1. Introduction
Though high-quality precipitation information is critical for understanding water budget variability over regional and global extents, precipitation datasets have high uncertainties, especially in complex environments. In such areas, ground measurements are difficult to obtain, and stations are burdensome to maintain and monitor. As a result, the spatial extent and coverage and temporal consistency of in situ measurement networks are limited. Also, gauge-based estimates lack the ability to capture the strong heterogeneities of precipitation and suffer from errors and uncertainties. Gridded precipitation products derived from reanalysis and/or satellite observations are another source of precipitation information that can provide continuous long-term estimates in both time and space (Hersbach et al. 2020; Molteni et al. 1996; Rienecker et al. 2011; Yatagai et al. 2012). Therefore, they constitute the only practical precipitation products that could be used in modeling efforts in remote regions. Despite the existing numerous gridded precipitation products (Funk et al. 2015; Hersbach et al. 2020; Huffman et al. 2015; Rienecker et al. 2011; Yatagai et al. 2012), precipitation remains one of the most difficult forcing to estimate accurately. In particular, lack of reliable observations of atmospheric moisture and vertical motion makes precipitation estimation particularly challenging over complex mountainous terrains (Ebert 2001).
High Mountain Asia (HMA), a high-elevation geographical area, includes the mountain ranges surrounding the Tibetan Plateau (Fig. 1). HMA is a complex region due to the myriad of hydrologic processes that control the terrestrial water budget including cryospheric sources of water (snow, glacier, and permafrost), monsoon dynamics, and anthropogenic activities such as irrigation and pumping. HMA basins, which include the Ganges–Brahmaputra, the Indus, and the Yangtze, provide the water supply for around a billion people (Pritchard 2019; Viviroli et al. 2007). The complex topography of HMA leads to large uncertainties in meteorological forcing (Müller Schmied et al. 2016; Yoon et al. 2019). In HMA, gridded precipitation products provide different intensities and trends, making hydrologic analyses difficult despite their importance in understanding the evolution of water resources in such populated region (Palazzi et al. 2013; You et al. 2015). Many studies have assessed the performance of these gridded precipitation datasets in an attempt to identify the most accurate ones in the region (Andermann et al. 2011; Tong et al. 2014; Yoon et al. 2019; Hong et al. 2021). Nevertheless, because ground measurements are very scarce, there is still a lack of an agreed-upon accurate precipitation dataset.
Extent of the High Mountain Asia (HMA) domain with the elevation variations as the background. The black lines represent the boundaries of hyrologic basins whose names are indicated in black within each basin.
Citation: Journal of Hydrometeorology 23, 9; 10.1175/JHM-D-21-0196.1
Considering an ensemble of estimates is often used to overcome the inconsistencies and differences of individual precipitation products (Clark et al. 2009; Ebert 2001; Newman et al. 2015; Schwartz et al. 2017; Surcel et al. 2014). The main advantage of the ensemble is that the averaging procedure filters out features that individual ensemble members fail to agree on (Surcel et al. 2014; Warner 2010). Comparisons against observations have shown that ensemble precipitation is generally more skillful than any of the constituent ensemble members (Buizza and Palmer 1998; Clark et al. 2009; Ebert 2001; Weusthoff et al. 2011).
Among the different ensemble techniques, the ensemble mean (EM) based on the arithmetic average of gridded precipitation fields is widely used. However, the averaging process tends to overpredict light precipitation, underpredict heavy precipitation, and create artificially smooth features. The probability matched mean (PM) proposed by Ebert (2001) recalibrates the EM field to capture the extreme values and to provide a more realistic precipitation field with greater accuracy than the EM. The resampling process of the PM discards the outliers and singularities and gives more weight to regions with strong similarities among the ensemble members. Though probability matching methods were found to be superior to EM (Qiao et al. 2020; Snook et al. 2020), PM could also lead to unrealistic precipitation fields because the recalibration procedure incorporates spatially distant information. Moreover, the performance of the PM could be significantly reduced in regions with strong changes in topography. Clark (2017) proposed the localized PM mean (LPM) approach restricting the use of locations within a certain distance of each grid point to contribute to the reassignment.
In this study, we describe the development of a consensus ensemble precipitation product across HMA using the LPM method with three gridded precipitation datasets: the ECMWF’s fifth generation of atmospheric reanalysis of the global climate (ERA5; Hersbach et al. 2020), the Final product of the Integrated Multi-satellitE Retrievals for Global Precipitation Measurement (GPM) (IMERG; Huffman et al. 2015), and the Climate Hazards Group Infrared Precipitation with Station data (CHIRPS; Funk et al. 2015). As noted above, the individual precipitation products over HMA have large differences in yearly averages and trends (Andermann et al. 2011; Ma et al. 2009; Müller Schmied et al. 2016; Song et al. 2016; Yoon et al. 2019; You et al. 2015). The primary contribution of this paper is the description of the development of a dataset with realistic trends and yearly averages by reducing the discrepancies across individual precipitation products. The manuscript provides evaluations of the LPM-based product using both direct comparisons with limited in situ datasets and strategies such as extended triple collocation. The advantage of the LPM product is highlighted by examining the consistency of long-term trends and averages and by comparing it against an equivalent EM product. The development of the consensus dataset is expected to be a critical input modeling studies aimed at reconstructing and understanding the hydrology of the region over the past decades.
2. Study area: High Mountain Asia
Our study area ranges from 20° to 46°N and from 60° to 111°E. HMA comprises high-elevation zones such as the Tibetan Plateau, Hindu Kush, and Tien Shan, which are sources of many major rivers (e.g., the Ganges, the Indus, and the Yangtze). Eleven hydrologic basins are included in the domain: Yangtze, Si, Song Hong, Irrawaddy, Hwang Ho, Ganges–Brahmaputra, Indus, Tarim, Ili, Amu Darya, and Syr Darya (Fig. 1). These basins play a critical role in sustaining the economy, agriculture, and energy of more than a billion people living in about 10 countries, including China, Myanmar, Bhutan, Nepal, Bangladesh, India, Pakistan, Afghanistan, and Kyrgyzstan. Besides being populated, HMA is characterized by strong heterogeneities in topography, land cover, and land use. As a result, the atmospheric dynamics are also heterogeneous with the eastern and southern regions characterized by the East Asian and Indian monsoons and the western areas including the Hindu Kush and the Karakoram subject to the westerlies. The abrupt changes in elevation lead the precipitation patterns to be highly contrasted with some areas having 4 times more precipitation than others.
3. Data and methodology
a. Gridded-precipitation products: ERA5, IMERG, and CHIRPS
We assessed seven gridded precipitation datasets (see appendix A) in High Mountain Asia to ultimately select three gridded precipitation products (ERA5, IMERG, and CHIRPS). ERA5 provides hourly estimates of precipitation by combining satellite and in situ data into global estimates using advanced modeling and data assimilation systems on a 30-km grid (Hersbach et al. 2020). IMERG uses information from the GPM satellite constellation to estimate precipitation over Earth’s surface at a spatial resolution of 0.1° (approximately 10 km) and 30-min temporal resolution (Huffman et al. 2015). We used the final run product which incorporates multisatellite and gauge data. CHIRPS incorporates infrared data and gauge products and provides a daily quasi-global precipitation dataset at a resolution of approximately 5 km (Funk et al. 2015).
When both IMERG and ERA5 do not estimate precipitation within the day, the coefficient αh is equal to αh = 1/24.
b. Computing the ensemble precipitation
We compute two different ensemble-based precipitation products using EM and LPM as described by Clark (2017). Specifically, 1) EM is based on a simple average of the three ensemble members at each grid point and 2) the LPM method is used to replace the ensemble mean amounts with amounts sampled from the distribution of ensemble forecasts. LPM involves the following steps:
-
Compute the arithmetic mean of the three ensemble members at each grid point.
-
Sort all the values of the arithmetic mean grid from lowest to highest within a given radius.
-
Record the rank of the arithmetic mean of the center point Rmean.
-
Sort of the values of the three ensemble members from lowest to highest within the same radius.
-
Replace the value of the neighborhood center point with the sorted ensemble member value ranked Rens withwhere “nint” is the nearest integer and M = 3 the number of ensemble members, N is equal to the number of points within the selected radius, and σ is a coefficient equal to 1.05 as in Clark (2017). As shown in appendix B, low values (<1) of σ lead to very localized precipitation patterns on contrary to large values (>1.5) of σ.
The computational overhead associated with LPM increases significantly with larger radii values. For example, it requires 90 min to generate 1 h of data using a 50-km radius. In addition, large radii could lead to unrealistic precipitation patterns (Clark 2017). Therefore, a 25-km radius was chosen in this study after experimenting with several radii (please refer to appendix B for more details) as a reasonable compromise between computation time and accuracy.
More details about the LPM method and the differences with the classical PM could be found in Clark (2017). The application of LPM to large domains such as HMA is computationally expensive and has not been performed before. To overcome this limitation, we generated a lookup table in which each cell is associated with all cells located within the 25-km radius. The lookup table indicates for each cell the total number of cells located within the 25-km radius and their associated indices. This facilitates the ranking as the search of the cells located in the defined radius is done only once.
c. Statistical analyses of the products
The performances LPM are assessed by comparing the spatial and temporal variations of their averages, trends, standard deviations, and ratio power, as described below.
1) Extended triple collocation
2) Ratio power
3) Averages and standard deviations
4) Trends
4. Results
a. Analysis of the RMSE and correlation coefficient of three selected ensemble members
Figure 2 illustrates the spatial distributions of the RMSE and r of the three ensemble members from the ETC analysis. IMERG has the lowest RMSE, indicating that this product has the lowest random errors compared to the other two products, and the highest r. ERA5 RMSE shows values twice as high as the IMERG RMSE. For all the three ensemble members, large RMSEs are mainly found at the intersection of the Ganges–Brahmaputra and Irrawaddy basins, an area characterized by sharp changes in elevation and high precipitation. High RMSE values are also found in the Himalayas ranges and high-elevation zones of the Yangtze, suggesting the influence of complex orography on the higher precipitation uncertainties.
b. Performance of LPM at a selected time
Figure 3a shows the spatial distributions of precipitation of the two ensemble products (EM and LPM) and their members at 2200 local time (LT) 6 August 2001, a day where the precipitation was high and has a strong spatial variability. At this time, the three ensemble members have different spatial distributions. The EM spatial distribution is very similar to ERA5 because ERA5 has high precipitation values. Also, the effects of the coarse resolution of ERA5 are also prominent in EM. On the contrary, LPM does not have such distribution patterns (high precipitation values are very localized) and contains smaller areas with light precipitation. EM creates artificial precipitation by assigning more spatially persistent nonzero precipitation values, which is not realistic. LPM also provides more consistent precipitation representations in areas where there are large disagreements across the ensemble members. For example, over the Tarim basin, ERA5 estimates heavy precipitation, CHIRPS does not indicate a precipitation event, and IMERG captures localized precipitation. In this area, EM precipitation patterns include nonzero and high precipitation similar to ERA5 whereas LPM precipitation is more localized and shows lower precipitation values. As such, LPM significantly reduces ERA5 biases. The three ensemble members also show different spatial distributions over the Yangtze basin. Consequently, EM precipitation field contains light precipitation scattered almost everywhere whereas LPM restricts such patterns. Another advantage of the LPM technique is accounting for the information provided by both high- and low-resolution ensembles. Due to the model physics and parameterizations especially in complex terrains, high- and low-resolution simulations have different biases. Therefore, merging different resolutions could lead to more realistic precipitation patterns when using LPM. In such instances, EM does not account for the detailed and localized features and structures of the high-resolution product and provides a precipitation field similar to the product with a coarse resolution. This limitation makes EM less suitable for generating a consensus ensemble product when the members have different spatial resolutions. Moreover, because ERA5 has very high precipitation values, the temporal variation of EM is very akin to the ERA5 temporal variations (Fig. 3b). LPM has precipitation when both IMERG and ERA5 forecasts precipitation at this point, whereas EM has precipitation even when only ERA5 indicates nonzero precipitation (Fig. 3c).
Spatial distribution of RMSE and correlation coefficients r of the three selected datasets. IMERG has the lowest random errors and high correlation coefficient. The white color indicates areas with no valid data.
Citation: Journal of Hydrometeorology 23, 9; 10.1175/JHM-D-21-0196.1
(a) Maps of precipitation at 2200 LT 6 Aug 2001 of the three ensemble members (ERA5, CHIRPS, and IMERG) and the two ensemble products (EM and LPM) and (b),(c) temporal variation of precipitation at a single point (located at 27.5°N, 101.25°E) at two selected times.
Citation: Journal of Hydrometeorology 23, 9; 10.1175/JHM-D-21-0196.1
c. Quantifying the aggregated precipitation amounts
Figure 4 illustrates the spatial distributions of the relative differences (Δ) of the yearly averages of precipitation of EM, LPM, and their constituent members. Because the relative differences obtained with EM and LPM are similar, we only show the differences with LPM and focus on the differences between EM and LPM. An analysis of the spatial distributions of the yearly averages (not shown here) has shown that the three ensemble members have similar spatial distributions, areas with high precipitation (the Himalayas ranges and the southeast of the domain) remain the same across the different products, which is consistent with the results of Yoon et al. (2019). In general, the precipitation of EM is lower than that of ERA5 and higher than the CHIRPS and IMERG precipitation.
Maps of the relative differences [(Δ; defined in Eq. (7)] of the yearly average precipitation between LPM and the three ensemble members (ERA5, CHIRPS, and IMERG) and EM and LPM.
Citation: Journal of Hydrometeorology 23, 9; 10.1175/JHM-D-21-0196.1
Areas with large relative differences correspond to the high-elevation areas where abrupt changes in orography make precipitation estimates more uncertain and create disparities among the different products. In these areas, EM tends to smoothen the large differences in precipitation whereas LPM accounts for these singularities. Although the low-elevation areas of the Ganges–Brahmaputra have large precipitation amounts, the relative differences are low due to the absence of elevation gradients. Because EM overpredicts light precipitation, its yearly average precipitation is slightly higher than the yearly average of LPM mean, which is close to the IMERG yearly precipitation as shown in Fig. 5.
Yearly average domain precipitation of the three ensemble members (ERA5, CHIRPS, and IMERG) and the two ensemble products (EM and LPM): (a) time series, (b) distribution, and (c) as a function of elevation.
Citation: Journal of Hydrometeorology 23, 9; 10.1175/JHM-D-21-0196.1
The histogram of ERA5 yearly precipitation averages is more dispersed than that of IMERG and CHIRPS (Fig. 5b). EM reproduces the distribution of IMERG, whereas LPM includes more extremes and has a more dispersed distribution. The relationship between elevation and precipitation is the same for EM and LPM because the three ensemble members have similar changes of precipitation with respect to elevation. The differences between EM and LPM are mainly in high-elevation areas due to the inconsistencies of the three members.
d. Statistical analyses of the precipitation fields: analysis of the filtering effects of LPM
As expected, EM has a standard deviation lower than LPM (Fig. 6a). In the generation of an ensemble consensus product, the standard deviation is important because it could be used to create a set of ensembles which encompasses all the ranges of precipitation of the different products used to generate the ensemble product. Overall, LPM better captures the ranges of precipitation variability.
Maps of (a) standard deviation of the two ensemble products (EM and LPM) and the difference between the two and (b) ratio power between the two ensemble products (EM and LPM) and the three ensemble members (ERA5, CHIRPS, and IMERG). Areas in red indicate zones where the ensemble has a variance lower than the ensemble member (within the 25-km radius), whereas the blue areas represent the areas where the ensemble has high variance.
Citation: Journal of Hydrometeorology 23, 9; 10.1175/JHM-D-21-0196.1
Figure 6b illustrates the spatial distributions of the ratio power computed between EM and LPM and the three ensemble members, based on yearly average precipitation fields. The ratio power is computed using a radius equal to 25 km because previous studies have shown that these differences are more preponderant at a small-scale where the ensemble members would disagree (Surcel et al. 2014). Figure 6 shows that the large-scale spatial distributions are generally consistent among EM and LPM. In general, EM has locally less spatial variance than ERA5 despite the coarse resolution of this product. In low-elevation areas, EM contains more small-scale heterogeneities than ERA5. In contrast, the complexities and localized precipitation patterns due to the complex orography in high-elevation areas within ERA5 are smoothed out in EM. IMERG, on the other hand, has more heterogeneities in low elevation than in the Tibetan Plateau compared to EM because of its higher resolution. Even though IMERG has smaller random errors than the two other ensemble members, the product shows less precipitation heterogeneity in the plateau. The spatial distribution of the ratio power of CHIRPS with respect to EM has dispersed areas with higher and lower variances. In general, low-elevation areas of the Ganges–Brahmaputra, Indus, Yangtze, Hwang Ho, and the Tibetan Plateau have more local heterogeneities in CHIRPS than EM whereas the product tends to smoothen the precipitation in the Himalayas ranges. LPM overcomes these disparities and provides a product that has almost everywhere more spatial heterogeneities than any ensemble members. Therefore, LPM preserves the localized precipitation patterns of low-elevation areas as described by CHIRPS and IMERG and the refined high-elevation precipitation of ERA5.
e. Long-term trends
Temporal trends of precipitation are critical to reproducing and assessing trends of water cycle variables. The uncertainties in gridded precipitation products also extend to long-term trends, particularly over complex terrain such as HMA. Here we test the ability of EM and LPM to generate reasonable trends of precipitation. Figure 7 depicts the spatial distributions of the trends of precipitation over the past two decades for the three ensemble members, EM, and LPM. ERA5, CHIRPS, and IMERG present very different spatial distributions of trends. The three ensemble members agree on the increasing trends of precipitation in the Indus and western part of the Ganges–Brahmaputra and the decreasing trends over eastern Ganges–Brahmaputra. While ERA5 indicates that the precipitation is increasing in the Himalayas ranges, IMERG and CHIRPS generally show a decrease, although some local areas have increasing trends. The three ensemble members also agree on the low increasing trends of precipitation in the central portion of the Tibetan Plateau. IMERG and CHIRPS precipitation has increased over the past two decades in the Yangtze and Hwang Ho, whereas ERA5 estimates a decrease. Similarly, precipitation trends of the basins located in the northwest differ according to the product; ERA5 and CHIRPS show a decrease whereas IMERG exhibits an increase. The spatial distributions of the precipitation trends of EM and LPM are similar, and as for the yearly averages, ensemble generation filters out areas where only one product differs from the other two. Because IMERG and CHIRPS have in general similar trends, the relative differences between the LPM trends and the IMERG and CHIRPS trends are inferior to the differences between LPM and ERA5, the outlier (Fig. 7b). EM has mainly the trend sign and spatial distribution of CHIRPS and IMERG whereas LPM distribution incorporates more negative trends from CHIRPS and ERA5. For example, the consensus products allow changing the ERA5 trend sign of the Yangtze and Hwang Ho basins.
Maps of (a) yearly trends of precipitation of the three ensemble members (ERA5, CHIRPS, and IMERG) and the two ensemble products (EM and LPM), (b) relative difference of trends between LPM and the three ensemble members (ERA5, CHIRPS, and IMERG) and EM and LPM, and (c) changes in trends signs between LPM and the three ensemble members (ERA5, CHIRPS, and IMERG) and EM and LPM.
Citation: Journal of Hydrometeorology 23, 9; 10.1175/JHM-D-21-0196.1
LPM trends are higher than the EM trends and the differences between the spatial distributions of these two trends are mainly prevalent in areas where the ensemble members differ such as the northwest basins, Himalayas ranges, Irrawaddy, Yangtze, and Hwang Ho. By averaging the three products, EM tends to give more weight to ERA5 trends especially at low elevations (inferior to 4000 m), despite its high uncertainty, because ERA5 trends have high values (Fig. 8). LPM corrects this issue by making the trends more realistic and closer to CHIRPS and IMERG. As opposed to the yearly averages, LPM trends are higher than the EM trends. This is because the high random errors of ERA5 characterized by high total precipitations and low trends are carried out in EM.
Yearly domain-precipitation trends of the three ensemble members (ERA5, CHIRPS, and IMERG) and the two ensemble products (EM and LPM): (a) distribution and (b) as a function of elevation.
Citation: Journal of Hydrometeorology 23, 9; 10.1175/JHM-D-21-0196.1
f. Comparisons with ground measurements
We collected ground-based measurements from three networks providing daily (but discontinuous) precipitation values. There are 29 stations in the study domain (Fig. 9). Most of the stations are in the Ganges–Brahmaputra basin, the other stations are in the Tarim, the Indus and the Tibetan Plateau. Data of the stations located in the Tarim basin are provided by the Chinese Surface Stations for Global Exchange Version 3.0 product collected by the Chinese Meteorology Administrative (CMA). The stations scattered in the high-elevation area of the Ganges–Brahmaputra and the Indus are from the Nepalese Department of Hydrology and Meteorology (DHM) and the Pakistan Meteorology Department (PMD) respectively. Figure 9 indicates the average number of daily measurements per year for each station. The stations located in the Ganges–Brahmaputra have the highest number of measurements whereas the ones located in the northern areas of HMA have less than 50 daily measurements per year. The lack of ground measurements in HMA makes the comparison between the gridded precipitation products and the measurements difficult.
Ground stations in HMA and their average number of daily measurements.
Citation: Journal of Hydrometeorology 23, 9; 10.1175/JHM-D-21-0196.1
Generally, the ground measurements indicate high precipitation in the Ganges–Brahmaputra and low precipitation in the Tarim and the northern areas (Fig. 10). This is similar to the spatial distributions of the annual averaged precipitation shown in Fig. 4a. Although this pattern is well captured by all the three ensemble members, the patterns of IMERG and CHIRPS are closer to the measurements than ERA5. As a result (because at least two ensemble members capture the pattern), EM and LPM also have this pattern. Ground measurements have higher precipitation than all the three ensemble members and the two ensemble products (EM and LPM). Only ERA5 precipitation comes close to the measured precipitation while CHIRPS and IMERG (though good at capturing the spatial patterns) are dry compared to the ground measurements. The precipitation magnitudes of EM and LPM are closer to the measurements than CHIRPS and IMERG, though they are drier than ERA5 (and ground measurements). Combining the ability of CHIRPS and IMERG to capture the spatial pattern and the high precipitation of ERA5 allows generating ensembles that provide better match to patterns and magnitudes of the ground measurements. Because EM gives more weight to ERA5, which has a high precipitation and creates a precipitation field with nonzero values in many areas, it has higher annual averages of precipitation which are closer to the measurements than LPM. As shown in appendix B, because the other products such as MERRA2 and GSMAP are drier than CHIRPS, IMERG, and ERA5, the precipitation associated with these products are very dry compared to the measurements.
(a) Spatial distributions of the average annual precipitation at ground stations and (b) annual average precipitation for each station and product.
Citation: Journal of Hydrometeorology 23, 9; 10.1175/JHM-D-21-0196.1
Figure 11 shows the trends of precipitation at the 29 ground stations. According to these measurements, the precipitation is decreasing in northern HMA and increasing in the Ganges–Brahmaputra. Note that this is contrary to GSMAP and MERRA2 that generally indicate an increasing trend of precipitation everywhere. Precipitation trends computed using measurements have magnitudes higher than the trends of the other products. ERA5 trends are the closest to the trends of the measurements. From station 1 to 15 (Fig. 11b), measurements show increasing trends, and only ERA5 captures these trends whereas IMERG and CHIRPS depict decreasing trends. Because ERA5 positive trends are very high, the resulting ensemble products have positive trends (consistent with the measurements) despite the negative trends of IMERG and CHIRPS. LPM trends are even closer to measurements than EM because LPM performs better than EM at 21 out of the 29 stations. From station 20 to 25, the measurements indicate a decreasing trend as IMERG and CHIRPS whereas ERA5 has an increasing trend. Again, the ensemble products show consistency with the measurements as they follow the trends of IMERG and CHIRPS.
(a) Spatial distributions of the yearly trends of precipitation at ground stations and (b) yearly trends of precipitation for each station and product.
Citation: Journal of Hydrometeorology 23, 9; 10.1175/JHM-D-21-0196.1
5. Summary and conclusions
Because of its complex orography and harsh environments, ground measurements of precipitation in HMA are difficult and the majority of gridded precipitation products fail to provide consistent estimates. Assessment of the accuracy of these products in HMA is a very active topic and there is no general agreement in the community on the best and most accurate gridded precipitation product to date. HMA has low and flat elevation areas, steep gradients of elevation, and high-elevation zones; therefore, one product may perform well in low elevation and fail to estimate high-elevation precipitation. Ensemble precipitation can overcome the disparities and inconsistencies of individual gridded products. This manuscript presents the performance of LPM-based consensus ensemble precipitation product generated using IMERG, CHIRPS, and ERA5 on long-term averages and trends in HMA. Additionally, an EM-based estimate was also developed to contrast the utility of the LPM against a simpler approach. An extended triple collocation evaluation shows that IMERG has smaller errors than CHIRPS and ERA5, with the latter showing precipitation twice higher than the two others. Our analysis also concludes the following:
-
Because ERA5 has high random errors and precipitation amounts, both temporal and spatial variations of EM are similar to ERA5, indicating that these biases are carried out in EM, whereas LPM significantly reduces them. LPM provides a more realistic precipitation field especially when one member has high biases and has precipitation values more than twice the other products.
-
LPM accounts for the small-scale features and heterogeneities embedded in the high-resolution data while considering the global average provided by the low-resolution data. Therefore, LPM may be a good option for merging data with different resolutions.
-
In terms of yearly averages, differences between the two ensemble techniques (EM and LPM) are mainly prominent in high-elevation zones where there are high inconsistencies among the different members.
-
By filtering disparities among the different products, the ensemble products overcome the problem of a single product performing well only in a particular area. Areas with low accurate precipitation estimates are significantly reduced in these products. Both EM and LPM improve precipitation estimation especially in high-elevation areas, but the LPM improvement is more significant.
-
Ensemble products eliminate inconsistent trends and keep only trend signs that two ensemble members agreed on. However, EM trends are closer to ERA5 than LPM which further improves the estimate of trends in areas with significant disagreements between the three ensemble members.
-
Combining the ability of CHIRPS and IMERG to capture the spatial pattern of the measured precipitation and the high precipitation of ERA5 allows generating ensembles that come close to both the spatial pattern and the value of the measured precipitation.
This work is motivated by the need for reliable estimates of precipitation over HMA, which is a critical input for quantifying hydrological storage in HMA, an area with both regional and global water supply implications. The conclusions presented above highlight the utility of the LPM-based consensus ensemble precipitation product, with reasonable spatial patterns, consistent long-term trends, and consistency with ground measurements. We expect this product to contribute to studies for characterizing the hydrology of HMA.
Acknowledgments.
This research was supported by the grant from the National Aeronautics and Space Administration High Mountain Asia program (19-HMA19-0012). Computing was supported by the resources at the NASA Center for Climate Simulation. The authors declare no conflict of interest.
Data availability statement.
Datasets used in this study can be found in the following websites: ERA5 forcing: https://www.ecmwf.int/en/forecasts/datasets/reanalysis-datasets/era5; IMERG Precipitation: https://gpm.nasa.gov/taxonomy/term/1372; CHIRPS Precipitation: https://www.chc.ucsb.edu/data.
APPENDIX A
Analysis of Gridded Precipitation Datasets over High Mountain Asia
We selected seven widely used gridded precipitation products in High Mountain Asia (Table A1): 1) APHRODITE (Asian Precipitation–Highly Resolved Observational Data Integration Towards Evaluation of Water Resources) is a daily gridded precipitation dataset for Asia that is generated from a dense network of daily rain gauge data (Yatagai et al. 2012); 2) CHIRPS provides a quasi-global precipitation dataset at a resolution of approximately 5 km (Funk et al. 2015); 3) IMERG provides the estimation of precipitation at a spatial resolution of 10 km (Huffman et al. 2015); 4) ERA5 provides hourly estimates of meteorological forcing on a 30-km grid (Hersbach et al. 2020); 5) MERRA2 (Modern-Era Retrospective Analysis for Research and Applications, version 2) is the latest atmospheric reanalysis from the NASA Global Modeling and Assimilation Office and is produced with the Goddard Earth Observing System model version 5 (GEOS-5) data assimilation system (Gelaro et al. 2017); 6) GSMAP (Global Satellite Mapping of Precipitation), a product of the GPM mission, provides a global hourly rain rate with a 0.1° resolution; and 7) WRF-HMA is a reanalysis of HMA atmospheric dynamics using the Weather Research and Forecasting (WRF) Model at a 4-km and hourly resolution. We performed the analysis from 2001 to 2016.
Selected gridded precipitation products
Figure A1 depicts the monthly variations of average precipitation in HMA. ERA5 has the highest precipitation rates in the domain whereas MERRA2 is the driest precipitation product. The monthly variations of IMERG and CHIRPS are similar. From 2001 to 2008, the monthly variations of APHRODITE are similar to MERRA2, indicating that APHRODITE is also dry compared to CHIRPS, IMERG, and ERA5.
Monthly variations of average precipitation in HMA.
Citation: Journal of Hydrometeorology 23, 9; 10.1175/JHM-D-21-0196.1
We then computed average precipitation at different elevation zones (Fig. A2). In areas where the elevation is below 2000, 3000, and 4000 m, APHRODITE, GSMAP, and WRF-HMA are the driest precipitation products, whereas IMERG is the product with the highest precipitation followed by ERA5. In these low-elevation zones, MERRA2 has more precipitation than CHIRPS.
Average annual precipitation by elevation zones: (a) low elevation below 2000, 3000, and 4000 m and (b) high elevation above 2000, 3000, and 4000 m.
Citation: Journal of Hydrometeorology 23, 9; 10.1175/JHM-D-21-0196.1
MERRA2 and GSMAP are the driest precipitation products in high-elevation zones. WRF-HMA and ERA5 have the highest precipitation followed by IMERG and CHIRPS. APHRODITE’s precipitation estimates indicate precipitation higher than in MERRA2 and GSMAP.
Although at the domain scale MERRA2 is the driest product, this product has high precipitation estimates in low elevation. On the contrary, APHRODITE precipitation estimates are relatively low in low elevation and almost similar to CHIRPS and IMERG in high elevation. CHIRPS and IMERG have different behavior: CHIRPS is dry in low elevations whereas IMERG is dry in high elevations. In these high-elevation zones, only ERA5 and WRF-HMA predict high precipitation.
Precipitation significantly changes across the Himalayas region. We compare the different products in the region. We defined three zones: west, central, and east Himalayas (Fig. A3). All the products except WRF-HMA capture the precipitation gradient, i.e., the precipitation increases from west to east. The differences of precipitation are more pronounced in the east than in the west due to the orographic complexities of the eastern region of the Himalayas. GSMAP is very dry in the Himalayas whereas ERA5 is the product with the highest precipitation. The difference in precipitation between the east and the west is very high in ERA5 compared to the other products. APHRODITE, CHIRPS, IMERG, and ERA5 have similar total precipitation in the west.
Average annual precipitation of High Mountain Asia and the west, central, and east Himalayas.
Citation: Journal of Hydrometeorology 23, 9; 10.1175/JHM-D-21-0196.1
Because the spatial coverage of WRF-HMA is smaller than the actual HMA domain and the preliminary analyses discussed above have shown that this product is very dry and not suitable for low elevation, we did not show the results of this product in the following paragraphs. As mentioned previously, all the products capture the spatial distribution of the precipitation (Fig. A4). Only the rates of precipitation are different from one product to another. As shown in Fig. A4, the spatial distributions of precipitation in winter where it is dominated by the westerlies is different from summer where the precipitation is a result of the monsoon. Nonetheless, all the products capture these spatial distributions.
Spatial distribution of (a) the annual (b) winter and summer average of precipitation.
Citation: Journal of Hydrometeorology 23, 9; 10.1175/JHM-D-21-0196.1
On the contrary to the spatial distributions of the annual average precipitation, the spatial distributions of the annual trends differ from one product to another (Fig. A5). The magnitudes of the trends are also different. MERRA2, IMERG, CHIRPS, and GSMAP estimate an increasing trend of precipitation in the Yangtze basin located in the eastern part of the domain whereas ERA5 and APHRODITE indicate a decrease. MERRA2 and GSMAP indicate in general that the precipitation is increasing in HMA only sparse areas are characterized by decreasing trends. The increase in precipitation in MERRA2 is mostly concentrated in the Himalayas region. ERA5 and APHRODITE indicate decreasing precipitation in the foothill region of the Himalayas and an increasing trend in the high elevation of the region.
Spatial distribution of the annual trends of precipitation.
Citation: Journal of Hydrometeorology 23, 9; 10.1175/JHM-D-21-0196.1
We evaluated the random errors associated with the different products using the extended triple collocation method (Fig. A6). Previous analyses have shown that GSMAP is very dry and has trends not consistent with the other products; therefore, we computed the random errors using five products: MERRA2, ERA5, APHRODITE, CHIRPS, and IMERG. IMERG is the product with low random errors followed by APHRODITE whereas the two products derived from reanalysis ERA5 and MERRA2 have the highest random errors.
Spatial distribution of the RMSE calculated using the extended triple collocation method.
Citation: Journal of Hydrometeorology 23, 9; 10.1175/JHM-D-21-0196.1
Our comparisons have shown that MERRA2 and GSMAP are very dry and are characterized by increasing trends of precipitation, IMERG and CHIRPS have similar annual averages and have the lowest random errors. Although APHRODITE has low random errors, the product does not have data to the present day which limits its use for our analysis.
APPENDIX B
Sensitivity of the Spatial Distributions of LPM Precipitation to the Values of σ and Radii
We analyze the sensitivity of the LPM precipitation fields to the values of σ and the radii (Fig. B1).
Spatial distributions of LPM precipitation at 0700 LT 1 Jan 2001 with different values of σ and computed using different radii.
Citation: Journal of Hydrometeorology 23, 9; 10.1175/JHM-D-21-0196.1
REFERENCES
Andermann, C., S. Bonnet, and R. Gloaguen, 2011: Evaluation of precipitation data sets along the Himalayan front. Geochem. Geophys. Geosyst., 12, Q07023, https://doi.org/10.1029/2011GC003513.
Buizza, R., and T. N. Palmer, 1998: Impact of ensemble size on ensemble prediction. Mon. Wea. Rev., 126, 2503–2518, https://doi.org/10.1175/1520-0493(1998)126<2503:IOESOE>2.0.CO;2.
Clark, A. J., 2017: Generation of ensemble mean precipitation forecasts from convection-allowing ensembles. Wea. Forecasting, 32, 1569–1583, https://doi.org/10.1175/WAF-D-16-0199.1.
Clark, A. J., W. A. Gallus, M. Xue, and F. Kong, 2009: A comparison of precipitation forecast skill between small convection-allowing and large convection-parameterizing ensembles. Wea. Forecasting, 24, 1121–1140, https://doi.org/10.1175/2009WAF2222222.1.
Ebert, E. E., 2001: Ability of a poor man’s ensemble to predict the probability and distribution of precipitation. Mon. Wea. Rev., 129, 2461–2480, https://doi.org/10.1175/1520-0493(2001)129<2461:AOAPMS>2.0.CO;2.
Funk, C., and Coauthors, 2015: The climate hazards infrared precipitation with stations—a new environmental record for monitoring extremes. Sci. Data, 2, 150066, https://doi.org/10.1038/sdata.2015.66.
Gelaro, R., and Coauthors, 2017: The Modern-Era Retrospective Analysis for Research and Applications, version 2 (MERRA-2). J. Climate, 30, 5419–5454, https://doi.org/10.1175/JCLI-D-16-0758.1.
Hersbach, H., and Coauthors, 2020: The ERA5 global reanalysis. Quart. J. Roy. Meteor. Soc., 146, 1999–2049, https://doi.org/10.1002/qj.3803.
Hong, Z., Z. Han, X. Li, D. Long, G. Tang, and J. Wang, 2021: Generation of an improved precipitation dataset from multisource information over the Tibetan Plateau. J. Hydrometeor., 22, 1275–1295, https://doi.org/10.1175/JHM-D-20-0252.1.
Huffman, G. J., D. T. Bolvin, and E. J. Nelkin, 2015: Integrated Multi-SatellitE Retrievals for GPM (IMERG) technical documentation. NASA/GSFC Code, 612, 47 pp., https://gpm.nasa.gov/sites/default/files/document_files/IMERG_doc.pdf.
Kendall, M. G., 1948: Rank Correlation Methods. Oxford University Press, 160 pp.
Ma, L., T. Zhang, O. W. Frauenfeld, B. Ye, D. Yang, and D. Qin, 2009: Evaluation of precipitation from the ERA-40, NCEP-1, and NCEP-2 reanalyses and CMAP-1, CMAP-2, and GPCP-2 with ground-based measurements in China. J. Geophys. Res., 114, D09105, https://doi.org/10.1029/2008JD011178.
Mann, H. B., 1945: Nonparametric tests against trend. Econometrica, 13, 245–259, https://doi.org/10.2307/1907187.
McColl, K. A., J. Vogelzang, A. G. Konings, D. Entekhabi, M. Piles, and A. Stoffelen, 2014: Extended triple collocation: Estimating errors and correlation coefficients with respect to an unknown target. Geophys. Res. Lett., 41, 6229–6236, https://doi.org/10.1002/2014GL061322.
Molteni, F., R. Buizza, T. N. Palmer, and T. Petroliagis, 1996: The ECMWF ensemble prediction system: Methodology and validation. Quart. J. Roy. Meteor. Soc., 122, 73–119, https://doi.org/10.1002/qj.49712252905.
Müller Schmied, H., and Coauthors, 2016: Variations of global and continental water balance components as impacted by climate forcing uncertainty and human water use. Hydrol. Earth Syst. Sci., 20, 2877–2898, https://doi.org/10.5194/hess-20-2877-2016.
Newman, A. J., and Coauthors, 2015: Gridded ensemble precipitation and temperature estimates for the contiguous United States. J. Hydrometeor., 16, 2481–2500, https://doi.org/10.1175/JHM-D-15-0026.1.
Palazzi, E., J. von Hardenberg, and A. Provenzale, 2013: Precipitation in the Hindu-Kush Karakoram Himalaya: Observations and future scenarios. J. Geophys. Res. Atmos., 118, 85–100, https://doi.org/10.1029/2012JD018697.
Pritchard, H. D., 2019: Asia’s shrinking glaciers protect large populations from drought stress. Nature, 569, 649–654, https://doi.org/10.1038/s41586-019-1240-1.
Qiao, X., S. Wang, C. S. Schwartz, Z. Liu, and J. Min, 2020: A method for probability matching based on the ensemble maximum for quantitative precipitation forecasts. Mon. Wea. Rev., 148, 3379–3396, https://doi.org/10.1175/MWR-D-20-0003.1.
Rienecker, M. M., and Coauthors, 2011: MERRA: NASA’s modern-era retrospective analysis for research and applications. J. Climate, 24, 3624–3648, https://doi.org/10.1175/JCLI-D-11-00015.1.
Roebeling, R. A., E. L. A. Wolters, J. F. Meirink, and H. Leijnse, 2012: Triple collocation of summer precipitation retrievals from SEVIRI over Europe with gridded rain gauge and weather radar data. J. Hydrometeor., 13, 1552–1566, https://doi.org/10.1175/JHM-D-11-089.1.
Schwartz, C. S., G. S. Romine, K. R. Fossell, R. A. Sobash, and M. L. Weisman, 2017: Toward 1-km ensemble forecasts over large domains. Mon. Wea. Rev., 145, 2943–2969, https://doi.org/10.1175/MWR-D-16-0410.1.
Snook, N., F. Kong, A. Clark, B. Roberts, K. A. Brewster, and M. Xue, 2020: Comparison and verification of point-wise and patch-wise localized probability-matched mean algorithms for ensemble consensus precipitation forecasts. Geophys. Res Lett., 47, e2020GL087839, https://doi.org/10.1029/2020GL087839.
Song, C., B. Huang, L. Ke, and Q. Ye, 2016: Precipitation variability in high mountain Asia from multiple datasets and implication for water balance analysis in large lake basins. Global Planet. Change, 145, 20–29, https://doi.org/10.1016/j.gloplacha.2016.08.005.
Surcel, M., I. Zawadzki, and M. K. Yau, 2014: On the filtering properties of ensemble averaging for storm-scale precipitation forecasts. Mon. Wea. Rev., 142, 1093–1105, https://doi.org/10.1175/MWR-D-13-00134.1.
Tong, K., F. Su, D. Yang, L. Zhang, and Z. Hao, 2014: Tibetan plateau precipitation as depicted by gauge observations, reanalyses and satellite retrievals. Int. J. Climatol., 34, 265–285, https://doi.org/10.1002/joc.3682.
Viviroli, D., H. H. Dürr, B. Messerli, M. Meybeck, and R. Weingartner, 2007: Mountains of the world, water towers for humanity: Typology, mapping, and global significance. Water Resour. Res., 43, https://doi.org/10.1029/2006WR005653.
Warner, T. T., 2010: Numerical Weather and Climate Prediction. Cambridge University Press, 526 pp, https://doi.org/10.1017/CBO9780511763243.
Weusthoff, T., D. Leuenberger, C. Keil, and G. C. Craig, 2011: Best member selection for convective-scale ensembles. Meteor. Z., 20, 153–164, https://doi.org/10.1127/0941-2948/2011/0211.
Yatagai, A., K. Kamiguchi, O. Arakawa, A. Hamada, N. Yasutomi, and A. Kitoh, 2012: APHRODITE: Constructing a long-term daily gridded precipitation dataset for Asia based on a dense network of rain gauges. Bull. Amer. Meteor. Soc., 93, 1401–1415, https://doi.org/10.1175/BAMS-D-11-00122.1.
Yoon, Y., and Coauthors, 2019: Evaluating the uncertainty of terrestrial water budget components over high mountain Asia. Front. Earth Sci., 7, 120, https://doi.org/10.3389/feart.2019.00120.
You, Q., J. Min, W. Zhang, N. Pepin, and S. Kang, 2015: Comparison of multiple datasets with gridded precipitation observations over the Tibetan Plateau. Climate Dyn., 45, 791–806, https://doi.org/10.1007/s00382-014-2310-6.
Yue, S., P. Pilon, and G. Cavadias, 2002: Power of the Mann–Kendall and Spearman’s rho tests for detecting monotonic trends in hydrological series. J. Hydrol., 259, 254–271, https://doi.org/10.1016/S0022-1694(01)00594-7.
