Precipitation over Monsoon Asia: A Comparison of Reanalyses and Observations

Reanalysisproductsrepresentavaluable sourceofinformationfordifferent impactmodelingandmonitoring activities over regions with sparse observational data. It is therefore essential to evaluate their behavior and their intrinsic uncertainties. This study focuses on precipitation over monsoon Asia, a key agricultural region of the world. Four reanalysis datasets are evaluated, namely ERA-Interim, ERA-Interim/Land, AgMERRA (an agricultural version of MERRA), and JRA-55. APHRODITE and the Climate Hazards Group Infrared Precipitation with Stations (CHIRPS) dataset are the two gridded observational datasets used for the evaluation; the former is based on rain gauge data and the latter on a combination of satellite and rain gauge data. Differences in seasonality, moderate-to-heavy precipitation events, daily distribution, and drought characteristics are analyzed. Results show remarkable differences between the APHRODITE and CHIRPS observational datasets as well as between these datasets and the reanalyses. AgMERRA generally achieves the best performance, but it is not updated at near–real time. ERA-Interim/Land shows good spatial performance, but when the interest is on the temporal evolution JRA-55 is recommended, as it exhibits the most stable temporal behavior. This study shows that the use of reanalyses for impact modeling and monitoring over monsoon Asia requires an accurate evaluation and choices to be tailored to the speciﬁc needs.


Introduction
Impact modeling, such as agricultural modeling and monitoring activities, is heavily based on the availability of daily meteorological variables, such as precipitation and mean temperature.In regions with sparse meteorological stations and/or with restricted data accessibility and availability, reanalysis products offer a convenient way to work.It is therefore essential to evaluate their behavior as well as their intrinsic uncertainties.Here, we focus on precipitation in the core monsoon region of Asia (Noska and Misra 2016;Ding and Chan 2005), a key agricultural region of the world.This region is characterized by high population and intensive agricultural land use.Precipitation variability within the monsoon season can cause adverse impacts on agricultural yields (Singh et al. 2014, and references therein).For example, prolonged dry spells during the Indian summer monsoon season can substantially reduce yields of Kharif (monsoon) crops when these events occur during soil preparation, transplanting, or at critical crop growth phases.The Indian summer monsoon is responsible for 85% of the annual precipitation (Singh et al. 2014) and it is vital for Indian agriculture.Sensitivity of agricultural production to rainfall dynamics is generally higher over rain-fed areas (e.g., Ceglar et al. 2016).Although irrigation is intensively practiced in monsoonal Asia (Hatcho et al. 2010), irrigation water availability still depends on precipitation and could become an issue in a changing climate context (e.g., Elliott et al. 2014).Monitoring droughts during the monsoon season is therefore necessary to manage natural resources and it has high importance for policy makers (Shah and Mishra 2014).
Several evaluation studies have been performed on monsoon precipitation by using different observational and reanalysis datasets.Recently, Shah and Mishra (2014) reported positive biases in the Indian monsoon precipitation affecting the ECMWF interim reanalysis (ERA-Interim, ERA-I; Dee et al. 2011), the NASA reanalysis MERRA (Rienecker et al. 2011), and the NCEP reanalysis CFSR (Saha et al. 2010).Huang et al. (2016) reported overall good performance of reanalyses in reproducing the climatology and the variability of the East Asia summer monsoon precipitation.Rana et al. (2015) analyzed reanalyses (CFSR and ERA-I) and rain gauge-and satellite-based datasets; reanalyses reproduced the seasonality in the Indian subcontinent well, especially in low-elevation areas, while more uncertainties remained in high-elevation areas and areas with poor weather station coverage.It is also worthwhile highlighting that, as precipitation is a diagnostic variable of reanalyses, model physics and simulated circulation play an important role (Bosilovich et al. 2008).
As for agricultural monitoring and modeling, precipitation characteristics at different time scales (from daily to seasonal) are relevant and must be evaluated.As discussed by Huang et al. (2016), there are large uncertainties in nonrainfall and heavy rainfall classes among different reanalyses in eastern Asia.Therefore, this study aims at evaluating four latest-generation reanalysis datasets (some of them offering bias-corrected precipitation data) in terms of various precipitation characteristics at different time scales.Seasonality, distributional characteristics of daily precipitation, precipitation extremes, and drought during the monsoon season are analyzed by using recently proposed methods and approaches.

Data and methods
Four reanalysis datasets and two observational gridded datasets (one rain gauge based and another one satellite and rain gauge based) have been used for this study (Table 1).These datasets have been selected due to their spatial resolution and their spatiotemporal coverage.The analyzed area extends over 08-408N, 658-1258E, covering the core Asian monsoon areas, except the northeastern part of the region affected by the East Asian summer monsoon (Ding and Chan 2005).To perform a quantitative comparison, all datasets have been regridded on the same 0.758 reference grid (i.e., the one of ERA-I) by applying a conservative remapping procedure (Chen and Knutson 2008, and references therein).Among the chosen reanalyses, ERA-Interim/Land (the land version of ERA-I, hereafter ERA-I-Land; Balsamo et al. 2015) and AgMERRA (the agricultural version of MERRA; Ruane et al. 2015) provide bias-corrected precipitation.ERA-I-Land is a single 32-yr simulation with the latest ECMWF land surface model driven by meteorological forcing from ERA-I (Dee et al. 2011).It is characterized by a monthly bias adjustment of precipitation based on the Global Precipitation Climatology Project (GPCP) monthly precipitation dataset (Adler et al. 2003).AgMERRA (Ruane et al. 2015) is a NASA reanalysis based on MERRA and MERRA-Land reanalysis (Rienecker et al. 2011;Reichle et al. 2011).It includes a monthly bias adjustment of precipitation based on the Climate Research Unit (CRU) wet days dataset (Harris et al. 2014), the Global Precipitation Climatology Centre (GPCC) dataset (Schneider et al. 2014), the University of Delaware precipitation dataset (WM; Willmott and Matsuura 1995), and satellite-based products [TRMM (Huffman et al. 2007), CMORPH (Joyce et al. 2004) and PERSIANN (Hsu and Gao 1997)].The other two reanalyses, ERA-I (Dee et al. 2011) and JRA-55 (Kobayashi et al. 2015), are not bias-corrected.Concerning the gridded As the temporal availability of the datasets is not the same, the common period from 1981 to 2007 has been selected for the analysis.Moreover, the period from June to August (JJA) has been chosen as the monsoon period.The onset of the monsoon depends on both the year and the region, but JJA can be thought as a conservative choice to avoid the complexity and the uncertainties of dealing with the spatiotemporal nonstationarity.

a. Seasonality
There exists a wide range of seasonality metrics that were developed for specific purposes and locations, with the aim of assessing magnitude, timing, and duration of wet/dry seasons (Feng et al. 2013;Zhang and Qian 2003).To evaluate the seasonality of precipitation in monsoonal Asia and compare its representation in the different datasets under investigation, a recently proposed approach combining relative entropy and the dimensionless seasonality index (Pascale et al. 2015, and references therein) is here applied.The relative entropy D is defined as where p m (x) is the monthly precipitation ratio with respect to (hereafter w.r.t.) the annual total precipitation at location x.The relative entropy measures the degree of similarity to a uniformly distributed precipitation throughout the year.Thus, it reaches its maximum value when precipitation is concentrated in one single month, while it is equal to 0 when there is perfect uniform distribution in the 12 months.The dimensionless seasonality index S is defined as where R(x) is the total annual precipitation at location x and R 0 is a constant, usually chosen as the maximum R(x) over the whole spatiotemporal domain, to make the index dimensionless.This index combines information on seasonality and total accumulated precipitation.It is zero when either the location is dry or there is a perfectly uniformly distributed monthly precipitation during the year, and reaches its maximum when precipitation occurs in a single month and a region has high accumulated precipitation.On the contrary, when high accumulated precipitation is well spread out throughout the year, the index is lower due to the low D(x).The advantages of both indices are discussed in detail in Pascale et al. (2015).Because of the abovementioned attributes, the dimensionless seasonality index is very useful to identify and compare monsoonal regions (Wang and Ding 2008).

b. Distributional analysis
Concerning precipitation during JJA, the different datasets are evaluated first by looking at the seasonal accumulated values and then at the daily precipitation distribution.Concerning the latter, a recently proposed nonparametric approach (Toreti and Naveau 2015), particularly suitable when the right tail of the distribution (i.e., the extremes) is of interest, is applied to compare (with a special focus on extremes) the daily precipitation distribution functions of reanalyses and observations.The approach is based on a two-sample modified Anderson-Darling statistic, A, combined with the Kullback-Leibler direct divergence, I. Let X and Y be an n-and m-sample, respectively (here, the daily precipitation during JJA coming from observations and reanalysis), with survival empirical distribution functions F and G (i.e., 1 minus the empirical cumulative distribution function) of X and Y (and associated probability density functions f and g), respectively.Then, where H denotes a measure defined by the weighted average of F and G, N 5 n 1 m, and E f denotes the expectation under the probability density function f.Also, A measures how well the distribution of daily precipitation from reanalysis reproduces the distribution of daily observed precipitation, especially with respect to moderate-to-heavy precipitation.Smaller values of this statistic indicate a similarity between the two distributions of daily precipitation during JJA.The statistical significance of these differences between the distributions is assessed by using a test based on where E(A) denotes the mean of A and s is its standard deviation.The null hypothesis of equal distributions is rejected when T is greater than the approximated critical value at the 95% significance level.The Kullback-Leibler divergence I is then used to assess the sign (i.e., the direction of change), since A only provides the absolute difference of two distributions.For more details the reader is referred to Toreti and Naveau (2015).
Here, the approach is applied to the rescaled daily precipitation, obtained by dividing all values by the mean intensity of precipitation on the wet days during JJA.Spatial significance is evaluated, as proposed by Toreti and Naveau (2015), by using the approach of Genovese and Wasserman (2004).

c. Precipitation events
To provide a complementary evaluation in terms of moderate-to-heavy precipitation events, a forecastverification approach is applied.The fractions skill score (FSS; Roberts and Lean 2008) is used.The FSS is based on transforming precipitation data into binary values (1 or 0) by using thresholds (here the 75th and the 90th percentiles); 1 is assigned to those grid points where precipitation exceeds the given threshold.Then, for each grid point and its fixed neighborhood the fractions of events (i.e., the number of 1's) is counted, both for reanalysis and observations.Thus, the FSS can be derived by where the FBS is the mean square difference between the observed and the forecast precipitation fractions, defined as with O i and F i representing, respectively, the observed (either APHRODITE or CHIRPS) and the reanalysis fractions at each grid point and N is the number of grids in the domain.The term FBS worst represents the largest FBS that can be obtained; the reader is referred to Roberts and Lean (2008) for more details.FSS ranges between 0 (no skill) and 1 (perfect reanalysis).Here, the FSS is computed for different neighborhood sizes: from the model grid size (0.758) up to 12.758.Then, the smallest neighborhood size, L, at which a reanalysis provides useful information is derived.L is such that the FSS at scales equal or greater than L is greater or equal to 0:5 1 f /2, where f is the observed fractional precipitation coverage over the domain (wet-area ratio; Roberts and Lean 2008).

d. Drought
To characterize droughts, the Standardized Precipitation Index (SPI; Mckee et al. 1993) is here applied.SPI is a dimensionless index based on the probability distribution of precipitation.It can be generated for different time scales, such as monthly and seasonal.Drought events are categorized as follows: extreme (SPI less than 22), severe (SPI between 22 and 21.5), and moderate (SPI between 21.5 and 21.0).For the assessment of monsoon season droughts, the summer season SPI is used here (i.e., SPI on three months' time scale for August).
The area affected by drought is estimated as the proportion of grid cells over land affected by drought.Taylor diagrams (Taylor 2001) are used to assess the performance of different reanalyses to reproduce the spatial patterns of drought for each specific year in the analyzed period.Thus, in the Taylor diagrams, the reference dataset (either APHRODITE or CHIRPS) is a function of the year.In addition, the Taylor skill score S is calculated (Hirota et al. 2011;Taylor 2001): where R represents the spatial pattern correlation between the reanalysis and the observations and SDR represents the ratio of spatial standard deviations between reanalysis and observations.This score therefore quantifies the similarity of the distribution and amplitude of the spatial patterns between reanalysis and observations.

a. Seasonality
As observational gridded datasets also suffer from uncertainties and can fail in reproducing precipitation field over areas with poor data coverage, it is worthwhile to compare APHRODITE and CHIRPS.In terms of seasonality, APHRODITE shows a more uniformly distributed precipitation through the year in western China w.r.t.CHIRPS (Fig. 1).In northwestern India/Pakistan, CHIRPS shows a broader area with precipitation being concentrated in a few months (Fig. 1).Concerning reanalyses, all of them show more uniformly distributed monthly precipitations when compared to CHIRPS (Fig. 2).However, differences can be observed in the dimensionless seasonality index S (Fig. 3) as, for instance, a broader and more pronounced monsoonal behavior in Myanmar in ERA-I (Fig. 3).When compared to APHRODITE, JRA-55 and ERA-I overall show a similar spatial pattern in terms of seasonality with precipitation being more uniformly distributed during the year, except in some areas such as northwestern India/Pakistan and western China (Fig. 2).However, ERA-I is characterized by a broader and more pronounced monsoonal behavior in the northeastern side of the Bay of Bengal and along the southern side of the Himalaya, while JRA-55 overall shows a less pronounced monsoonal behavior (Fig. 3).Both AgMERRA and ERA-I-Land show less pronounced differences (Fig. 2), although they both have higher monsoonal activity over Myanmar.

b. JJA differences
During JJA, CHIRPS shows higher accumulated precipitation (especially in the area under the influence of both the Indian summer monsoon and the west Pacific summer monsoon) except over the northern part of the domain and along the western coast of India (Fig. 4).This behavior is almost stationary during the investigated period and differences are always more pronounced in the zonal band comprised between 108 and 158N (Fig. 4).As for reanalyses, AgMERRA is generally characterized by the smallest differences in summer mean precipitation cumulates, except over Myanmar and along the southeastern side of the Tibetan Plateau where it significantly overestimates JJA precipitation (w.r.t.APHRODITE; Fig. 5).Large significant positive biases affect ERA-I especially in the eastern part of the domain (Fig. 5) whereas these biases are considerably reduced in ERA-I-Land, especially over southeastern China.JRA-55 has a completely different behavior w.r.t. the other reanalyses, being largely affected by significant underestimation of JJA accumulated precipitation (Fig. 5).These differences contribute to explain and better understand the identified behavior in the dimensionless seasonality index discussed in the previous subsection.In terms of spatial correlation of JJA precipitation, AgMERRA achieves the best performance, being characterized by correlation values higher than 0.9 with both APHRODITE and CHIRPS (Fig. 6).JRA-55 and ERA-I-Land have similar performance (with correlation values varying around 0.8), but ERA-I-Land is affected by a sudden drop in 1990.ERA-I is the worst in terms of spatial correlation with values always lower than 0.8 (Fig. 6).Concerning the temporal evolution of the identified JJA-biases, it is worthwhile to highlight the sudden decrease in the positive bias affecting ERA-I (w.r.t.APHRODITE) in the mid-1990s over the zonal band comprised between 108 and 158N (Fig. 7) as well as the changes affecting ERA-I-Land.The other two reanalyses do not show remarkable temporal changes (Fig. 7).
Besides the identified difference in the JJA precipitation of the two observational datasets, the (rescaled) distributions of APHRODITE and CHIRPS are still significantly different (although the differences are not too large), with the former being characterized by fatter tails (not shown).Concerning the reanalyses, they all show more complex differences w.r.t.observations than a simple bias in the mean.Indeed, they all tend to underestimate the (rescaled) distributions and ERA-I shows the largest differences (except over the Tibetan Plateau; Fig. 8).

c. Spatial and temporal variation in the skill of daily precipitation from reanalyses
Figure 9 shows the FSS as a function of the spatial scale for the two selected thresholds.The skill increases with increasing size of spatial scale, since rainfall events are more accurately captured over larger areas and models often tend to misplace localized intense rainfall events (Roberts and Lean 2008).Moreover, higher skill can be observed for lower precipitation thresholds (here, 75th percentile), including less intense and spatially broader events that are more accurately captured by reanalyses.Higher skill at the model grid scale can be observed for AgMERRA; however, for the 90th percentile the skill is generally below 0.5.With respect to APHRODITE, the FSS becomes comparable between ERA-I, JRA-55, and AgMERRA at spatial scales higher than 2.258.When using CHIRPS as the reference dataset, AgMERRA and JRA-55 show similar skill throughout all spatial scales, whereas slightly lower performance is identified for ERA-I.The skill of ERA-I-Land is below the skill of all the other reanalyses, most probably due to sudden change after 1990, as described in the previous subsection.It should be emphasized that before that year the skill of ERA-I-Land is higher than the ones estimated for both ERA-I and JRA-55, especially when considering the 75th percentile threshold.
Results suggest that skillful spatial scale is temporally varying.A scale of 2.258 appears to be a skillful scale for AgMERRA using the 75th percentile threshold.For ERA-I and JRA-55, the skillful spatial scale varies between 2.258 between 1990 and 2000 and 3.758 during the rest of the analyzed period.Considering the 90th percentile, the skillful spatial scale of 3.758 prevails in ERA-I, JRA-55, and AgMERRA; ERA-I-Land has a comparable skillful spatial scale before 1989, but it drops substantially after that year.

d. Drought
In terms of interannual drought variability, APHRODITE and CHIRPS tend to agree except over Myanmar, Cambodia, and the northwestern part of the domain (not shown).Consistency between drought-affected areas can be found for events such as the one in 1987 (northwestern India, Pakistan, and the Tibetan Plateau), 1997 (central India, central-western part of China, and parts of southeastern Asia), and 2002 (northwestern India, Pakistan, central China, and the North China Plain) (e.g., Zou et al. 2005;Wang 2006).The drought events shown by APHRODITE in 1999 Figure 10 shows the time series of the area affected by extreme, severe, and moderate drought (as defined in the previous section).Overall, drought-affected areas are well represented by all the reanalyses.However, higher differences from observations after 1990 can be observed in the time series of ERA-I-Land (especially for moderate events).
Taylor diagrams shown in Fig. 11 represent the estimated differences in the spatial patterns of drought events.AgMERRA clearly outperforms the other reanalyses.Its spatial correlation with the observational datasets generally exceeds 0.8.Lower performance can be seen in Fig. 11 associated with ERA-I and JRA-55, which are characterized by similar behavior, while worse spatial representation is observed for ERA-I-Land.In this case, two well-separated groups of years can be identified.Spatial patterns tend to be better represented before 1990 (when ERA-I-Land outperforms ERA-I).Afterward, the spatial representativeness of drought patterns tends to be poor, with nonsignificant correlations and RMS, when compared to both observational datasets.The Taylor skill score S exhibits high interannual variability (not shown) but generally highly resembles the spatial correlation patterns of Fig. 11.A substantial drop of the skill score characterizes ERA-I-Land after 1990.ERA-I and JRA reach lower skill than AgMERRA, varying respectively between 0.2 and 0.4 and between 0.4 and 0.7.The skill score of ERA-I-Land drops from 0.4 to 0.1 after 1990.Generally, the skill drops in AgMERRA, ERA-I, and JRA after 1999.
To better understand and characterize the different drought representation, a specific analysis of severe drought events over selected regions of monsoon Asia is performed.Figure 12 shows areal averages of SPI over central India (region 1), Southeast Asia (region 2), and the North China Plain (region 3).Severe drought events occurred during the analyzed period in central India in 1987 and 2002 according to the observational datasets.The 1987 drought event is well represented by all the reanalysis datasets.On the contrary, the 2002 event is underestimated by ERA-I and ERA-I-Land.Reanalyses and CHIRPS also point to a drought event in 1992, the drought signal is not well reproduced after 2000.The most intensive drought events after 1990 (1997 and 2002) are well reproduced by AgMERRA and JRA-55, whereas ERA-I and ERA-I-Land fail to capture those events.

Discussion and conclusions
The comparison between the two observational precipitation datasets has shown remarkable differences (e.g., in the represented seasonality of precipitation as well as in the distribution of JJA daily precipitation).This highlights the uncertainties affecting all gridded observational datasets associated both with the methodological interpolation procedures and data spatiotemporal homogeneity and availability.Significant differences have been also identified in the comparison of all the reanalyses with the two observational dataset.An interesting sudden change in the 1990s seems to affect ERA-I in a specific zonal band (between 108 and 158N) and ERA-I-Land (affected by another change at the beginning of the twenty-first century).In terms of droughts, ERA-I-Land fails to reproduce drought occurrence patterns after 1990, although its performance is higher than JRA-55 and ERA-I before that year.In terms of spatial correlation of JJA-accumulated precipitation, a drop is clearly affecting ERA-I-Land in 1990 (Fig. 6).However, the spatial correlation of seasonal precipitation of ERA-I-Land with the observational datasets remains still higher than the one observed between ERA-I and observations.This sudden change can be attributed to a change in the ERA-I assimilation rain-affected radiances from SSM/I (Geer et al. 2008), but also the observed interdecadal shift that occurred over the region at the beginning of the 1990s (Huang et al. 2016, and references therein) could have played a role.Given that the spatial correlation between observations and ERA-I remains stable, the shifts in JJA precipitation might have been amplified by modified land modeling component in ERA-I-Land.This shift in precipitation translates into an amplified drought shift.Indeed, Balsamo et al. (2015) have shown that the changes in various simulated components of the hydrological cycle over Asia are mainly a result of model changes, whereas the bias correction in precipitation has been shown to be rather neutral over Eurasia.
Overall, AgMERRA outperforms the other three reanalyses.This performance of AgMERRA is somewhat expected due to the high level of statistical postprocessing (involving several different observational datasets) of the MERRA reanalysis.
This study has shown that the use of reanalyses for impact modeling and monitoring over monsoonal Asia should be based on an accurate evaluation process and choices should be tailored to the specific needs.Indeed, AgMERRA achieves the best performance, but it is not a near-real time (or almost near real time) updated dataset.ERA-I-Land shows good spatial performance as well, but when the interest is on the temporal evolution (such as drought analysis) JRA-55 should be preferred, as it exhibits the most stable temporal behavior.

FIG. 1 .
FIG. 1. Relative entropy (DX) of (a) APHRODITE and (b) CHIRPS and (c) their difference, and the dimensionless seasonality index (DSI) of (d) APHRODITE and (e) CHIRPS and (f) their difference.

FIG. 2 .
FIG. 2. Difference in relative entropy between the four different reanalyses and (top) APHRODITE and (bottom) CHIRPS observational datasets.

FIG. 5 .
FIG. 5. Differences in the mean summer accumulated precipitation of four different reanalyses and (top) APHRODITE and (bottom)CHIRPS observational datasets.Regions where differences are statistically significant ( p , 0.05) are denoted with white crosses.

FIG. 7 .
FIG. 7. Mean zonal differences in the summer accumulated precipitation of the four reanalyses w.r.t.(top) APHRODITE and (bottom) CHIRPS.Values (in mm) are presented for period between 1981 and 2007.

FIG. 9 .
FIG. 9.The distribution of FSS as a function of spatial scale for (left) 75th and (right) 90th percentiles chosen as precipitation thresholds.The box plot at each of the spatial scales represents the temporal distribution of FSS (i.e., the skill is assessed for each day of all summer seasons).