Evaluation of Precipitation in Multi-Generation Reanalyses with Land Observations over Zhejiang Province

An east coastal province of China, Zhejiang (ZJ) is located in the subtropical region adjacent to the Western Pacific with mean annual precipitation of around 1500 mm. It features a rainy season from May to September when flood disasters occur and lead to huge economic losses and casualties (Zhang et al. ^[1]; Gao et al. ^[2]). Apart from forecasts, a comprehensive understanding of the precipitation characteristics in ZJ would also help achieve more timely preparations for flood disasters. This requires a long-term continuous precipitation dataset; however, the precipitation observed at weather stations in ZJ is spatially inhomogeneous and sometimes absent, which hinders our knowledge of the spatiotemporal characteristics of the precipitation over ZJ.

Atmospheric reanalysis datasets, generated by numerical models and observation assimilation schemes, could fill this gap by providing long-term continuous gridded data. Up to now, reanalyses have been through three generations, e. g., the National Centers for Environmental Prediction (NCEP)-National Center for Atmospheric Research (NCAR) Reanalysis I (R1) (Kalnay et al. ^[3]) and the NCEP-Department of Energy (DOE) Reanalysis II (R2) (Kanamitsu et al. ^[4]) as the first generation, the ECMWF 40-year Reanalysis (ERA-40) (Uppala et al.^[5]) and the Japanese Meteorological Agency (JMA) 25-year Reanalysis (JRA-25) as the second generation (Onogi et al. ^[6]), and the NCEP Climate Forecast System Reanalysis (CFSR) (Saha et al. ^[7]), the ECMWF Interim Reanalysis (ERA-Interim) (Dee et al. ^[8]), the National Aeronautics and Space Administration (NASA) Modern-Era Retrospective analysis for Research and Applications (MERRA) (Rienecker et al. ^[9]), the JMA 55-year Reanalysis (JMA-55) (Kobayashi et al. ^[10]) and the ECMWF Reanalysis 5 (ERA5) (Hersbach et al. ^[11]) as the third generation. Reanalysis datasets have been utilized widely in climate research, initialization and verification of climate prediction models, etc. (Wang et al.; Wheeler and Hendon ^[14]; Duliè et al. ^[15]; Carvalho and Jones ^[16]; Guo et al. ^[17]; Solman and Orlanski ^[18]).

Despite being a rather good representation of atmospheric states, reanalyses are not equivalent to real observations since their reliability could be affected by the qualities and quantities of observations, numerical models as well as assimilation schemes (Decker et al. ^[19]). Thus, it is a necessity to evaluate them before appropriate use (Bengtsson et al. ^[20]), especially for reanalyzed precipitation which, to a large extent, stems from model outputs instead of observations and is thus of relatively low credibility (Kistler et al. ^[21]). Past reanalysis evaluations have primarily concentrated on near-surface temperature (Ma et al. ^[22]; Mao et al. ^[23]; Mooney et al. ^[24]; He and Zhao ^[25]; Zhao et al. ^[26]), surface fluxes (Decker et al. ^[19]; Zib et al. ^[27]; Chen. et al. ^[28]; Wen et al. ^[29]), soil temperature (Yang and Zhang ^[30]), relative humidity (Wang and Zeng ^[31]; Bao and Zhang ^[32]; Brunke et al. ^[33]), greenhouse gases (Davis et al. ^[34]; Guo et al. ^[35]; Zhang et al. ^[36]), oceanic components (Liu et al. ^[37]; Xue et al. ^[38]; Chaudhuri et al. ^[39]), meteorological events (Hodges et _{al. ;} ^[40] Shah and Mishra ^[41]; Chen et al. ^[42]), etc., whereas limited evaluations have been performed with respect to precipitation.

By evaluating daytime and nighttime precipitation in eight reanalyses over China, Zhou et al. pointed out that JRA-55 and CFSR gave the best performances in reproducing the observed nighttime-daytime contrast in precipitation intensity compared with ERA-Interim ^[43]. Huang et al. reported a generally good capture of the climatology and interannual variability of the East Asian summer monsoon precipitation in five reanalyses (R1, R2, JRA-25, ERA-Interim, and MERRA) ^[44]. Bosilovich et al. argued that precipitation in R1 over some oceanic regions exhibited spatial patterns closer to observed precipitation than R2 ^[45]. The study by Lin et al. revealed that five reanalyses (R2, JRA-25, ERA-40, ERA-Interim, and MERRA) could reasonably represent the climatology, long-term trend and interannual variability of global monsoon precipitation with R2 exhibiting the lowest skill, but they all failed to reproduce the increasing tendency of the monsoon precipitation over the North African ^[46]. You et al. implied that most reanalyses (R1, R2, ERA-Interim, ERA-40, MERRA, CFSR, etc.) were able to generally capture the spatiotemporal characteristics and variabilities of mean precipitation over the Tibetan Plateau while overestimating precipitation in the southeast of the Tibetan Plateau ^[47].

Apparently, evaluations of reanalysis datasets are typically carried out on a large spatial scale, e.g., over the globe (Wang et al. ^[13]; Bosilovich et al. ^[45]; Lin et al. ^[46]; Lee and Biasutti ^[48]; Auger et al. ^[49]), East Asia (Huang et al. ^[44]), China (Zhou and Wang ^[43]), and other large parts of the global ocean and land (Wang and Zeng ^[31]; You et al. ^[47]; Zhang et al. ^[50]; Bromwich et al. ^[51]), which is of great significance in comprehensively assessing the quality of reanalysis products. However, such evaluation results could hardly be applied to provincial-level regions where localized climate diagnostics receive great attentions. On the other hand, it has only been around a decade since the third-generation reanalyses emerged, and their performances have not yet been extensively understood, particularly for variables like precipitation that depends mostly on model outputs. As a successor to previous reanalyses NCEP has participated, CFSR remains a doubtful product in terms of its superiority over older-generation products even though being an unprecedented effort of a coupled atmosphere-ocean-sea ice-land framework. For instance, He et al. and Zhao et al. pointed out R2 gives more realistic temperatures over certain areas of China than CFSR does ^[25-26]. Additionally, Auger et al. suggested that most reanalyses show large discrepancies over particular regions despite performing similarly over large domains ^[49]. Consequently, there is much blank in the performances of precipitation in reanalyses of different generations, especially on a local scale. This study aims to comprehensively evaluate the credibility of precipitation in R1, R2, and CFSR over ZJ against land observations from 1979 to 2010 by analyzing multi-scale spatiotemporal characteristics of precipitation, with the hope to enhance our understanding of the performances of multi-generation reanalyses particularly over local areas.

The observed daily and monthly precipitation from 1979 to 2010 at 66 meteorological stations were obtained from ZJ Meteorological Service, which has undergone a series of strict quality control and homogenization measures. The spatial distribution of 66 land stations is shown in Fig. 1. Precipitation in R1, R2 and CFSR are assessed against observations during the same period. Daily and monthly precipitation in R1 and R2 were derived from the daily and monthly precipitation rate in R1 and R2 from the NOAA Physical Sciences Laboratory (https://psl.noaa.gov/data/gridded/reanalysis/). The resolution of R1 and R2 is 2.5° × 2.5° horizontally with 17 pressure levels (Kalnay et al. ^[3]; Kanamitsu et al. ^[4]). Daily and monthly precipitation in the CFSR were derived from the hourly precipitation rate and monthly precipitation rate in the CFSR from the NCAR Research Data Archive (https://rda.ucar.edu/). CFSR is constructed with a high-resolution fully coupled model with the atmospheric component at a horizontal resolution of 0.5° × 0.5° with 64 pressure levels (Saha et al. ^[7]). To facilitate the comparison of the gridded reanalyses with observations, precipitation in three reanalyses is interpolated to each land station via bivariate spline approximation (Nürnberger and Zeilfelder ^[52]).

Figure 1.  Spatial distribution of 66 meteorological stations (dots) over ZJ province.

In addition to a variety of traditional tools in climate diagnostics such as the time series analysis, correlation coefficient, Empirical Orthogonal Function (EOF), root-mean-square error (RMSE), etc., some specific methods are constructed and employed here to further quantify the differences between the reanalyses and observations.

On a monthly basis, the bias level of precipitation in reanalyses (REA) against the observed precipitation (OBS) is categorized into three cases: positive bias when (REA — OBS)/OBS≥0.2, negative bias when (REA—OBS)/OBS≤-0.2, and neutral bias when -0.2 < (REA—OBS)/OBS < 0.2. Then the percentage of each case is computed and compared among all reanalyses. On a daily basis, the observed precipitation is classified into five intensity categories: slight rain (0.1mm-9.9 mm), moderate rain (10.0mm-24.9mm), heavy rain (25.0mm-49.9mm), rainstorm (50mm-99.9mm), and heavy rainstorm (100mm-249.9mm). The extreme rainstorm (≥250 mm) does not appear in all chosen reanalyses and thus are not taken into account. Then the concept of"false alarm"and"miss" in weather forecast is utilized within each precipitation intensity category at all stations. Specifically, the case is marked as"hit"when the precipitation intensity is the same in the reanalyses and observations. The case is marked as "false alarm" when the precipitation intensity in the reanalyses is at least one level higher than that in observations. The case is marked as "miss" when the precipitation intensity in the reanalyses is at least one level lower than that in observations. Then the rate of the three cases within each precipitation intensity category could be acquired and the ability of these reanalyses to reproduce observed precipitation intensity could be assessed.

The spatial distribution of 32-year averaged (1979-2010) monthly precipitation in observations and three reanalyses is presented in Fig. 2 (top panel). The observed precipitation climatology exhibits a wetter pattern (≥140 mm) in the southeast coast and southwest of ZJ, primarily associated with the occurrence of typhoons and the transport of warm and humid air from the southwest, respectively. The observed precipitation also suggests a peak (≥180 mm) in the southernmost part of ZJ and a decreasing trend northward with less than 120 mm over most northern ZJ. The climatological precipitation in three reanalyses all exhibit some discrepancies. Generally, R2 shows the least bias while CFSR exhibits the most prominent wet bias, as seen in the bias percentages (bottom panel in Fig. 2). The mean absolute bias percentages for three reanalyses are 20% (R1), 10% (R2) and 37% (CFSR). To quantify the ability to depict spatial details, pattern correlation coefficients are calculated between climatological precipitation in three reanalyses and in observations, which show 0.54, 0.76 and 0.60 for R1, R2 and CFSR respectively. R1 gives the lowest pattern correlation as it fails to reproduce the wetter features observed at coastal areas while overestimating the precipitation in the majority of the rest areas. R2 is spatially most correlated with observations for it well captures the meridional distribution of precipitation and the precipitation magnitudes over most ZJ areas. As for the CFSR, although it displays the wettest bias, it is spatially more correlated with observations than R1 due to a better depiction of the spatial characteristics of the observed precipitation, such as the wetter features in the southeast coast and southwest and the dryer feature in the central and northeast. Therefore, the CFSR captures the most spatial details of the observed climatological precipitation in spite of having the wettest bias.

Figure 2.  Spatial distribution of 32-year averaged (1979-2010) monthly precipitation (shaded, units: mm) over ZJ in observations, R1, R2, and CFSR (top panel). Bias percentages of precipitation climatology for R1, R2 and CFSR against observations (bottom panel).

The time series of provincial-averaged annual mean precipitation from 1979 to 2010 is presented in Fig. 3. Both R1 and R2 match well with observations with the correlation coefficient (CC) of 0.72 and 0.73, respectively, except for the year 1997 and 1998 in which they both significantly deviate from the observed values, especially R2 (Fig. 3a, c). The CFSR exhibits an improvement over R1 and R2 in a higher CC of 0.84 but shows a prominent systematic wet bias throughout the whole period. The anomalous wet deviation in 1997 and 1998 is also noticed in the CFSR. To explore the impact brought about by these two years, we exclude them in another set of time series (Fig. 3b, d, f). After the exclusion of year 1997 and 1998, the CCs between three reanalyses and observations are raised to 0.81 (R1), 0.81 (R2) and 0.87 (CFSR), strongly supporting the fact that the two years have an impact on attenuating the temporal correlations between the annual mean precipitation in three reanalyses and observations, particularly for R1 and R2.

Figure 3.  Time series of provincial-averaged annual mean precipitation over ZJ province for the period 1979-2010 for R1 (a), R2 (c), and CFSR (e) (red lines, units: mm) and the corresponding observational values (black lines). Same time series (b, d, f) except that the year 1997 and 1998 are excluded.

To further investigate the correlation between reanalyses and observations, the spatial distribution of the CCs at all stations between monthly (Fig. 4a, b, c) and yearly precipitation (12-month mean precipitation; Fig. 4d, e, f) in three reanalyses and that in observations are presented. It is shown that the CFSR displays much higher CC than R1 and R2 do over most areas in ZJ in both cases. For monthly precipitation, the CC reaches 0.8-0.9 for the CFSR over western ZJ where the CCs are only 0.5-0.7 for R1 and R2. Moreover, there are discrepancies in the spatial distribution of the CCs among three reanalyses. Relatively higher CCs (r > 0.7) are distributed over eastern coastal regions (R1), most southern and partial northern areas (R2), and almost entire ZJ (CFSR). For the CFSR, the highest CCs (r > 0.8) appear over the northwest and southwest of ZJ. As for yearly precipitation, the CCs are notably impaired over the southeast coastal areas for all reanalyses, decreasing from 0.6-0.9 to 0.2-0.5 for R1 and R2 and to 0.3-0.5 for the CFSR.

Figure 4.  Spatial distribution of the correlation coefficients (shaded) between monthly (a, b, c) and yearly (d, e, f) precipitation in R1 (a, d), R2 (b, e), CFSR (c, f) and that in observations over ZJ province.

Another important aspect in assessing the quality of reanalyses is the capability to reproduce the principal spatial and temporal modes, which is investigated here via the Empirical Orthogonal Function (EOF) analysis. The EOF analysis is performed on the anomaly of annual-mean precipitation for each dataset. The spatial structures and associated principal components (PCs) of the first three leading EOF modes are compared and discussed as below.

The first EOF mode (EOF1) of observations, explaining 56.7% of the total variance, is characterized by a consistent variability all over ZJ with the largest amplitude in western ZJ (Fig. 5a). The associated PC (PC1) time series displays interannual to decadal fluctuation before 2003 and a rising trend after that (Fig. 5e). All reanalyses agree well with observations in terms of the homogeneous spatial structure of EOF1, except for the inconsistent anomaly center. However, the associated explained variances of all three reanalyses are much larger than observations, accounting for 84.4%, 83.0% and 78.4% of the total variance for R1, R2 and the CFSR, respectively (Fig. 5b, b, d), implying that the variability consistency over ZJ is overestimated for all reanalyses. For the PC1 time series, the observed rising trend after 2003 is well reproduced in the CFSR with a CC of 0.84 which is impaired in R1 (r=0.72) and R2 (r=0.73).

Figure 5.  Spatial patterns (a-d, i-l, q-t, shaded, units: mm) and temporal coefficients (e-h, m-p, u-x, black lines) of the first three EOF modes of precipitation anomaly over ZJ province for observations (the first column), R1 (the second column), R2 (the third column), and the CFSR (the fourth column).

The second EOF mode (EOF2) of observations features a dipole pattern with consistent variability over most northern, western and central ZJ but oppositely consistent variability over the rest areas, explaining for 16.3% of the total variance (Fig. 5i). The corresponding PC (PC2) time series is dominated by decadal variability with negative phases before 2000 and positive phases after that (Fig. 5m). The dipole spatial pattern of EOF2 is well represented by all reanalyses with R2 bearing the most resemblance. The explained variances account for 8.4% (R1), 8.6% (R2) and 11.4% (CFSR) of the total variance, which are underestimated compared to observations (Fig. 5j, k, l). The PC2 time series for all reanalyses fail to reproduce the aforementioned decadal variability (Fig. 5n, o, p) with CCs of 0.41 (R1), 0.51 (R2) and 0.61 (CFSR).

The third EOF mode (EOF3) of observations also displays a dipole pattern where the anomalies over the southwestern ZJ opposes the anomalies over the rest areas, explaining for 6.6% of the total variance (Fig. 5q). The third PC (PC3) time series is dominated by interannual variability throughout the whole period (Fig. 5u). The spatial structures of EOF3 derived from three reanalyses are somewhat different from observations (Fig. 5r, s, t), especially the CFSR which indicates a triple pattern with the same sign over central ZJ but the opposite sign over northern and southern ZJ. The portions of total variance explained by EOF3 are 4.4%, 5.7% and 3.0% for R1, R2, and the CFSR, respectively, which are relatively close to that in observations. As for PC3, both R1 and R2 (Fig. 5v, w) exhibit evident rising trends which disagree with observations while the CFSR (Fig. 5t) exhibits a neutral trend but with different variability compared with observations. The CC of PC3 for CFSR (r=0.43) is substantially higher than that for R1 (r=-0.05) and R2 (r=-0.12).

The root-mean-square error (RMSE) is utilized here as a direct factor in assessing the credibility of the reanalyses. In our case, the RMSE is applied to the time series of monthly precipitation for acquiring the spatial distribution of the temporal RMSE (TRMSE) as well as applied to yearly-mean precipitation based on all stations for acquiring the time series of the spatial RMSE (SRMSE). The top panel in Fig. 6 shows the spatial distribution of TRMSE of monthly precipitation from January 1979 to December 2010 for three reanalyses. R2 gives the most modest TRMSE with 10-30 mm over most ZJ areas and 30-40 mm over east coastal regions, while the CFSR gives the most significant TRMSE with > 30 mm over most areas and the extreme of > 80 mm, in concert with the aforementioned obvious systematic wet bias in CFSR. The TRMSE for R1 falls somewhere in the middle with 10-50 mm over most ZJ areas. The lower panel in Fig. 6 presents the time series of the SRMSE of yearly-mean precipitation for three reanalyses. In general, R2 and CFSR exhibit the slightest and the largest SRMSE, respectively, with R1 in between. After 2001, the SRMSE for CFSR suddenly decreases and keeps smaller than R1 and R2, possibly associated with the enhanced quality of CFSR due to increasing observations assimilated in the CFSR(Saha et al. ^[7]; Xue et al. ^[37]).

Figure 6.  Spatial distribution of the TRMSE (top panel, shaded) over ZJ province of monthly precipitation calculated from January 1979 to December 2010 and time series (lower panel) of SRMSE from 1979 to 2010 of yearly-mean precipitation calculated from all stations for R1, R2, and CFSR (units: mm).

The time series of seasonal-mean SRMSE of precipitation in each season (Fig. 7a-c) indicate that for all reanalyses, the SRMSE is generally the smallest in winter and the largest in summer, and the SRMSE in spring and autumn are rather close to each other while only slightly larger than that in winter. The differences of SRMSE are small among all reanalyses in all seasons except summer when the differences get amplified. In summer, the number of years when the SRMSE surpasses 150 mm is ten for the CFSR but only six for R1 and three for R2, and there are two years (1989 and 1998) when the SRMSE even surpasses 200 mm for the CFSR but none for R1 and only one year (1998) for R2. In other words, the summer deviation of CFSR from observations is much larger than those of R1 and R2 and is the major cause of the significant wet bias in the CFSR. Looking further into the SRMSE in each month, it is noted that the maxima of SRMSEs for R1 and R2 occur in August with the second largest SRMSE in July while for teh CFSR the maximum of SRMSE occurs in July followed by June. Thus, the larger SRMSE in June for CFSR is the main factor contributing to the larger bias in summer compared with R1 and R2.

Figure 7.  Time series of seasonal-mean precipitation SRMSE in spring (green line), summer (red line), autumn (orange line), and winter (blue line) during 1979-2010 for R1 (a), R2 (b), and CFSR (c). Multiyear-mean precipitation SRMSE (black line) in each month during the same time-span for R1 (d), R2 (e), and CFSR (f) (units: mm).

Until now we have investigated the relative quality of three reanalyses, but to what extent these reanalyses depart from observations remains unknown. To elucidate this question, we divided the biases of monthly precipitation into three categories as introduced in section 2.2: positive bias, negative bias and neutral bias. Then we computed the percentage of each case based on all stations and acquired the annual time series as shown in Fig. 8. The result implies that the cases of neutral bias make up similar percentages (around 50% - 60%) for all reanalyses. The case of positive bias for the CFSR makes up much higher percentage (about 40%) than that in R1 and R2 (about 30% for both), and the case of negative bias for the CFSR only accounts for less than 10% compared with 10% - 20% for R1 and R2, which works as an important cause for the significant wet bias occurring in the CFSR. After 2001, the cases of positive bias accounts for less percentages for all reanalyses while the cases of negative bias accounts for more percentages for R1 and R2 and less percentage for the CFSR, consequently leading to slightly higher percentages of neutral bias cases for the CFSR.

Figure 8.  Time series of annual percentages of the cases of positive bias, positive bias, and neutral bias at all stations of ZJ province during 1979-2010 for R1, R2, and CFSR.

Using five precipitation categories and the concepts of "false alarm", "miss" and "hit" as introduced in section 2.2, the deviations of three reanalyses from observations could be quantitatively computed on a daily scale. Fig. 9 presents the time series of annual false alarm rate, miss rate and hit rate of five precipitation intensities (light, modest, heavy, rainstorm and heavy rainstorm) assessed at all stations. For light rain, all reanalyses well capture observations with hit rates of > 80%, false alarm rates of < 20% and zero miss rates (which may result from none-zero values at all stations due to interpolation). For moderate rain, CFSR gives a higher hit rate of around 40% than R1 and R2 (both are around 20% - 40%), which may contribute to a lower miss rate (40%-60%) for the CFSR than R1 and R2 (> 60%) in spite of a slightly higher false alarm rate for CFSR. For heavy rain, the CFSR is still superior to R1 and R2 for having a higher hit rate of around 20% followed by R2 and R1, as a result of a lower miss rate of < 80% than R1 (> 80%) and R2 (> 90%). When it comes to the categories of rainstorm and heavy rainstorm, both R1 and R2 almost lose the ability to capture the reality with hit rates of 0% - 10% for rainstorm and around zero for heavy rainstorm, whereas the CFSR still exhibits a hit rate of 10%-20% for rainstorm and > 5% in most of the years, sometimes reaching 10% - 20%, for heavy rainstorm. All three reanalyses tend to overestimate light rain and underestimate heavier precipitation, which is similar to what Zhou et al. ^[43] have concluded. In summary, three reanalyses give similar performances in reproducing light rain but the CFSR performs better in terms of capturing the precipitation intensities of moderate rain, heavy rain, rainstorm, and heavy rainstorm. Overall, the CFSR has the highest quality in portraying daily precipitation. Besides, it is interestingly noted that the significant wet bias in the CFSR is instead conducive to its capability to reproduce large amount of precipitation, which is difficult for R1 and R2.

Figure 9.  Time series of station -averaged annual false alarm rate, miss rate, and hit rate (black lines) for light rain, moderate rain, heavy rain, rainstorm, and heavy rainstorm during 1979-2010 over ZJ province for R1, R2, and CFSR, with grey error bars denoting the maxima and minima of the rates.

Assessment of reanalyses is an essential prerequisite for appropriate use of the data in climate studies. In spite of substantial assessment performed on various reanalyses since 1990s, limited attentions were paid on the performances of different generations of reanalyses. Although the third-generation reanalyses have been an improvement over older-generation reanalyses in terms of profoundly enhanced horizontal and vertical resolutions, it should not be assumed that the third-generation reanalyses are superior to older reanalyses in all aspects. Comprehensive examination of different generations of reanalyses is needed for an understanding of their detailed advantages and deficiencies. Moreover, most of the previous assessment work was based on large-scale regions, so the conclusion, especially for variables like precipitation, which features a strong dispersed pattern and is regionally dependent, could hardly be applied to small-scale regions, which is exactly the attention of localized meteorological service. This study systematically evaluates multi-scale performances of R1, R2 and the CFSR, which belong to different generations, in reproducing precipitation over ZJ province for the period of 1979-2010. The following conclusions are reached.

1) The spatial pattern of climatological precipitation derived from all three reanalyses exhibit discrepancies compared with observations. The mean absolute bias percentages for three reanalyses are 0.20 (R1), 0.10 (R2) and 0.37 (CFSR). The pattern correlation coefficients between three reanalyses and observations are 0.54 (R1), 0.76 (R2) and 0.60 (CFSR). Overall, R2 (R1) gives the best (worst) depiction of the spatial characteristics of the observed precipitation climatology. Although significant wet bias is present in the CFSR, it captures the most spatial details of the observed climatological precipitation.

2) The time series of annual provincial-mean precipitation derived from three reanalyses agree well with observations with correlation coefficients of 0.72 (R1), 0.72 (R2) and 0.84 (CFSR). However, prominent wet biases are present in the years of 1997 and 1998 in all three reanalyses. The CFSR presents a persistent wet bias throughout the whole stage while being best correlated with the observed time series.

3) The EOF analysis implies that all reanalyses could generally reproduce the spatial pattern of the first two leading EOF modes of the annual mean precipitation, while evident deviations occur in their third modes. The CFSR gives the most relevant PC time series in all modes and hence has the best capability to portray spatiotemporal characteristics of the observed precipitation.

4) For both TRMSE and SRMSE, the CFSR and R2 present the largest and the smallest values, respectively, with R1 in between. After 2001, however, the SRMSE of CFSR decreases more distinctly than those of R1 and R2. In terms of the seasonal cycle, the SRMSEs in summer are far over the rest of seasons for all reanalyses, with the CFSR exhibiting significantly higher SRMSE than R1 and R2. By further examining the monthly variation of annual mean SRMSE, it is found that the CFSR surpasses R1 and R2 to the most extent in June, and thus the SRMSE in June makes the greatest contribution to the large deviation of CFSR in summer, which eventually leads to the prominent systematic wet bias in the CFSR.

5) On a monthly basis, the neutrally biased cases make up similar percentages of the total cases for all reanalyses. The positively biased (negatively biased) cases for the CFSR accounts for percentage higher (lower) than R1 and R2, respectively, which works as an important factor for the significant wet bias in the CFSR. After 2001, the negatively-biased cases make up larger percentages for R1 and R2 and smaller percentage for the CFSR, consequently leading to slightly higher percentage of neutrally-biased cases for the CFSR.

6) On a daily basis, all reanalyses give similar performances in capturing light rain, while the CFSR performs better in terms of reflecting the precipitation intensities of moderate rain, heavy rain, rainstorm, heavy rainstorm. In general, the CFSR has the highest quality in portraying daily precipitation. The significant wet bias in the CFSR is instead conducive to its better capability to reproduce large amount of precipitation.

Overall, despite presenting many superiorities over R1 and R2 in our assessment, the CFSR still exhibits several deficiencies such as the remarkable systematic wet bias over the majority of ZJ domain. Hence the prior conclusion should not be drawn without doubt that newer-generation reanalyses must be perfectly better products than older-generation reanalyses. Our analysis indicates that the CFSR does not perform better than R1 and R2 in all aspects over ZJ province. Additionally, it is worth noting that the systematic wet bias in the CFSR evidently weakens after 2001, indicating a more reliable period of 2001-2010 for the CFSR. It is important to realize that the credibility of reanalyses may vary with the subject of research, the spatial scale as well as the temporal scale, and thus the choice of reanalyses is worth great consideration before practical use.