Evaluation and Error Analysis of Multi-Source Precipitation Datasets during Summer over the Tibetan Plateau

Abstract

Due to the scarcity of meteorological stations on the Tibetan Plateau (TP), owing to the high altitude and harsh climate, studies often resort to satellite, reanalysis, and merged multi-source precipitation data. This necessitates an evaluation of TP precipitation data applicability. Here, we assess the following three high-resolution gridded precipitation datasets: the China Meteorological Forcing Dataset (CMFD), the European Center for Medium-Range Weather Forecasts Reanalysis V5-Land (ERA5-Land), and Integrated Multi-satellitE Retrievals for Global Precipitation Measurement (IMERG) during TP summers. Using observations from the original 133 China Meteorological Administration stations on the TP as a reference, the evaluation yielded the following conclusions: (1) In summer, from 2000 to 2018, discrepancies among the datasets were largest in the western TP. The CMFD showed the smallest deviation from the observations, and the annual summer precipitation was only overestimated by 12.3 mm. ERA5-Land had the closest trend (0.41 mm/y) to the annual mean summer precipitation, whereas it overestimated the highest precipitation (>150 mm). (2) The reliability of the three datasets at annual and monthly scales was in the following order: CMFD, ERA5-Land, and IMERG. The daily scales exhibited a lower accuracy than the monthly scales (correlation coefficient CC of 0.51, 0.38, and 0.26, respectively). (3) The CMFD assessments, referencing the 114 new stations post-2016, had a notably lower accuracy and precipitation capture capability at the daily scale (CC and critical success index (CSI) decreased by 0.18 and 0.1, respectively). These results can aid in selecting appropriate datasets for refined climate predictions on the TP.

1. Introduction

Precipitation monitoring is critical in disaster warning, environmental monitoring, and climate system regulation [1,2,3]. Obtaining precise precipitation data applicable to different spatial-temporal features remains a challenge in meteorology and hydrology [4,5]. Currently, precipitation data are primarily acquired through various methods, including direct measurements with rain gauges and indirect approaches like interpolation/reconstruction observations, satellite remote sensing, numerical modeling, and merged multi-source datasets [6].

Rain gauges are currently the most used and reliable method for observing precipitation at the point scale; however, obtaining continuous and widespread observed precipitation data in complex terrain is difficult [7,8]. There are gridded precipitation products that have been generated by the interpolation of station observations, such as the Asian Precipitation Highly Resolved Observational Data Integration Towards Evaluation of Water Resources (APHRODITE) [9] and Global Precipitation Climatology Centre (GPCC) datasets [10]. Exercising caution when using these products for studies involving extreme events is important, and the input observations used for interpolation are not homogenized. Satellite remote sensing detects precipitation under all weather conditions and is, therefore, widely used in estimating large-scale precipitation. Notable satellite precipitation products include the Integrated Multi-satellitE Retrievals for Global Precipitation Measurement (IMERG) [11] and the Tropical Rainfall Measuring Mission (TRMM) [12]. However, satellite precipitation products are not without limitations. The challenges include the difficulty in estimating precipitation at high altitudes over a complex terrain [13,14] and the misidentification of precipitation events at high latitudes and on icy/snowy surfaces [15]. Reanalysis products, which integrate historical atmospheric information with numerical modeling results, offer another valuable resource. Prominent examples include the European Centre for Medium-Range Weather Forecasts (ECMWF) Reanalysis Version 5-Land (ERA5-Land) [16] and the National Center for Environment Prediction and the National Center for Atmospheric Research (NCEP/NCAR) reanalysis products [17]. Reanalysis products have advantages in capturing long-term precipitation changes [18], although there has been comparatively less emphasis on quality improvement endeavors [19]. Precipitation data merged from multiple sources of satellite products, reanalysis datasets, and station observations, with continuous temporal coverage and consistent quality, have also been widely applied [20]. However, different precipitation datasets have wide variations in describing the frequency, intensity, and geographic distribution of precipitation events. Therefore, their applicability should be evaluated before applying precipitation products to research [21].

The Tibetan Plateau (TP), known as the “Asian water tower,” is a pivotal source of forcing for global atmospheric movements. Changes in precipitation on the TP have crucial implications for the climate and ecology [22,23,24]. However, the TP’s challenging natural environment and complex topography result in an uneven and limited network of meteorological stations. This underscores the critical need for accurate precipitation datasets to facilitate the study of climate variability in the TP and its surrounding regions. Andermann et al. [25] focused on evaluating the performance of various datasets concerning Himalayan frontal precipitation. Their findings have highlighted the capability of the TRMM-2B31 dataset to accurately depict the distribution of topographic precipitation across large scales along the Himalayan fronts. The IMERG and the CMORP can objectively reflect the precipitation information for the Yarlung Zangbo River basin [26]. Zhan et al. [27] reported that both calibrated and uncalibrated products of the Global Precipitation Measurement (GPM)-Era do not accurately recreate the frequency–intensity structure of hourly precipitation over the western Tibetan Plateau (WTP). Regarding precipitation extremes, both the China Meteorological Forcing Dataset (CMFD) and APHRODITE slightly underestimate the extreme precipitation indices on the TP, whereas CHIRPS overestimates most indices [28]. None of the current conventional gridded precipitation products can recreate the occurrence of the Yushu snowstorm on a daily scale [29].

While numerous recent studies have evaluated cold-season precipitation datasets on the TP [29,30], there remains a paucity of systematic assessments regarding the suitability of high-resolution precipitation datasets for TP summers. During the summer months, the southern TP experiences increased precipitation owing to ample water vapor transport, alongside the presence of active plateau low vortices and shear lines [31,32]. The intricate interplay of complex topography and the Asian monsoon engenders intricate spatial and temporal patterns of summer precipitation on the TP [33,34,35]. Therefore, it is imperative to undertake further evaluations of multi-source summer precipitation datasets tailored to the TP’s unique characteristics. Furthermore, the absence of current evaluations of gridded precipitation datasets referencing new TP stations added post-2016 necessitates attention.

This study selected three conventional high-resolution precipitation datasets—CMFD, ERA5-Land, and IMERG—for evaluation. The comprehensive nature of these gridded precipitation products, in comparison to direct observations on the TP, has been scrutinized across annual, monthly, and daily scales. Our objective analysis of the applicability of the multi-source precipitation datasets to the TP during the summer is a valuable reference for enhancing climate predictions in the region.

2. Materials and Methods

2.1. Study Area

Figure 1 shows the extent of the TP within China. The digital elevation model of the TP was derived from Shuttle Radar Topography Mission data, featuring a resolution of 250 m. Situated at an average elevation exceeding 4000 m, the TP experiences a characteristic plateau mountain climate, characterized by significant daily temperature fluctuations. The precipitation on the TP follows a pattern of decreasing intensity from the southeast to the northwest, resulting in an exceptionally irregular spatial distribution. Notably, the disparity in annual precipitation between the Yarlung Zangbo River Canyon (28°–30° N, 94°–96° E) and the Qaidam Basin (35°–39° N, 90°–99° E) surpasses 500 mm.

2.2. Data

2.2.1. Gridded Precipitation Datasets

In this study, the following three high-resolution gridded precipitation datasets were used for the summers of 2000–2018: CMFD, ERA5-Land, and IMERG. The spatial resolution of all three gridded datasets was 0.1°.

The CMFD was generated by merging routine meteorological observations from the CMA with a wide range of international reanalysis and satellite remote sensing precipitation products as background fields [20,36]. A precipitation rate variable with a temporal resolution of 3 h was selected for this study.

ERA5-Land is a reanalysis dataset generated by repeating the land component of the ERA5 reanalysis [16]. It employs the H-TESSEL land surface model, covering the entire global land area. Unlike ERA5, ERA5-Land offers an improved horizontal resolution, transitioning from 31 to 9 km, making it better suited for intricate terrains. For this study, we utilized a total precipitation variable with a temporal resolution of 1 h.

The IMERG represents the latest generation of multi-satellite fused-inversion precipitation data, specifically designed for the GPM [11]. Version 06 of IMERG incorporates the estimated precipitation data obtained during the operations of the TRMM and GPM satellites. The dataset comprises three sets of satellite precipitation products presented chronologically. For this study, we relied on The Final Run version, which has been calibrated using monthly GPCC precipitation products and operates at a temporal resolution of 0.5 h.

2.2.2. Observed Data

The daily observed precipitation data from the China Meteorological Administration (CMA) stations were provided by the National Meteorological Information Center. The Chinese standard precipitation gauges are employed by CMA stations for precipitation collection. To ensure observed data quality, a total of 15 CMA stations with long-term missing data (>5% missing data) were eliminated, and 247 CMA stations were finally selected on the TP, comprising 133 original CMA stations and 114 newly constructed stations post-2016. The temporal span of the original station data extended from the summers of 2000 to 2018, while the newly added station data spanned from the summers of 2016 to 2018. If the daily observations contained missing values, we adjusted the corresponding daily gridded precipitation data to the missing values [27,37]. Notably, these missing data only accounted for 0.64% of the total observed data.

A total of 10 rain gauge stations from the hourly rain gauge data for the warm season in the central and western TP datasets were selected to further validate the accuracy of the three gridded datasets for estimating precipitation across the Qiangtang Plateau. Hobo tipping-bucket rain gauges are utilized by the rainfall stations. Figure 1 illustrates the distribution of all meteorological and rain gauge stations.

2.3. Methods

2.3.1. Spatial Interpolation

The utilization of Ordinary Kriging for interpolating station observations onto the grid enables spatial comparisons among the datasets. Ordinary Kriging interpolation offers the most accurate linear unbiased prediction of regionalized variables within a finite area [38]. This spatial interpolation method had some errors for the northwestern TP with sparse stations, whereas the effect was better for the eastern TP with more stations.

To facilitate the evaluation of the gridded datasets on multiple timescales, the gridded data values corresponding to the observation stations were extracted using the bilinear interpolation method. While interpolation inherently introduces errors in the evaluation outcomes, numerous studies have validated the suitability of bilinear interpolation for extracting gridded precipitation data at the corresponding stations [19,39].

2.3.2. Linear Tendency Estimation

To demonstrate the long-term trend of precipitation, a linear tendency estimation method was adopted [40,41]. Below, p is a variable with a sample size of n, t_i is the corresponding time series, a is the regression constant, and b is the regression coefficient.

$$p_i=a+b{t}_i(i=1,2,\dots ,n)$$

(1)

$$b=\frac{{\displaystyle \sum _{i=1}^n{p}_i{t}_i}-\frac{1}{n}({\displaystyle \sum _{i=1}^n{p}_i})({\displaystyle \sum _{i=1}^n{t}_i})}{{\displaystyle \sum _{i=1}^n{t_i}^2}-\frac{1}{n}{({\displaystyle \sum _{i=1}^n{t}_i})}^2}$$

(2)

2.3.3. Evaluation Method

The Taylor diagram simultaneously presents the correlation coefficient (CC), standard deviation ratio (σ), and centered root-mean-square (RMSD) of the spatial distributions of the model and the reference field [42]. It visually quantifies the agreement between the station precipitation observations and the grid estimates. The calculation of CC is shown in

Table 1. Evaluation metrics for the high-resolution precipitation datasets.

Evaluation Metric	Formula 1	Perfect Value
Correlation coefficient (CC)	$\mathrm{CC}=\frac{{\displaystyle \sum _{i=1}^n(x_i-\overline{x})}(y_i-\overline{y})}{\sqrt{{\displaystyle \sum _{i=1}^n{(x_i-\overline{x})}^2}}\sqrt{{\displaystyle \sum _{i=1}^n{(y_i-\overline{y})}^2}}}$	1
Relative bias (bias)	$\mathrm{bias}=\frac{{\displaystyle \sum _{i=1}^n(x_i-y_i)}}{{\displaystyle \sum _{i=1}^n{y}_i}}\times 100\%$	0
Root-mean-square error (RMSE)	$\mathrm{RMSE}=\sqrt{\frac{1}{n}{\displaystyle \sum _{i=1}^n{(x_i-y_i)}^2}}$	0
Probability of detection (POD)	$\mathrm{POD}=\frac{H}{H+M}$	1
False alarm ratio (FAR)	$\mathrm{FAR}=\frac{F}{H+F}$	0
Critical success index (CSI)	$\mathrm{CSI}=\frac{H}{H+M+F}$	1

¹ Notation: H, observed precipitation event correctly identified by the gridded precipitation datasets; M, observed precipitation event not detected by the gridded precipitation datasets; F, precipitation event detected by the gridded precipitation datasets but not observed.

and the formulae for σ and RMSD are as follows:

$$\sigma =\frac{\sqrt{\frac{1}{n}{\displaystyle \sum _{i=1}^n{(x_i-\overline{x})}^2}}}{\sqrt{\frac{1}{n}{\displaystyle \sum _{i=1}^n{(y_i-\overline{y})}^2}}}$$

(3)

$$\mathrm{RMSD}=\sqrt{\frac{1}{n}{\displaystyle \sum _{i=1}^n{[(x_i-\overline{x})-(y_i-\overline{y})]}^2}}$$

(4)

where $x_i$ is the gridded precipitation estimate, $\overline{x}$ is the mean gridded precipitation estimate, $y_i$ is the observed precipitation, $\overline{y}$ is the mean observed precipitation, and n is the number of samples.

To comprehensively evaluate the accuracy of the three precipitation datasets for estimating the summer precipitation and the capability to capture the precipitation events on the TP, certain evaluation metrics—CC, relative bias (bias), root-mean-square error (RMSE), probability of detection (POD), false alarm ratio (FAR), and critical success index (CSI)—were used in this study. The calculation formulae and perfect values are listed in Table 1. A threshold of 1 mm/d was used to distinguish between the precipitation events. On the same day, if the observed precipitation and the precipitation estimated by the gridded dataset are both greater than 1 mm, then the gridded dataset has correctly detected the precipitation event.

Figure 2 summarizes the methods and research route.

3. Results

3.1. Spatial and Temporal Patterns of Summer Precipitation on the TP

The spatial patterns of the mean annual summer precipitation during 2000–2018 on the TP are presented in Figure 3, utilizing the data from the CMA station observations, the CMFD, ERA5-Land, and the IMERG. While all of the datasets indicated that high precipitation was primarily concentrated in the southeastern TP, notable differences emerged in the WTP, where observation data were scarce. The CMFD closely approximated the observed precipitation pattern (Figure 3b). ERA5-Land estimated an annual precipitation exceeding 350 mm in the western Kunlun Mountains and the Qilian Mountains (Figure 3c). In contrast, the IMERG suggested that a large area extending from the Kunlun Mountains to the east of the Qaidam Basin received less than 50 mm of annual precipitation (Figure 3d).

To provide a detailed view of the differences between the observations and the three gridded datasets, spatially continuous station data were generated through Ordinary Kriging interpolation (Figure 4). The CMFD displayed a less-than-50-mm difference from the observed precipitation in over 80% of the area (Figure 4a). ERA5-Land tended to overestimate the precipitation across the TP, except in the Qaidam Basin, with an average overestimation exceeding 150 mm (Figure 4b). The IMERG underestimated the precipitation significantly in the Qiangtang Plateau and the Qilian Mountains but overestimated it by more than 150 mm in the regions along the Himalayas and the Gangdise mountains (Figure 4c). Additionally, all three of the gridded datasets exhibited overestimations exceeding 300 mm in the Yarlung Zangbo River Canyon, which was likely due to a lack of station observations in that region.

The spatial pattern of summer precipitation trends over the last two decades for the various datasets is shown in Figure 5. Generally, the summer precipitation on the TP had an insignificant and weakly increasing trend. The Nianqing Tanggula Mountains, extending to the Hengduan Mountains, exhibited a reduction trend at a rate of approximately 1 mm/y. The CMFD indicated a significant decreasing trend in precipitation on the Qiangtang Plateau, with a rate of decrease of 8 mm/y (Figure 5b). Notably, the precipitation trend indicated by the CMFD in this region significantly deviated from that of the other datasets, which contrasts with the findings of Luan et al. [43]. This error may be due to the different study periods selected. Given the potential for interpolation errors due to the lack of observations on the Qiangtang Plateau, further verification of the CMFD precipitation trend in this region is necessary. The spatial trends in summer precipitation in ERA5-Land (Figure 5c) and the IMERG (Figure 5d) resembled the observed trends, but their increasing rates were notably higher (approximately 5 mm/y).

Figure 6 depicts the annual changes in precipitation from the station observations, CMFD, ERA5-Land, and IMERG. The summer precipitation changes from all four of the datasets consistently exhibited upward trends, but the CMFD, ERA5-Land, and IMERG all overestimated precipitation (overestimated by 12.29, 165.33, and 38.46 mm, respectively). Despite ERA5-Land’s overestimation of precipitation by at least 150 mm, its precipitation change pattern aligned with those of the other datasets, with its increasing rate being the closest to the observations. This suggests that ERA5-Land effectively captured informative precipitation changes on the TP, but grossly overestimated precipitation. The CMFD offered precipitation values closest to the observations among the three high-resolution datasets. However, both the CMFD and the IMERG overestimated the rate of summer precipitation increase on the TP in the past 20 y, with an average rate exceeding 1 mm/y.

3.2. Evaluation with Reference to the Original Stations from 2000 to 2018

3.2.1. Evaluation at an Annual Scale

Bilinear interpolation was used to obtain the mean annual summer precipitation for the three high-resolution datasets corresponding to the original 133 stations. These were compared with the station observations (Figure 7). The CMFD, ERA5-Land, and the IMERG overestimated the summer precipitation on the TP, with bias values of 4.37, 58.69, and 13.66%, respectively. The CMFD showed the greatest agreement with the observed precipitation, with a CC of 0.88, a bias of 4.37%, and an RMSE of 65.33 mm/y. ERA5-Land had the most unfavorable estimation, with nearly all of the stations overestimating precipitation to varying degrees. The IMERG had a CC of 0.75 and an RMSE of 102.92 mm/y.

3.2.2. Evaluation at a Monthly Scale

Table 2. Evaluation metrics of high-resolution precipitation datasets for summer months from 2000 to 2018.

Month	Evaluation Metric	CMFD	ERA5-Land	IMERG
Jun.	CC	0.932	0.692	0.846
	Bias (%)	2.375	56.006	8.113
	RMSE (mm/month)	19.921	71.310	30.099
Jul.	CC	0.837	0.624	0.743
	Bias (%)	4.996	58.462	15.523
	RMSE (mm/month)	34.519	90.711	46.040
Aug.	CC	0.868	0.633	0.736
	Bias (%)	5.494	61.442	16.676
	RMSE (mm/month)	28.488	80.821	43.258

lists the CC, bias, and RMSE values of the high-resolution precipitation datasets for each summer month. The CMFD maintained the greatest consistency with the directly observed precipitation on a monthly scale. The optimum performance was in June. The CC value of ERA5-Land was only 0.624 in July, with the largest RMSE among the three datasets. The IMERG was significantly more accurate in June. Additionally, none of the three datasets exhibited an apparent bias in the precipitation estimation in June, July, or August (figure omitted). The overall accuracy of the three datasets was ranked in the following order: June > August > July.

The Taylor diagram, illustrated in Figure 8, offers an insight into the correlation between the three precipitation datasets and the station observations during the summer months. In the diagram, four distinct points are employed to represent each dataset. The distance between each gridded dataset point and observation point reflects the magnitude of the RMSD. Here, σ is the distance between the gridded dataset points and the origin. As σ approaches 1, the distance between the two points diminishes, and the pinch angle approaches 0, indicating that the two datasets are more correlated [42]. The CMFD had the highest accuracy of the three datasets, with σ approximating 1 in all of the months, and had a CC greater than 0.8. The IMERG was second only to the CMFD, with a σ of approximately 1 and an RMSD of approximately 0.7–0.8. Similarly, its accuracy for precipitation estimation was optimal in June. ERA5-Land performed the most unfavorably on a monthly scale, with a σ greater than 1.25 and an RMSD of approximately 1, while its CC was 0.63 in both July and August. Generally, the CC of precipitation for each month was greater than 0.6, indicating good accuracy and stability.

3.2.3. Evaluation at a Daily Scale

Figure 9 presents box plots of the CC, bias, RMSE, POD, FAR, and CSI for each gridded dataset at a daily scale on the TP in summer. Notably, the correlation and accuracy of the three gridded datasets at the daily scale were inferior compared to their performance at the monthly scale. Specifically, the mean CC values for the CMFD, ERA5-Land, and IMERG were 0.5, 0.38, and 0.26, respectively. The considerable spread between the upper and lower CC bounds for the CMFD indicates substantial variability in the correlation between the CMFD and the observed data at individual stations. Regarding bias, all three of the gridded datasets displayed significant outliers, primarily skewed toward larger values on the right side of the distribution. Among the three precision evaluation metrics, the IMERG exhibited a notably lower accuracy at the daily scale compared to the annual and monthly scales. Regarding the precipitation capture capability, the POD values of the CMFD were primarily distributed from 0.8 to 0.9, and the mean value of FAR was less than 0.3, which was an accurate estimation of the precipitation events. The IMERG had the lowest POD values and the most severe underestimation of precipitation events among all of the datasets. Although ERA5-Land had the highest mean POD value, its FAR was higher than that of the other datasets. This suggests that ERA5-Land significantly overestimated the number of precipitation events during the summer months on the TP. The CMFD significantly outperformed ERA5-Land and the IMERG. In summary, the CMFD had the largest CC and CSI values and the smallest bias and RMSE values, rendering it the most accurate precipitation dataset on a daily scale.

The evaluation of the three gridded datasets in capturing daily precipitation events, referencing the original 133 stations, is presented in Figure 10. The POD values for the CMFD were above 0.8 at almost all of the stations, except for a few stations in the northwestern TP (Figure 10a). The spatial pattern of the POD values for ERA5-Land was similar to that of the CMFD, with the features being high in the south and low in the north (Figure 10b). The FAR values for ERA5-Land were almost above 0.3 or more, with the probability of incorrectly capturing precipitation events being greater than that of the other datasets (Figure 10e). Regarding the CSI, the CMFD had the highest mean CSI value, and its ability to capture precipitation was the best among the three datasets on the daily scale (Figure 10g). The CSI values for ERA5-Land (Figure 10h) and IMERG (Figure 10i) were both from 0.5 to 0.7. However, the ability of ERA5-Land to capture small precipitation events was better than that of the IMERG.

3.3. Evaluation with Reference to the Additional Stations from 2016 to 2018 at a Daily Scale

An additional 114 stations were added to the original 133 stations on the TP post-2016, totaling 247 stations. A comparison of the daily precipitation at all of the stations showed that more than 20 stations had a mean daily precipitation greater than 5 mm in summer, with certain individual stations recording values exceeding 20 mm. The average daily precipitation at stations near to the Qaidam Basin was less than 0.5 mm. These stations, with maximum and minimum daily precipitations, were located among the 114 newly added stations. A total of 3 gridded precipitation datasets at the corresponding 114 newly added stations were extracted using bilinear interpolation and evaluated on a daily scale.

Figure 11 displays the spatial patterns of the CC, bias, and RMSE for the daily scale precipitation data from the CMFD, ERA5-Land, and IMERG, interpolated to the additional 114 stations. It has been observed that the accuracy of ERA5-Land remained relatively consistent when compared to the reference with the original 133 stations. In contrast, the quality of the CMFD decreased notably. Specifically, the mean CC value for the CMFD (Figure 11a) decreased from 0.5 to 0.35, which was identical to that of ERA5-Land (Figure 11b). The CMFD (Figure 11d) and the IMERG (Figure 11f) showed an increase in stations with bias values of less than −30% compared to the previous assessments. The RMSE values for the three datasets showed a low-to-high gradient from north to south on the TP (Figure 11g–i).

The evaluation of the three gridded datasets in capturing daily precipitation events, referencing the newly added 114 stations, is presented in Figure 12. Notably, the POD values for ERA5-Land remained similar to those referenced for the original 133 stations (Figure 12b). Overall, ERA5-Land demonstrated a high capacity to capture precipitation events, surpassing the CMFD and the IMERG. The POD values of the IMERG significantly increased near to the Gangdise-Nianqing Tanggula Mountain Range (Figure 12c). In contrast, more than 30% of the CMFD stations had a POD value below 0.8 (Figure 12a), and the mean FAR value increased from 0.28 to 0.4 (Figure 12d), signifying a significant decrease in the precipitation capture capability. Regarding the CSI (Figure 12g–i), the areas with high precipitation detection accuracy for all three of the datasets corresponded to stations with substantial precipitation, while a lower accuracy was noted in the Qaidam Basin, which experiences less precipitation.

4. Discussion

4.1. Dataset Differences

Owing to the different retrieval algorithms and data sources, different precipitation products have variable applicability [44]. Meanwhile, the biased distributions of the precipitation products have high spatial heterogeneity. The evaluation results for the same product may vary, depending on factors such as topography, elevation, and precipitation [45,46,47].

The three datasets have strengths and weaknesses concerning spatial and temporal resolutions, coverage area, and data quality (

Table 3. Differences between the gridded precipitation datasets used in this study with their parameters.

Dataset	Spatio-Temporal Resolution	Temporal Coverage	Spatial Coverage	Addition
CMFD	0.1°, 3 h	1979 to 2018	China	release latency, CMA station calibration
ERA5-Land	0.1°, 1 h	1950 to present	Global land	release latency, GPCC calibration
IMERG	0.1°, 0.5 h	2000 to 2021	60°N-S	release latency, NCEP Stage IV calibration

). First, all three of the gridded datasets offer a high spatial resolution of 0.1°, which is well-suited for investigating complex terrains. In terms of temporal resolution, while the CMFD exhibits superior accuracy at annual, monthly, and daily scales, its temporal resolution is limited to 3 h. Therefore, for research requiring hourly scale data, it is advisable to select ERA5-Land or IMERG, which provide time resolutions of 1 and 0.5 h, respectively. Secondly, if the study region is primarily within China, the CMFD dataset is recommended due to its superior accuracy compared to satellite remote sensing products and reanalysis data. However, we note that ERA5-Land covers global land areas, but it may have data gaps at the land–ocean interface. In contrast, the IMERG is suitable for large-scale studies involving land–ocean precipitation at non-high latitudes. If up-to-date summer precipitation products for the TP are required, the CMFD dataset was last updated in 2018, whereas the IMERG and ERA5-Land provide more recent data. IMERG data have been available since 2000, although it lacks historical data, while ERA5-Land offers the longest temporal coverage, dating back to 1950, and continuing to the present day. Finally, it is essential to acknowledge that all three of the datasets have been calibrated with other precipitation data and released latency, rendering them unsuitable for near-real-time hydro-geo-climatological studies [19].

4.2. Observational Limitations

In this study, notable discrepancies in the spatial and temporal variations of multi-source precipitation datasets were primarily concentrated in the Qiangtang Plateau and the Yarlung Zangbo River Canyon. This phenomenon can be attributed to the absence of observation stations in these regions. The spatial scale mismatch between the sparse stations and high-resolution gridded datasets makes it difficult for individual stations to characterize regional precipitation [5,48]. In contrast, interpolating sparsely distributed observations onto the high-resolution grids tends to introduce substantial errors. To ascertain the reliability of the gridded precipitation datasets in these data-scarce areas, 10 rain gauge stations on the Qiangtang Plateau were selected.

Table 4. Precipitation estimations at gauge stations on the Qiangtang Plateau during the summer of 2018 from the three gridded datasets.

Station	Latitude (°E)	Longitude (°N)	OBS (mm)	CMFD (mm)	ERA5-Land (mm)	IMERG (mm)
EW01	32.51	79.68	88.6	95.4	286.1	139.6
EW03	32.37	80.61	99.6	93.5	246.3	177.3
EW05	32.21	81.56	181.2	211.4	332.3	200.2
EW06	32.22	82.26	163.6	233.7	232.0	321.7
EW07	32.53	82.59	114.2	247.2	250.0	400.6
EW08	32.4	83.42	173.4	252.8	330.2	346.5
EW09	32.29	84.07	166.0	196.3	286.8	237.1
EW11	32.02	85.48	250.8	203.4	281.9	376.0
EW16	31.87	88.15	296.8	283.3	457.3	436.9
SN28	32.38	89.15	312.0	289.5	371.5	332.5

presents the precipitation estimation at 10 rain gauge station locations for the summer of 2018, utilizing three distinct gridded datasets. Among these datasets, the CMFD demonstrated the highest level of accuracy in estimating precipitation.

Unfortunately, despite utilizing the most comprehensive and accurate observations as references for evaluation, substantial portions of the northwestern TP still lack observed data. Given the incorporation of the gridded data values corresponding to the observation stations, the lack of observations in the northwestern TP did not significantly affect the results of the evaluation on multiple timescales. Nonetheless, the spatial and temporal variability patterns of the gridded datasets in this region require further verification.

4.3. Error Analysis

The calibration of the three high-resolution gridded datasets using distinct precipitation data sources is the primary factor for the discrepancy, which is numerous among the 133 CMA stations that participated in both the production and the calibration processes of the CMFD [20]. ERA5-Land assimilated NCEP Stage IV precipitation, while the IMERG used GPCC monthly precipitation product correction. Both of these datasets were produced with the rare use of CMA stations [11,16]. Although we did not exclude these stations (it was not possible to determine which stations were involved in the production of the three datasets), the new 114 stations (added post-2016) certainly did not participate in the production. Consequently, referencing solely the newly added stations mitigated the impact of the CMA stations involved in calibrating the gridded precipitation datasets. On the daily scale, the evaluation results with the newly added stations as references showed that ERA5-Land and the IMERG maintained a stable performance, while the CMFD demonstrated a significant decrease in both accuracy and precipitation capture capability (CC and CSI decreased by 0.18 and 0.1, respectively). This phenomenon suggested that the calibration of CMFD datasets using CMA station data plays a crucial role in determining their quality. Additionally, the IMERG product was significantly affected by the GPCC monthly precipitation data calibration [49,50], resulting in markedly inferior daily scale performance compared to its monthly scale performance.

Figure 6 shows that the CMFD, ERA5-Land, and IMERG overestimate precipitation by 12.29, 165.33, and 38.46 mm, respectively. Combined with Figure 4, the overestimated precipitation originates primarily in the southern and southeastern TP, where the topography is more undulating. Notably, the rationales for the overestimations in this region may vary among the gridded datasets. The deficiencies in the cumulus parameterizations on the steep terrain led to the overestimation in ERA5-Land [51]. The satellites may have mistaken the rising water vapor with the terrain for precipitation signals, leading to the overestimation of the IMERG [52]. The CMFD benefitted from calibration by having more CMA stations in the region, performing better than the other datasets.

Meanwhile, there were other distinctive features of the satellite, and the reanalyzed precipitation products affected the capability of each dataset to capture the precipitation events. The low accuracy of the CMFD and IMERG in detecting precipitation in the regions with low precipitation reflected the shortcomings of the two datasets in capturing minor precipitation events. Previous studies have shown that precipitation of less than 0.2 mm/h is difficult to detect, due to deficiencies in satellite observing equipment [39,53]. Hersbach et al. [54] highlighted a notable characteristic of ERA5-Land, which tends to provide non-zero estimates, leading to the incorrect recording of several precipitation events that never occurred. This behavior could account for the general overestimation of precipitation across the TP and the resulting high values for both POD and FAR associated with ERA5-Land. Additionally, the deviation of the grid center coordinates in the IMERG and ERA5-Land precipitation datasets was 0.05° [19]. Notably, as the CMA stations are not very dense on the TP, the evaluation error caused by a 0.05° deviation in the grid center coordinates could be almost negligible.

The above discussions can facilitate the selection of appropriate precipitation datasets for various applications. It is essential to acknowledge that various gridded datasets hold significant potential for enhancement. For instance, the CMFD should incorporate more CMA stations for calibration in upcoming updates. Improvements in data assimilation and physical process parameterization are necessary in order to elevate the quality of ERA5-Land. The use of calibration precipitation data sources on daily and hourly scales can be considered for the IMERG to further enhance its applicability on these scales.

5. Conclusions

High-resolution precipitation datasets play a crucial role in enhancing climate predictions for the TP. This study conducted a comprehensive evaluation of the CMFD, ERA5-Land, and IMERG high-resolution gridded precipitation datasets at annual, monthly, and daily scales using TP station observations from the summers of 2000 to 2018 as a reference. The key findings are as follows:

(1) Spatially, significant variations in the estimated precipitation were observed among the WTP datasets, where precipitation observations were limited. The CMFD exhibited the closest agreement with the station observations, while ERA5-Land generally overestimated precipitation (>150 mm). In areas with a lack of observation stations, all three of the gridded datasets showed a substantial overestimation of precipitation, exceeding 300 mm in the lower Yarlung Zangbo River region.

(2) On annual and monthly scales, the quality of the three gridded summer precipitation products for the TP ranked as follows: CMFD > IMERG > ERA5-Land. When considering individual months, June exhibited the highest accuracy, followed by August and July. However, all three of the datasets performed less accurately on a daily scale compared to a monthly scale, with the CMFD being the most accurate and the IMERG showing the lowest accuracy at this scale (CC of 0.51 and 0.26, respectively).

(3) On a daily scale, when assessing the datasets with reference to the original 133 stations and the 114 new meteorological stations established after 2016, ERA5-Land and the IMERG demonstrated a relatively stable performance, while the CMFD showed a notable decrease in both accuracy and precipitation capture capability (CC and CSI decreased by 0.18 and 0.1, respectively).

As of now, the CMFD stands out as the most accurate dataset for recording the summer precipitation characteristics on the TP across various and multiple scales. However, the meteorological satellites and numerical models are continually updated, while more precipitation observations are being integrated into the TP. Therefore, future evaluations should focus on the latest precipitation products, especially their application on the northwest TP.