change between climate zones. In addition, the high wind speed error in the coastal areas often leads to a significant *ET*<sup>0</sup> error.


**Table 4.** Statistical indicators of reference evapotranspiration in different climate zones of China.

**Figure 5.** Spatial distribution of *ET*<sup>0</sup> statistical indicators.

To explore the differences in the CLDAS data in different climate regions, one station from each climate zone was randomly selected to fit the correlation between the calculated *ET*<sup>0</sup> based on CLDAS and the FAO56-PM *ET*<sup>0</sup> (Figure 6). Although there were a few outliers, the scatter points in climate zone 1 were more concentrated to the 1:1 line than those in other climate zones. Scatter points in climate region 2 were slightly more dispersed than in climate zone 1 and showed some obvious overestimations when *ET*<sup>0</sup> was more significant than 6. In climate zone 3, the accuracy was excellent when the value of *ET*<sup>0</sup> was low (<2 mm) but showed a decline as the following scatter points started to discretize. However, no overestimation or underestimation existed. In climate zone 4, the error was relatively large when *ET*<sup>0</sup> ranged from 3 mm to 6 mm. When the *ET*<sup>0</sup> of climate zone 5 was less than 2, the

problem of underestimation appeared, and then the points were scattered in the 1:1 line for two measurements, but the distance from the 1:1 line was far. In climate zone 6, the error was significant when *ET*<sup>0</sup> was greater than 3 mm, and some scatter points were obviously overestimated or underestimated. Although the points were not as discrete as those in climate zones 4 and 5, the *ET*<sup>0</sup> of climate zone 7 showed a significant underestimation.

Figure 7 shows the *ET*<sup>0</sup> box diagram of a station randomly selected from each climate zone. From the median value, there are differences in the performance of different climate regions. Among them, the *ET*<sup>0</sup> prediction bias of climate region 2 and climate region 3 is slight. The bias of climate region 6 and 7 is large. In addition, from the extreme value, the bias of *ET*<sup>0</sup> estimated in climate zone 4 and climate zone 6 is small, and other regions have overestimated or underestimated in varying degrees. From the quartile line, there are significant differences in *ET*<sup>0</sup> estimation in climate regions 5, 6, and 7. The predicted *ET*<sup>0</sup> performance of climate zones 1, 2, 3, and 4 is relatively good.

**Figure 6.** Scatter plots of CLDAS and FAO56 PM values of *ET*<sup>0</sup> in different climates.

#### *3.3. Seasonal Performance of Reference Evapotranspiration from CLDAS*

Since the demand for water resources varies significantly between seasons, it is necessary to assess the performance of the CLDAS dataset in different seasons. Figure 8 shows the RMSE performance of CLDAS *ET*<sup>0</sup> in the four seasons. In spring (March–May), stations with RMSE smaller than 1.5 mm d−<sup>1</sup> accounted for more than 85% of all stations across China. The RMSE was lower in the south of climate zone 1 and the middle and north of climate zone 3, ranging from 0.5 to 1 mm d<sup>−</sup>1. For most stations of climate zones 2, 4, 5, and 6, RMSE values ranged from 1–1.5 mm d−1. Stations with errors greater than 1.5 mm d−<sup>1</sup> are mainly located in climate zones 1 and 7.

**Figure 7.** The box diagram of CLDAS and FAO56 PM values of *ET*<sup>0</sup> in different climates.

In summer (June–August), the RMSE of CLDAS *ET*<sup>0</sup> was generally higher than that of spring. More than 80% of the stations had RMSE ranging between 1 mm d−<sup>1</sup> and 1.5 mm d<sup>−</sup>1. Stations with RMSE smaller than 1 mm d−<sup>1</sup> were mainly concentrated in the southern part of China and near the boundary area between climate zones 5 and 6. Stations with RMSE greater than 1.5 mm d−<sup>1</sup> were distributed in all climatic regions, of which climate zone 7 had the largest RMSE, followed by climate zone 1. Especially for the southwest area of climate zone 7, stations in this area were sparse, and the error was relatively large, with the value of RMSE larger than 2 mm d<sup>−</sup>1.

In autumn (September–November), RMSE was less than 1 mm d−<sup>1</sup> in 80% of all stations, and stations with a significant error were still mainly concentrated in climate zone 7. It is worth mentioning that there were also many stations with RMSE greater than 1 mm d−<sup>1</sup> in the coastal areas of climatic zone 6. This is mainly due to the relatively high temperature of this area in autumn, resulting in a relatively large RMSE.

In winter (December–February), RMSE in northern regions (i.e., climate zones 1–3) was lower than 0.5 mm d−<sup>1</sup> due to the minimal *ET*<sup>0</sup> value. RMSE of most stations in climate

zones 4 and 5 was less than 1 mm d<sup>−</sup>1. However, the values of RMSE in the southern part of climate zone 7, the coastal part of climate zone 6, and the western part of climatic zone 5 were more outstanding than 1.5 mm d<sup>−</sup>1.

**Figure 8.** Seasonal RMSE of *ET*<sup>0</sup> calculated from CLDAS dataset.

Among all seasons, summer had the most significant RMSE error, followed in order by spring, autumn, and winter. The CLDAS dataset performed well in climate zones 2, 3, 4, and 5, but performed poorly in all seasons in climate zone 7. In addition, the coastal areas of climate zone 6 also did not perform well in autumn and winter.

Because the demand for water resources varies significantly in different seasons, it is also necessary to evaluate the specific overestimation or underestimation of the CLDAS dataset in different seasons. This provides a more detailed reference for practical production and life applications. Figure 9 shows the PBias distribution of *ET*<sup>0</sup> calculated by CLDAS in the four seasons. In spring, the sites with PBias between 0.2 and 0.2 accounted for about 70% of all sites in the country, and the overall forecast stability was good. The values of *ET*0CLD in the southern regions of climate zone 1, climate zone 2, the southern part of climate zone 3, most of climate zone 4, and the central and northern parts of climate zone 5 are within 10% of the local station data. In climate zone 7 (Underestimated), numerical biases are generally greater than 30%. The prediction of *ET*0CLD in the junction area of climate zone 7 and other climate zones is not very stable, and most of them are underestimated. In addition, the *ET*0CLD in coastal areas will have a relatively large bias.

In summer, the bias of *ET*<sup>0</sup> calculated by CLDAS is generally smaller than that in spring, but some sites have large fluctuations (the bias is greater than 60%), and the PBias of more than 60% of the sites is between −10% and 10%. *ET*0CLD is in the climate zone 3. The southeastern coastal areas (overestimated), the southern part of climate zone 5 (overestimated), and the western coastal areas of climate zone 6 (overestimated) have large biases from the local weather station data, with a gap of about 10% to 30%. Numerical bias with zone 7 (underestimation) is generally greater than 30%. The prediction of *ET*0CLD for meteorological stations in the junction of climate zone 7 and other climate zones is not very stable, and most of them are underestimated.

In autumn, the bias of *ET*<sup>0</sup> calculated by CLDAS is generally larger than that in spring and autumn, and only about 50% of the sites have PBias between 10% and 10%. The western region (underestimated), the central and western regions of climate zone 5 (underestimated), the southern coastal region of climate zone 6 (underestimated), and the climate zone 7 (underestimated) have large biases from the data of local meteorological stations, with a gap of more than 30%. In addition, the prediction accuracy of the *ET*0CLD of the meteorological stations at the junction of climate zone 3 and other climate zones decreased significantly. Most showed an overall underestimation.

In winter, the bias of *ET*<sup>0</sup> calculated by CLDAS is generally the largest, among which the bias of *ET*0CLD from the local station data in the southern region of climate zone 1, the central region of climate zone 2, the central region of climate zone 3, and the central and eastern regions of climate zone 5 is 10%. Within %; *ET*0CLD in the northern region of climate zone 1 (overestimated), the southern coastal region of climate zone 3 (overestimated), most of climate zone 4 (underestimated), the central and western regions of climate zone 5 (underestimated), the southern part of climate zone 6 Coastal areas (underestimated), and climate zone 7 have large biases from local weather station data, with a gap of more than 30%. In addition, the prediction accuracy for the *ET*0CLD of the meteorological stations in the transition areas of different climatic zones will drop significantly, and both overestimation and underestimation exist.

Figure 10 shows a boxplot of the calculated PBias for the CLDAS dataset. From the median value of PBias, there are biass in the performance of different seasons. Among these, the estimated bias in spring and summer is smaller, the performance in autumn is second, and the performance in winter is the largest. From the quartile line (aside from winter), the estimated differences in the other three seasons were small. From the perspective of extreme values, the estimated maximum and minimum values in winter are not good. The performance in summer is the best, and the estimated bias is the smallest. In conclusion, of all seasons, summer and spring have the slightest bias, followed by autumn and winter. The CLDAS dataset performs well in climate zones 2, 3, 4, and 5 but not in all seasons in climate zone 7. In addition, the coastal areas of climate zone 3 and climate zone 6 also

performed poorly in autumn and winter, and the performance at the interface of different climate zones was also relatively poor.

**Figure 9.** Seasonal PBias of *ET*<sup>0</sup> calculated from CLDAS dataset.

**Figure 10.** The box diagram of Seasonal PBias of *ET*<sup>0</sup> calculated from CLDAS dataset.

#### *3.4. Annual Performance of Reference Evapotranspiration from CLDAS*

China is a country with frequent droughts and floods. Water demand also varies widely between years. Therefore, it is necessary to evaluate the difference in CLDAS *ET*<sup>0</sup> error in different years. The RMSE of CLDAS *ET*<sup>0</sup> in 2017–2020 is shown in Figure 11. Overall, RMSE in 2019 was lower than that in other years, with more than 85% of all stations having a value less than 1 mm d<sup>−</sup>1. In 2020, stations with RMSE less than 1 mm d−<sup>1</sup> accounted for 60% of the total stations. Regarding the spatial distribution of errors, climate zones 3, 4, 5, and 6 overall performed better than other climate zones. Significant errors were in the southern part of climate zone 7 and the eastern part of climate zone 1, with RMSE generally more significant than 1.5 mm d−1. This may be related to the special geographical location of these stations, such as at the boundary of different climate zones. The above results further confirmed that the data set had good performance in some regions. At the same time, there were also significant uncertainties in other regions, which could bring certain risks to the application.

Figure 12 shows the PB distribution of *ET*<sup>0</sup> calculated by CLDAS in 2019–2020. In 2017, the sites with PBias between 0.1 and 0.1 accounted for about 60% of all sites in the country, and the overall forecast stability was good. The values of *ET*0CLD in the southern region of climate zone 1, climate zone 2, the central and northern regions of climate zone 3, the central region of climate zone 4, the central and northern regions of climate zone 5, and the central region of climate zone 6 are compared with local weather station data. The bias is within 10%; the bias of *ET*0CLD from the local weather station data is larger in the northern area of climate zone 1, the southern coastal area of climate zone 3 (overestimated), and the southern coastal area of climate zone 6 (underestimated), with a gap of 10% to 30%. However, the numerical bias of climate zone 7 (underestimated) is generally greater than 30%. The prediction of *ET*0CLD in the junction area between climate zone 7 and other climate zones is not very stable, and most of them are underestimated.

In 2018, the bias of *ET*<sup>0</sup> calculated by CLDAS was similar to that in 2017, with about 60% of sites having a PBias between 10% and 10%, and *ET*0CLD in the central region of climate zone 1 and a few in the southern part of climate zone 3. The coastal areas (overestimated), parts of the southern part of climate zone 5 (overestimated), and the southern coastal areas of climate zone 6 (overestimated) have large biases from local weather station data, with a gap of about 10% to 30%. The (underestimated) numerical bias is generally greater than 30%. Similarly, the prediction of *ET*0CLD for meteorological stations in the junction of climate zone 7 with other climate zones is not very stable, and most of them are underestimated.

**Figure 11.** Annual RMSE of *ET*<sup>0</sup> calculated from CLDAS dataset.

**Figure 12.** *ET*<sup>0</sup> annual PBias distribution calculated from CLDAS dataset.

In 2019, about 55% of the sites had PBias between −10% and 10%. The *ET*0CLD in the southern part of climate zone 1, most of climate zone 2, the middle part of climate zone 3, small parts of the central region of climate zone 4, the central and eastern regions of climate zone 5, and the central region of climate zone 6 within 10% of the data from local weather stations. *ET*0CLD is in the central region of climate zone 1, the southern coastal (overestimated) and northern regions (underestimated) of climate zone 3, the central and western regions of climate zone 4 (underestimated), and the southern part of climate zone 5 (overestimated). The bias between the southern coastal areas of climate zone 6 (overestimated) and the local weather station data is relatively large, about 10% to 30%. The numerical bias of climate zone 7 (underestimated) is generally greater than 30%.

In 2020, the bias of *ET*<sup>0</sup> calculated by CLDAS was generally the smallest, and about 60% of the sites have PBias between 10% and 10%. *ET*0CLD is in the southern part of climate zone 1 and the southwest of climate zone 2. The central region of climate zone 3, the central and southern regions of climate zone 4, most of climate zone 5, and most of climate zone 6 were within 10% of data from local weather stations. *ET*0CLD is in the central region of climate zone 1, a few southern coastal (overestimated) and northern regions (underestimated) of climate zone 3, the central and western regions of climate zone 4 (underestimated), and a small number of regions in climate zone 5 (underestimated). The bias of local weather station data is large, with a gap of about 10% to 30%. The numerical bias of climate zone 7 (underestimated) was generally greater than 30%. Similarly, the prediction of *ET*0CLD for meteorological stations in the junction of climate zone 7 with other climate zones is not very stable, and most of them are underestimated.

Figure 13 shows a boxplot of annual PBias calculated for the CLDAS dataset. From the median value of PBias, there is a bias in the performance of different years. The estimated bias in 2020 and 2017 is smaller, and the performance of the other two years is relatively poor. From the quartile line, the estimated bias in 2020 is the smallest and more compact, and the estimated differences in other years have different degrees of fluctuation. From the perspective of extreme values, there was a clear overestimation in 2017 and a clear underestimation in 2018. Additionally, 2020 had the best performance, and the estimated bias was the smallest. In conclusion, from 2017 to 2020, bias in 2019 and 2020 was the smallest. The CLDAS dataset performs well in climate zones 2, 3, 4, and 5 but not in all seasons in climate zone 7. In addition, the coastal areas of climate zone 3 and climate zone 6 also performed poorly except in 2020. The performance of the boundary areas of different climate zones was also relatively poor.

**Figure 13.** The box diagram of annual PBias of *ET*<sup>0</sup> calculated from CLDAS dataset.

#### *3.5. Main Factors Affecting Reference Crop Evapotranspiration*

The evapotranspiration process is affected by many factors [43], and its changes are mainly attributed to the changes in meteorological factors. The country's climate is complex and diverse. From a geographical point of view, the eastern part is mostly a monsoon climate zone with a complex and changeable climate. The air above it is severely polluted, affecting solar radiation and surface wind speed. Therefore, the performance of estimated *ET*<sup>0</sup> in coastal areas will decline. The northwest region is far from the sea, is a non-monsoon region, and belongs to a temperate continental climate. The ground topography in this region (the junction of climate zones 2, 4, and 5) is complex and changeable, and the wind speed is greatly affected by the roughness of the underlying surface. The direction of wind movement is accurately simulated, so the reduction in the accuracy of wind speed is likely to lead to a decline in the performance of estimating *ET*<sup>0</sup> in some areas. The Qinghai-Tibet Plateau belongs to the plateau climate zone. Due to its complex and changeable terrain, the climate itself on the Qinghai-Tibet Plateau will fluctuate depending on the region, which greatly affects the estimation of *ET*0. To sum up, the closer are to inland areas (such as climate zones 1, 2, and 3), the higher the accuracy of *ET*<sup>0</sup> estimation will be. The performance of *ET*<sup>0</sup> estimation in coastal areas, the Qinghai-Tibet Plateau, and the junction of climatic zones will be negative effects [44–46]. From the seasonal point of view, the summer is affected by the warm and humid air from the ocean, with high temperature, humidity, and rain. The climate is oceanic, so the estimation error in summer is the largest. In winter, affected by the dry and cold airflow from the continent, the climate is cold, dry, and less rainy, and the climate is continental, and the estimation error will be relatively small [47,48]. In addition, specific regions need to be further analyzed according to the actual situation.
