Evaluation of the Accuracy and Applicability of Reanalysis Precipitation Products in the Lower Yarlung Zangbo Basin

Abstract

The lower Yarlung Zangbo River Basin’s Great Bend region, characterized by extreme topography and intense orographic precipitation processes, presents significant challenges for accurate precipitation estimation using reanalysis products. Therefore, this study evaluates four widely used products (ERA5-Land, MSWEP, CMA, and TPMFD) against station observations (2014–2022) in this critical area. Performance was rigorously assessed using correlation analysis, error metrics (RMSE, MAE, RBIAS), and spatial regression. The region exhibits strong seasonality, with 62.1% of annual rainfall occurring during the monsoon (June-October). Results indicate TPMFD performed best overall, capturing spatiotemporal patterns effectively (correlation coefficients 0.6–0.8, low RBIAS). Conversely, ERA5-Land significantly overestimated precipitation, particularly in rugged northeast areas, suggesting poor representation of orographic effects. MSWEP and CMA underestimated rainfall with variable temporal consistency. Topographic analysis confirmed slope, aspect, and longitude strongly control precipitation distribution, aligning with classical orographic mechanisms (e.g., windward enhancement, lee-side rain shadows) and monsoonal moisture transport. Spatial regression revealed terrain features explain 15.4% of flood-season variation. TPMFD most accurately captured these terrain-precipitation relationships. Consequently, findings underscore the necessity for terrain-sensitive calibration and data fusion strategies in mountainous regions to improve precipitation products and hydrological modeling under orographic influence.

1. Introduction

The Yarlung Zangbo River, as a vital component of the Tibetan Plateau known as the “Water Tower of Asia,” provides water resources for nearly 2 billion people and significantly influences regional climate patterns and water resource distribution [1,2,3,4,5,6]. Due to the complex topographic and climatic conditions in the basin, particularly in the Great Bend downstream area, precipitation patterns are characterized by intricate formation mechanisms and extremely variable spatial-temporal distribution [7,8,9,10,11,12]. The region experiences diverse precipitation processes influenced by monsoon systems, orographic effects, and local circulation patterns, making accurate precipitation estimation particularly challenging [13,14,15,16].

The lower Great Bend region of the Yarlung Zangbo River faces severe data scarcity challenges due to complex terrain, frequent disasters [17,18,19,20,21], large elevation differences [22,23], and transportation difficulties, resulting in a long-standing lack of meteorological and hydrological observation data [24,25]. Existing research primarily relies on limited representative stations (such as Lazi and Zedang) to explore precipitation-runoff relationships and hydrological responses under climate change, but the sparse distribution of stations creates significant uncertainty in spatial-scale analysis results [26,27]. Under these circumstances, reanalysis precipitation products have emerged as crucial alternative data sources, providing spatiotemporally continuous and comprehensive precipitation data by integrating multi-source remote sensing data, ground observations, and numerical model simulations.

The development of satellite and reanalysis precipitation products has significantly improved our ability to monitor precipitation patterns in remote areas, becoming the primary data source for hydrological studies in regions lacking comprehensive observation networks. For the Great Bend region of the Yarlung Zangbo River, due to the absence of a complete runoff observation network, it is difficult to directly obtain actual runoff volumes and flood characteristics, making accurate precipitation products the only reliable basis for estimating regional runoff and assessing flood risks. However, the accuracy and applicability of these products vary significantly across different regions and time scales [28,29,30,31], with variations primarily stemming from complex terrain, climate variability, and limitations in data retrieval algorithms [32,33]. In recent years, various reanalysis precipitation datasets such as ERA5-Land, MSWEP, CMA, and TPMFD have been widely used in hydrological and climate research on the Qinghai–Tibet Plateau [34,35,36,37,38]. Recent methodological advances have introduced comprehensive evaluation frameworks, such as the entropy-based weighting approach, combined with novel spatial metrics (FMS, SCA, PSDI) for evaluating precipitation products over Eastern China, providing valuable insights for understanding precipitation product performance from spatial distribution perspectives [39]. Studies evaluating the accuracy of reanalysis precipitation products continue to deepen [40,41,42,43,44,45], showing significant differences in product performance across regions, elevations, and time scales, with multi-source fusion products like MSWEP generally outperforming single-source products, while ERA5-Land, TPHiPr, and CMFD demonstrate excellent performance in specific contexts, emphasizing the importance of accuracy assessment for specific regions.

The complex terrain and elevation gradients in the lower Great Bend region significantly affect the performance of reanalysis precipitation products, as interactions between monsoon systems and local topography generate unique precipitation patterns that are extremely difficult to accurately capture [46,47,48,49]. The monsoon-dominated precipitation patterns add another layer of complexity to accurate precipitation estimation [50,51,52], while the relationship between mountain precipitation patterns and geographical and circulation factors has been extensively studied, with elevation and slope aspect having significant effects on precipitation distribution in the Himalayan region [53,54,55]. Topographical factors significantly influence the estimation performance of satellite precipitation products in plateau regions, with satellite precipitation products lacking effective precipitation detection capabilities in high-elevation areas (>4500 m) of the southern Qinghai–Tibet Plateau [56]. Additionally, precipitation estimation results show significant correlation with terrain height, with performance particularly affected by topographical variations. The lower Great Bend region faces significant uncertainty in precipitation reanalysis product assessment due to complex terrain and scarce observational data [57,58], severely constraining hydrological simulation accuracy and water resource management decisions. Although some studies have evaluated precipitation products in different regions of the Qinghai–Tibet Plateau, few studies have specifically focused on the lower reaches of the Yarlung Zangbo River, where terrain and climate patterns are especially complex. Therefore, in-depth research on the performance of precipitation products in this region and their relationship with geographical factors is of great significance for improving the accuracy of precipitation estimation.

This study aims to (1) conduct the first systematic evaluation of the accuracy of four reanalysis precipitation products (ERA5-Land, MSWEP, CMA, and TPMFD) in the lower Great Bend region of the Yarlung Zangbo River based on newly acquired measured data, analyzing their performance at different time scales (annual, monthly); (2) study in depth the relationship between precipitation patterns and geographical factors; (3) evaluate and reanalyze the characterization ability of precipitation products for spatial heterogeneity under different geographical conditions. The research results will enhance the accuracy of evaluating the performance of different reanalyzed precipitation products in this region and provide valuable scientific basis for selecting appropriate precipitation products for hydrological research and water resource management in this region.

2. Data and Methods

2.1. Study Area

The Yarlung Zangbo River is the world’s highest international river, originating from the Jema Yangzong Glacier in Tibet, China, flowing through China, India, and Bangladesh, and finally emptying into the Bay of Bengal (Figure 1). The river has a total length of 2057 km and a basin area of 242,000 square kilometers, with approximately one-third of the basin located within China’s borders. The multi-year average runoff of the Yarlung Zangbo River Basin is 166.1 billion cubic meters. The precipitation-runoff processes exhibit diverse spatiotemporal distribution characteristics due to topographic features, monsoon climate, and warm humid airflow from the Indian Ocean [59,60,61]. In the lower basin, areas northwest of the Niyang River estuary, the Yigong Zangbo River estuary and the Palong Zangbubhumi line belong to temperate humid areas, while southern areas are subtropical humid regions [62,63]. The vertical climate differences and local climate effects make precipitation gradient changes more dramatic than in the middle and upper reaches, with significant spatial heterogeneity [64].

The climate in the lower reaches exhibits typical monsoon characteristics, with significantly increased summer precipitation due to warm humid airflow from the Indian Ocean, and decreased winter precipitation under cold northern airflow influence. Precipitation variations are regulated by both monsoon intensity and local topographic effects. The precipitation processes and hydrological characteristics are closely related in the lower reaches, with complex spatial differences in seasonal changes, spatiotemporal distribution, and precipitation intensity. Currently, constrained by regional hydrometeorological data conditions, research on the spatiotemporal characteristics and evolutionary patterns of key hydrometeorological elements is insufficient, and understanding of the hydrometeorological characteristics and runoff generation and confluence patterns in the Great Bend region of the lower Yarlung Zangbo River is weak. Therefore, it is necessary to select and construct precipitation datasets suitable for the Great Bend region from many existing datasets and to explore the relationship between precipitation distribution and geographical factors.

2.2. Data

This study utilizes four reanalysis precipitation products alongside ground-based observational data, covering various temporal scales and spatial resolutions to ensure the reliability and accuracy of the results. The ground observations were obtained from 18 local meteorological and rain gauge stations in the lower Yarlung Zangbo River Basin. These data underwent strict quality control procedures, including climatological limit checks, station extreme value checks, and spatiotemporal consistency checks. It should be noted that meteorological and most rain gauge stations in the lower Yarlung Zangbo Basin are located in river valleys and canyons due to accessibility and logistical constraints. This distribution pattern may introduce spatial representativeness limitations, as precipitation in mountainous regions exhibits significant spatial variability due to orographic effects, elevation gradients, and local topographic influences. The observed precipitation data may underrepresent the rainfall patterns in higher elevation areas and steep slopes where stations are difficult to install and maintain.

The reanalysis precipitation datasets include the Multi-Source Weighted-Ensemble Precipitation (MSWEP), the European Centre for Medium-Range Weather Forecasts Reanalysis 5-Land (ERA5-Land), the global land surface reanalysis products of the China Meteorological Administration (CMA-RA/Land) dataset, and the Third Pole long-term high-resolution surface meteorological dataset (TPMFD). The basic characteristics and attributes of each dataset are presented in

Table 1. Summary of five gridded precipitation products to be evaluated in this study.

Products	Temporal Resolution	Spatial Resolution	Temporal Coverage	Domain	Download Website
ERA5-Land	monthly	0.1° × 0.1°	1940–present	Global	https://cds.climate.copernicus.eu/datasets (accessed on 18 February 2025)
MSWEP V2	3-hourly	0.1° × 0.1°	1979–present	Global	https://www.gloh2o.org/mswep/ (accessed on 29 November 2024)
CMA-RA/Land	3-hourly	0.25° × 0.25°	1979–2020	Global	http://data.cma.cn/ (accessed on 9 December 2024)
TPMFD	1-hourly, daily, monthly	0.01° × 0.01°	1979–2023	Third Pole	https://www.tpdc.ac.cn/home (accessed on 7 July 2024)

.

ERA5-Land, developed by the European Centre for Medium-Range Weather Forecasts (ECMWF), is a high-resolution reanalysis dataset specifically designed to provide meteorological data over global land surfaces. With a spatial resolution of 0.1° × 0.1° and hourly temporal resolution, it is widely applied in hydrology, agriculture, and climate research [65].

MSWEP product estimates global precipitation with high temporal and spatial resolution by integrating multiple sources of observational and reanalysis data [34]. MSWEP is known for its high accuracy and broad spatial coverage, making it widely used in hydrological and climate studies.

CMA-RA/Land is the first-generation global land reanalysis product developed by the China Meteorological Administration, jointly produced by the National Meteorological Information Center and several other institutions. This dataset integrates a range of core technologies, including ensemble assimilation algorithms, multi-source data fusion methods, the Noah 3.3 land surface model, and optimized surface parameterization [66].

TPMFD [67] combines short-term high-resolution weather research and forecasting (WRF) model simulations, long-term ERA5 data, and station observations to provide high-accuracy meteorological variables, including precipitation, temperature, and specific humidity [68,69]. These data support climate analysis in the Third Pole region and offer high-quality inputs for land surface, hydrological, and ecological modeling [70].

Precipitation estimates in these data-scarce regions may exhibit certain biases, particularly in complex terrain such as mountainous or plateau areas, where their accuracy may be constrained by topographic effects.

The topographic data used for calculating terrain factors were derived from SRTM DEM (SRTM 3 Arc-Second Global, obtained from https://earthexplorer.usgs.gov/ (accessed on 3 November 2023)), with a spatial resolution of 90 m [71]. All topographic variables were extracted at the exact coordinates of each rain gauge station and grid using ArcGIS 10.8 and Python 3.12 processing tools.

2.3. Evaluation Methods

Directly interpolating meteorological station data into spatial datasets may introduce certain errors [72]. Therefore, this study adopts a point-to-point approach to establish correspondence between datasets [73]. Specifically, the precipitation value from each gridded dataset is extracted based on the coordinates of the meteorological stations. By using station observations as the reference, the accuracy and quality of each reanalysis precipitation dataset are evaluated through comparative analysis.

For spatial visualization and analysis of station-based precipitation patterns, the IDW interpolation was performed using ArcGIS 10.8 with a power parameter of 3. The interpolation resolution was set to 0.01° × 0.01° to ensure adequate spatial detail while maintaining computational efficiency. All available meteorological stations within the study area were included in the interpolation process.

To investigate the similarity in precipitation processes among datasets from different sources, as well as their temporal variability and spatial differences, the precipitation data were analyzed on both annual and monthly scales. Four quantitative evaluation metrics were employed to assess the accuracy of the reanalysis products: Pearson correlation coefficient (R), root mean square error (RMSE), mean absolute error (MAE), and relative bias (RBIAS). The R indicates the degree of linear correlation between reanalysis data and station observations; a value closer to 1 or −1 implies stronger correlation. RMSE measures the dispersion between the reanalysis and observed data, being particularly sensitive to large deviations—larger values indicate greater disparity, with the optimal value being 0. MAE reflects the average magnitude of absolute errors between the reanalysis and observed data, avoiding the issue of error cancellation; its optimal value is also 0. RBIAS provides a direct measure of the systematic error and relative deviation between reanalysis and observed precipitation, with an ideal value of 0. The formulas for each metric are as follows:

$$\mathrm{R}={\displaystyle \frac{Cov({\mathrm{PE}}_i-{\mathrm{GND}}_i)}{{\mathsf{\sigma }}_{PE}\cdot {\mathsf{\sigma }}_{GND}}}$$

(1)

$$\mathrm{RMSE}=\sqrt{{\displaystyle \frac{1}{N}}{\displaystyle \sum _{i=1}^N({\mathrm{PE}}_i-{\mathrm{GND}}_i}{)}^2}$$

(2)

$$\mathrm{MAE}={\displaystyle \frac{1}{N}}{\displaystyle \sum _{i=1}^N\left|{\mathrm{PE}}_i-{\mathrm{GND}}_i\right|}$$

(3)

$$\mathrm{RBIAS}={\displaystyle \frac{{\displaystyle \sum _{i=1}^N({\mathrm{PE}}_i-{\mathrm{GND}}_i})}{{\displaystyle \sum _{i=1}^N{\mathrm{GND}}_i}}}\times 100\%$$

(4)

In these formulas, N represents the length of the precipitation time series; $\mathrm{P}{\mathrm{E}}_i$ and $\mathrm{G}\mathrm{N}{\mathrm{D}}_i$ denote the reanalysis precipitation data and ground station precipitation data, respectively; ${\sigma }_{PE}$ and ${\sigma }_{GND}$ correspond to the square root of the sum of squares of the difference between the first two and their mean values, respectively.

To understand the relationship between precipitation and geographical factors in the topographically complex study area, we conducted spatial correlation analysis considering seven geographical factors: Altitude was the elevation extracted from digital elevation model at each rain gauge location. Slope was the slope gradient derived from DEM. Aspect was calculated as the compass direction (0–360°) that slopes face, derived from the digital elevation model. In our statistical analysis, aspect values were treated as a continuous variable representing slope orientation. Terrain undulation, or relative elevation difference, refers to the difference between the highest point within a radius of 5 miles from a point and the observation point’s elevation. Openness refers to the fan-shaped area within a 15-mile radius of the measuring point where there are no obstacles more than 1000 feet higher than the station [74]. Latitude and longitude were the geographic coordinates of each rain gauge station.

Multiple linear regression analysis was employed to quantify relationships between flood-season precipitation and geographical factors, incorporating spatial weight matrices based on the Euclidean distance between station coordinates to account for spatial autocorrelation effects. The same regression approach was applied to each precipitation product to evaluate its ability to capture spatial heterogeneity and terrain-precipitation relationships compared to station observations. Figure 2 shows the methodological diagram of the evaluation methods.

3. Results

3.1. Overall Accuracy of Precipitation Dataset

The precipitation in the downstream area of the Yarlung Tsangpo River exhibits significant seasonal differences, with flood season (May-October) precipitation making an especially prominent contribution to the annual total precipitation and extreme precipitation events occurring frequently during the flood season. Based on measured data statistics from 2014 to 2022, the average flood season precipitation in the study area accounts for 62.09% of the annual total precipitation and can reach as high as 73.80%, a conclusion that can also be found in Panday et al. [75] and Palash et al. [76].

This concentrated precipitation pattern leads to two key assessment needs: evaluating reanalysis products’ ability to capture heavy precipitation events during flood season and verifying products’ capability to characterize precipitation gradients influenced by terrain. Additionally, annual precipitation variations are primarily driven by flood season anomalies, making data accuracy during this period crucial for hydrological modeling reliability. Figure 3 provides scatter plots comparing observations from 18 ground stations with various precipitation datasets at annual, monthly, and flood-season scales, with the dashed line representing the 1:1 line.

Annual precipitation analysis (Figure 3a) reveals ERA5-Land significantly overestimates precipitation, particularly in the 1000–3000 mm/year range. CMA consistently underestimates precipitation throughout the entire range. MSWEP shows relatively good agreement with station data despite considerable dispersion. TPMFD performs comparatively well, with data points clustering closer to the reference line than ERA5-Land, despite some overestimation.

Monthly scale analysis (Figure 3b) shows similar patterns. ERA5-Land continues its systematic overestimation, with higher dispersion in the 0~200 mm/mon range. MSWEP demonstrates improved performance, with better agreement with observations. CMA maintains its systematic underestimation, particularly at low precipitation levels (0~300 mm/mon). TPMFD shows stable performance with high consistency with observational data.

Flood season precipitation (Figure 3c) analysis reveals significant differences among products. ERA5-Land exhibits obvious positive bias across the entire range, with particularly severe overestimation in high precipitation situations. Its flood season positive bias (+42.3%) significantly exceeds its non-flood season bias (+18.7%). CMA shows persistent underestimation, with its flood season error contribution rate (−68.9%) being 2.2 times that of the non-flood season (−31.1%), indicating fundamental deficiencies in characterizing monsoon precipitation. MSWEP performs moderately with data points showing medium dispersion and slight underestimation in the 1000~2000 mm range. TPMFD demonstrates the most balanced performance, clustering tightly around the reference line in the 1000–2500 mm interval, with particularly strong performance in terrain-induced heavy precipitation areas (annual precipitation > 3000 mm).

Based on the reanalysis precipitation products and the annual precipitation data of measured stations from 2014 to 2022, the multi-year average was calculated (Figure 4).

The divisions shown in Figure 4 represent sub-basins, which were delineated based on hydrological modeling. Spatial distribution analysis shows significant differences among products. Based on interpolated station data (Figure 4f), precipitation forms a local high-value center in the west-central study area, gradually decreasing northeastward, correlating with the regional “V”-shaped valley topography. Considering the different temporal coverage periods (2014–2022 for ERA5-Land, MSWEP, TPMFD, and station data; 2014–2020 for CMA data), ERA5-Land (Figure 4a) and TPMFD (Figure 4d) overestimate precipitation by approximately 35.22% and 11.9%, respectively, while MSWEP (Figure 4b) and CMA (Figure 4c) underestimate by 52.0% and 63.9%. ERA5-Land shows the most significant spatial variation, with precipitation decreasing from southwest to northeast, following the topographic gradient. TPMFD exhibits obvious spatial heterogeneity, with more pronounced local precipitation differences. MSWEP and CMA display more homogeneous spatial distribution, with less variability, potentially underestimating precipitation differences in complex topography.

Combined with DEM elevation data (536~7261 m) (Figure 1), the analysis shows that ERA5-Land and TPMFD are more sensitive to topography-induced precipitation variations, potentially exaggerating valley-mountain precipitation differences. Other products may not fully reflect topographic influences due to resolution limitations or retrieval algorithm constraints. Overall, each product shows varying degrees of uncertainty in spatial distribution and precipitation estimation, likely related to data sources, spatial resolution, retrieval algorithms, and methods for handling topographic influences.

3.2. Evaluation at Anunal Scale

The temporal evolution characteristics of four reanalyzed precipitation products on the annual scale (Figure 5) were evaluated using the annual precipitation of the stations from 2014 to 2022 and the error indicators (R, RMSE, MAE, RBIAS). The analysis reveals distinct performance patterns among the different products.

Correlation analysis (Figure 5a) indicates that TPMFD (yellow line) demonstrates the most stable temporal consistency, maintaining R values between 0.6 and 0.8 throughout the study period, with only a slight decrease to 0.55 in 2017. CMA (blue line) performed well in 2014–2015, followed by a fluctuating downward trend, reaching its lowest point in 2018 before recovering and stabilizing. ERA5-Land (black line) shows obvious fluctuations, peaking in 2015 and 2020, but with an overall unstable trend. MSWEP (red line) performed the worst, with consistently negative R values, indicating that its annual precipitation variation trend is negatively correlated with the measured annual precipitation.

Figure 5b shows the temporal variation of RMSE for each product. Overall, TPMFD has the smallest error, with annual RMSE remaining relatively stable between 600 and 800 mm/yr, indicating its high accuracy in the Great Bend region. CMA and MSWEP rank second, with RMSE fluctuating approximately between 1000 and 1800 mm/yr, with MSWEP showing an error peak in 2020. ERA5-Land has the largest error, with dramatic fluctuations, reaching a maximum value of 2528 mm/yr in 2019, followed by a downward trend. Figure 5c shows that the MAE variation trend is basically consistent with RMSE. TPMFD performs best, with MAE stable at around 500~700 mm/yr. CMA and MSWEP are intermediate, while ERA5-Land has the largest error, with notable fluctuations. It is worth noting that all products experienced error peaks between 2019 and 2020, which may be related to the special hydrometeorological conditions in the lower reaches of the Yarlung Zangbo River during this period.

The RBIAS analysis (Figure 5d) reveals systematic estimation tendencies. ERA5-Land consistently exhibits a pronounced positive bias, ranging from 40% to 120%, with peak values during 2018–2019. TPMFD maintains the smallest bias, remaining within ±20% throughout the study period. MSWEP and CMA show persistent negative biases, generally ranging from −40% to −80%.

By calculating the error indicators between the reanalysis precipitation products’ annual precipitation and the annual precipitation of various stations, the multi-year average value of the error indicators of each station is obtained, and the spatial distribution of the annual precipitation error of reanalysis precipitation products is further explored. The spatial distribution of correlation coefficients between the four reanalysis precipitation products and station observations exhibits distinct patterns (Figure 6).

In the study area, TPMFD exhibits good spatial consistency (Figure 6d), with correlation coefficients primarily ranging between 0.7 and 1.0, particularly strong in the western and central regions (R > 0.8). CMA shows similarly robust performance (Figure 6c), with most correlation coefficients above 0.7. MSWEP (Figure 6b) and ERA5-Land (Figure 6a) display moderate performance, with average correlation coefficients of 0.47 and 0.36, respectively. All products exhibit significant spatial heterogeneity—stations in the western part show higher correlations than those in the eastern part.

The spatial pattern of RMSE (Figure 7) further validates the performance characteristics revealed by the correlation analysis. TPMFD, which had the highest correlation coefficients, also demonstrated the best performance in terms of RMSE, with an average of 663.70 mm/yr. Its RMSE values remained relatively low (100~700 mm/yr) at most stations, particularly in the western and central regions, indicating consistent accuracy in precipitation estimation. Among all products, ERA5-Land exhibited the highest average RMSE (1832.89 mm/yr), with particularly large errors (2900~3500 mm/yr) in the northeastern areas, where correlation coefficients were lower. In contrast, RMSE values were relatively lower (700~1200 mm/yr) at western stations with higher correlations. MSWEP showed relatively good performance, with an average RMSE of 1081.52 mm/yr. RMSE values were lower (700~1200 mm/yr) in the western region, where correlations were higher. CMA had moderate accuracy, with an average RMSE of 1280.57 mm/yr, but its spatial distribution differed from the other products. Despite having the second highest average correlation coefficient, some stations in the central region exhibited unexpectedly high RMSE values (2300~2900 mm/yr), suggesting that although CMA effectively captured temporal variability (high R values), it may still exhibit significant magnitude bias at certain locations.

The spatial distribution of relative deviations (Figure 8 and Figure 9) shows that ERA5-Land exhibits a pronounced systematic overestimation (average RBIAS of 101.40%), with particularly large positive biases (280~350%) in the northeastern region, corresponding to areas with higher RMSE values and lower correlation coefficients. TPMFD exhibits the most balanced performance (average RBIAS of 8.80%), indicating the smallest system bias, with a mixed pattern of slight overestimation (40% of stations) and underestimation (60% of stations). This balanced performance is consistent with its superior correlation coefficients and low RMSE values across the study area.

The other two products both underestimate precipitation, but the degree and spatial pattern were different. CMA shows a high correlation coefficient; its average RBIAS is −54.51%, indicating a significant underestimation. All stations (100%) show negative bias, and the underestimation is relatively evenly distributed across the entire study area. CMA shows high correlation but significant underestimation (average RBIAS: −54.51%), with all stations showing negative bias relatively evenly distributed across the study area. MSWEP shows moderate underestimation (average RBIAS: −29.21%), with 82% of stations showing negative bias, performing slightly better at western stations.

3.3. Evaluation at Monthly Scale

The time evolution characteristics of four reanalysis precipitation products on the monthly scale (Figure 10) were evaluated by the monthly precipitation of the stations from 2014 to 2022 and error indicators, in order to reveal the obvious seasonal patterns and product-specific characteristics in precipitation estimation.

The correlation coefficient (R) shows significant seasonal variations across all products (Figure 10a). TPMFD exhibits the most stable and highest correlations throughout the year, with values ranging between 0.3 and 0.7, peaking in July (approximately 0.7), and showing relatively lower values during winter months. CMA displays a similar seasonal pattern but fluctuates between −0.2 and 0.6, with the highest correlations occurring during the monsoon season (June to September). ERA5-Land shows moderate correlations in summer and performs poorly in other seasons, with R values ranging between −0.1 and 0.3. MSWEP performs most unstably, maintaining correlation coefficients around 0.5 only from July to October, indicating significant challenges in capturing precipitation patterns during the post-monsoon period.

RMSE analysis (Figure 10b) similarly shows distinct seasonal cycles for all products, with larger errors during the monsoon season (June–September) and smaller errors in winter months. TPMFD maintains relatively stable low RMSE (50~150 mm/mon) throughout the year, with comparatively small seasonal variations. ERA5-Land and CMA exhibit similar seasonal characteristics during the monsoon season, with moderate RMSE values (150~250 mm/mon). MSWEP demonstrates medium performance, with RMSE ranging from 100 to 200 mm/mon. The MAE patterns (Figure 10c) largely reflect the RMSE trends.

The RBIAS analysis (Figure 10d) reveals distinct systematic biases among products. ERA5-Land shows consistent overestimation throughout the year, with RBIAS ranging between 30% and 210%, with particularly high values during winter months. In contrast, most other products exhibit negative biases with different seasonal patterns. MSWEP and CMA present relatively stable negative biases (−30%~−80%) throughout the year, while TPMFD demonstrates the most balanced performance, with RBIAS fluctuating around ±20%, showing slight overestimation in winter and slight underestimation in summer.

The spatial distribution of monthly-scale correlation coefficients present different patterns compared to the annual-scale analysis (Figure 11). CMA exhibits the highest average correlation (R = 0.76), followed by MSWEP (R = 0.72) and TPMFD (R = 0.70), while ERA5-Land (R = 0.62) performs relatively poorly. CMA exhibits strong correlations across the entire study area (R = 0.6~0.9 for most stations), showing significant improvement compared to its annual performance, especially in western regions. MSWEP’s spatial distribution is relatively uniform, with R values above 0.6 for most stations. TPMFD maintains stable performance similar to the annual scale (R = 0.6~0.9). ERA5-Land shows significant improvement at monthly scale compared to annual, with R values between 0.4 and 0.8 for most stations. All products show slightly lower correlation coefficients (0.4~0.6) at a few southwestern stations.

The spatial heterogeneity of RMSE values was influenced by terrain complexity and precipitation variability (Figure 12). TPMFD demonstrates superior performance, with RMSE values below 140 mm/mon at most stations except for a few in the southwestern region, which aligns well with its stable correlation coefficient (R = 0.70). MSWEP ranks second in overall performance, with its distribution of lower RMSE showing strong correlation (R = 0.72), especially in western and central regions.

However, some eastern stations display higher errors, suggesting that significant biases may still exist in this product’s quantitative precipitation estimates. CMA has the highest monthly correlation coefficient (R = 0.76) but shows higher RMSE values (280~350 mm/mon) in the western region, indicating effective capture of precipitation temporal patterns but potential systematic biases in quantitative estimation. ERA5-Land exhibits similar characteristics with CMA.

The spatial distribution of RBIAS (Figure 13 and Figure 14) reveals different patterns of systematic errors. ERA5-Land exhibits significant positive bias, with overestimation at 94% of stations, stronger in northern regions compared to southern regions. MSWEP and CMA display systematic underestimation, with all stations showing small negative biases. TPMFD demonstrates the most balanced performance, with 33% of stations showing overestimation and 67% showing underestimation, with slight overestimation at some northern stations and slight underestimation elsewhere.

3.4. Precipitation-Geography Correlations and Product Performance

Unlike other basins, the downstream area of the Yarlung Zangbo River has special geographical conditions, where surface conditions, primarily terrain, have a greater impact on precipitation. Therefore, to gain a deeper understanding of the relationship between precipitation and different geographical factors in the downstream area of the Yarlung Zangbo River, this study analyzed the ability of different precipitation products to respond to geographical heterogeneity at the spatial scale, based on station precipitation data from 2014 to 2022.

3.4.1. Correlation Analysis Between Precipitation and Geographical Factors

Given the complexity of winter precipitation influenced by meteorological factors such as snowfall, this study focuses exclusively on flood season (May to October) precipitation data from ground stations. We investigated the relationships between precipitation and multiple geographical factors (including altitude, slope, aspect, terrain undulation, openness, latitude, and longitude), while examining their spatiotemporal heterogeneity. A multiple linear regression analysis was conducted to quantify the relationship between flood-season precipitation and geographical factors. We incorporated spatial relationships by calculating spatial weight matrices based on Euclidean distance between station coordinates (threshold = 0.5 degrees) to account for potential spatial autocorrelation effects. The regression model yielded an R² value of 0.194, indicating that geographical factors explain 19.4% of the variation in precipitation during the flood season. This suggests that while geographical factors play a certain role in precipitation distribution, other influencing factors also exist. The regression results (

Table 2. Results of partial correlation analysis between precipitation and geographical factors.

Geographical Factors	Coef	p > |t|
Const	2.748 × 1015	1.000
Altitude	0.0856	0.131
Slope	−0.1676	0.000
Aspect	−0.1486	0.002
Latitude	−0.0793	0.128
Longitude	−0.1576	0.018
Terrain Undulation	0.0061	0.899
Openness	−0.0389	0.414

) demonstrate the following:

Slope has a significant negative impact on precipitation (coef = −0.16, p < 0.001), indicating that steeper slopes receive less precipitation, which may be related to two factors: (1) Rain gauge exposure effects—steeper slopes may experience stronger wind exposure, leading to wind-induced undercatch in precipitation measurements. (2) Topographic sheltering—rain gauges on steep slopes may be influenced by local topographic sheltering effects that reduce the representativeness of point measurements. Aspect also shows a significant negative correlation (coef = −0.149, p < 0.001), suggesting that aspect may influence precipitation distribution, especially in areas where the interaction between wind direction and topography is more pronounced. The negative correlation indicates that certain slope orientations in our study area (north- facing and east-facing slopes) tend to receive less precipitation than south- facing and west-facing slopes, which is consistent with typical windward-leeward precipitation patterns in mountainous regions. Latitude (coef = −0.197, p = 0.008) and longitude (coef = −0.112, p = 0.089) are both negatively correlated with precipitation, indicating that within the study area, the southern and western regions receive relatively higher precipitation, while precipitation decreases towards the northern and eastern regions, possibly due to the impact of topography on moisture transport. Although altitude shows a certain positive correlation trend, it does not reach a significant level (p > 0.1), suggesting that within the scale of this study area, the direct influence of altitude on precipitation is relatively weak. Terrain undulation and openness have higher p-values (p > 0.4), indicating that they have less influence on flood season precipitation.

To further investigate the spatial dependency between precipitation and geographical factors, spatial regression analysis methods were employed, constructing a spatial adjacency matrix. The regression coefficients of various geographical factors are shown in Figure 15, with the mean square error (MSE) of the spatial regression model = 0.806, indicating that the model has a certain degree of accuracy in predicting flood season precipitation and can effectively characterize the spatial dependency of precipitation. Analysis results show that the influence of geographical factors on precipitation exhibits significant spatial heterogeneity. In mountainous areas with more complex topography, terrain undulation and altitude have more significant impacts on precipitation, while in plain areas, the influences of slope, aspect, and openness are more prominent.

To investigate the temporal variations in the relationship between precipitation and geographical factors, a correlation matrix was computed for the monthly flood-season precipitation and the geographical factors. The results (Figure 16) indicate that the correlations between precipitation and geographical factors vary significantly across different months.

Altitude’s influence on precipitation gradually increases from July to October, reflecting the impact of interactions between atmospheric circulation and topography during this period. Slope exhibits a consistently stable negative correlation throughout the flood season, with particularly stronger effects in June and September. The negative correlation of aspect is more pronounced in May, July, and September. Latitude and longitude show the most significant negative influence on precipitation in May, with their impact diminishing in August, which is related to the seasonal variations of the monsoon and moisture transport pathways. Terrain undulation demonstrates a slight enhancement in its influence on precipitation from June to August, although its overall effect remains minimal. Openness has a more significant impact in May and October, suggesting that during these months, less open (or more enclosed) terrain favors the occurrence of precipitation.

3.4.2. Evaluation of Precipitation Products in Characterizing Spatial Heterogeneity

To evaluate the ability of different precipitation products to represent the spatial heterogeneity of the study area, building on the analysis of the relationship between station precipitation and geographical factors in Section 3.4.1, a similar multiple linear regression and monthly correlation analysis was performed for the four precipitation products (ERA5-Land, MSWEP, CMA, and TPMFD) to assess their consistency with the observed station data. The ordinary least squares (OLS) regression coefficients relating the four precipitation products to the geographical factors are presented in

Table 3. OLS regression coefficients of precipitation products and geographical factors.

Geographical Factors	ERA5-Land	MSWEP	CMA	TPMFD	Obs
Altitude	−0.0132	−0.0256	−0.0008	0.0228	0.0856
Slope	−0.3782	0.5422	−0.1344	−1.3451	−0.1676
Aspect	0.0281	0.0297	0.1014	−0.1830	−0.1486
Latitude	−454.8824	84.1474	−128.5879	−248.4819	−0.0793
Longitude	26.4383	−119.2298	−40.0767	84.9056	−0.1576
Terrain Undulation	−0.0035	0.0101	0.0002	−0.0184	0.0061
Openness	0.8779	−3.5020	−12.8327	13.7749	−0.0389

, reflecting the differences in how the various products respond to the geographical factors.

Comparison of the regression coefficients between each product and the station-observed data reveals significant differences in the responses to geographical factors. For slope effects, the station observations indicate a significant negative correlation. TPMFD exhibits the strongest response to this feature, maintaining the same direction but with a noticeably amplified magnitude. Both ERA5-Land and CMA also display negative correlations consistent with the observed data, whereas MSWEP shows a positive correlation, which is contrary to the observed trend. Regarding aspect influences, only TPMFD successfully captured the negative correlation observed in stations, whereas ERA5-Land, CMA, and MSWEP showed spurious positive relationships, with CMA exhibiting the largest deviation from ground truth.

Regarding the effect of altitude, station observations indicate a weak positive correlation, and only TPMFD captured this trend, while the other three products all exhibited negative correlations. In terms of longitude, station data show a negative correlation. Both MSWEP and CMA reflected this trend; however, ERA5-Land and TPMFD showed positive correlations. Concerning openness, station data indicate a weak negative correlation. MSWEP and CMA were in the same direction but with significantly amplified magnitudes, while ERA5-Land and TPMFD exhibited strong positive correlations.

To quantitatively evaluate the accuracy of the spatial heterogeneity characterization ability of each product, the monthly regression coefficient of the product was calculated, and the matching degree between the product and the monthly regression coefficient of the station’s precipitation and geography factor (MSE and Spearman CC) is shown in

Table 4. Evaluation of correlation coefficient characterization ability of monthly declining aquatic products to precipitation and geographical factors at stations.

Products	OLS MSE	OLS Spearman	Monthly Mean MSE	Monthly Mean Spearman
ERA5-Land	29,650.59	−0.07	0.04	0.60
MSWEP	3040.70	−0.21	0.11	−0.45
CMA	2610.25	0.25	0.03	0.35
TPMFD	9875.97	0.21	0.02	0.87

.

Based on the MSE derived from OLS regression, CMA exhibits the lowest error, indicating the closest agreement with station data in terms of responses to geographic factors at a global scale, followed by MSWEP. In contrast, ERA5-Land shows significantly higher error, suggesting a greater deviation from actual geographic factor responses. Regarding the Spearman correlation coefficients, CMA and TPMFD perform relatively well, whereas ERA5-Land and MSWEP exhibit negative correlations, highlighting fundamental mismatches in spatial response directionality relative to ground truth patterns.

At the monthly scale, TPMFD demonstrates the best performance, with the lowest MSE (0.02) and the highest Spearman correlation coefficient (0.87), indicating its strong ability to capture the intermonthly variability in the relationship between precipitation and geographic factors. ERA5-Land and CMA also perform relatively well at the monthly scale, with Spearman coefficients of 0.60 and 0.35, respectively. In contrast, MSWEP performs the worst, exhibiting a negative Spearman correlation coefficient (−0.45), which is opposite to the observed trend.

4. Discussion

4.1. Performance Assessment of Precipitation Products in Complex Terrain

The comparative evaluation of four reanalysis precipitation products in the Lower Yarlung Zangbo Basin reveals significant variations in their ability to capture precipitation patterns in complex mountainous terrain, consistent with findings from previous studies in similar environments [72,77]. It should be noted that this comparison involves different temporal coverages, with CMA data spanning 2014–2020, while other datasets cover the full period of 2014–2022. Our results demonstrate that TPMFD achieves superior performance across multiple evaluation metrics, with correlation coefficients ranging from 0.6 to 0.8 and bias within ±20% at the annual scale. This enhanced performance aligns with previous research highlighting the advantages of multi-source data integration in mountainous regions [78,79], where traditional reanalysis products often struggle due to sparse observation networks and complex topographic influences.

The contrasting performance characteristics observed among the products reflect fundamental differences in their underlying methodologies and data assimilation approaches. TPMFD’s balanced bias pattern and robust correlation suggest that its sophisticated data fusion techniques effectively capture the diverse precipitation mechanisms characteristic of the Tibetan Plateau region, supporting previous findings that emphasized the importance of region-specific calibration in complex terrain [80]. Conversely, ERA5-Land’s substantial positive bias (40–120%) and spatial heterogeneity in performance echo concerns raised in earlier studies about the limitations of global reanalysis products in representing orographic precipitation processes [81,82].

The systematic underestimation observed in CMA (−54.51%) and MSWEP (−29.21%) products, despite their relatively high correlation coefficients, highlights a critical challenge in precipitation product development for mountainous regions. This pattern is consistent with previous research demonstrating that while reanalysis products may capture temporal variability effectively, they often exhibit systematic biases that require correction before hydrological applications [83,84]. The finding that CMA’s performance deteriorated after 2015 suggests potential changes in observation networks or data assimilation procedures, emphasizing the importance of maintaining consistent methodologies in long-term climate datasets [85]. The shorter temporal coverage of CMA data (2014–2020) compared to other products may also contribute to the observed performance differences, as the missing years (2021–2022) could contain precipitation patterns that affect long-term statistical comparisons.

4.2. Orographic Controls and Precipitation Mechanisms

Our analysis reveals significant relationships between topographic characteristics and precipitation distribution, providing insights into the physical mechanisms governing precipitation in the Yarlung Zangbo Basin. The observed negative correlations between slope, aspect, and precipitation align with established theories of orographic precipitation, where windward slopes receive enhanced precipitation while leeward areas experience rain shadow effects [86]. The west-to-east decreasing precipitation gradient identified through longitude correlation reflects the dominant moisture transport pathways from the Indian Ocean, consistent with previous studies on Asian monsoon dynamics [14,87].

The seasonal variations in topographic influence on precipitation patterns reveal the dynamic nature of atmospheric-terrain interactions in this region. The strengthening altitude-precipitation relationship from July to October corresponds to the evolution of monsoon circulation patterns and the vertical structure of moisture transport, supporting previous research on seasonal precipitation mechanisms over the Tibetan Plateau [88]. These findings have important implications for understanding how climate change may alter precipitation patterns in this region, as shifting atmospheric circulation patterns could modify the relative importance of different topographic controls.

The spatial heterogeneity in precipitation-topography relationships across the basin reflects the complex interplay between local terrain characteristics and regional atmospheric dynamics. In mountainous areas, terrain undulation and altitude emerge as dominant controls, while in relatively flat regions, slope orientation and topographic openness play more significant roles. This spatial variability explains the differential performance of reanalysis products across the study area and supports previous findings emphasizing the need for spatially-aware precipitation estimation approaches in complex terrain [89,90].

4.3. Implications for Hydrological Applications and Water Resource Management

The performance characteristics identified for each precipitation product have significant implications for their application in hydrological modeling and water resource management in the Yarlung Zangbo Basin. TPMFD’s superior accuracy makes it most suitable for applications requiring precise quantitative estimates, such as flood risk assessment and reservoir operation planning. However, users should remain cognizant of its limitations during extreme events, as evidenced by the error peaks during 2019–2020, which coincided with anomalous climate conditions across the region.

For applications focused on trend analysis and climate variability studies, CMA’s high correlation coefficients suggest it may be valuable despite its systematic underestimation, provided appropriate bias correction procedures are implemented. This finding is particularly relevant for long-term hydrological assessments, where temporal patterns are more critical than absolute magnitudes [91]. The consistent biases identified in our study provide a foundation for developing region-specific correction algorithms that could enhance the utility of these products for local applications.

The challenges observed with ERA5-Land in complex terrain highlight the continued limitations of global reanalysis products in mountainous regions, despite recent improvements in spatial resolution and data assimilation techniques. These findings suggest that regional downscaling approaches or hybrid methods combining multiple data sources may be necessary to achieve acceptable accuracy for hydrological applications in similar environments [92,93].

4.4. Future Research Directions and Recommendations

The spatial representativeness of rain gauge observations is a key limitation in our study. The concentration of stations in valleys may lead to (1) potential underestimation of precipitation in high-altitude areas, where orographic precipitation is more pronounced; (2) limited representation of windward and leeward slope effects; and (3) possible bias toward lower elevation precipitation patterns. This limitation should be considered when interpreting the evaluation results of reanalysis products, as the ‘ground truth’ itself may not fully capture the spatial heterogeneity of precipitation across the entire basin. Future studies would benefit from additional high-altitude stations or alternative validation approaches, such as satellite-based precipitation estimates, to complement ground observations.

Our findings highlight several important avenues for future research in precipitation product development and evaluation. First, the development of ensemble approaches that combine the strengths of different products could potentially overcome individual limitations while providing uncertainty estimates crucial for risk-based decision-making. The contrasting performance characteristics observed among products suggest that optimal weighting schemes could vary spatially and temporally, warranting investigation of adaptive ensemble methods.

Second, the pronounced errors during extreme events (2019–2020) underscore the need for enhanced representation of extreme precipitation processes in reanalysis products. Future research should focus on improving the parameterization of convective processes and extreme weather systems in complex terrain, potentially through the integration of high-resolution numerical weather prediction models and improved observation networks.

Third, the significant spatial heterogeneity in product performance suggests that spatially-varying bias correction approaches may be more effective than uniform correction methods. The development of machine learning techniques that can account for local topographic and climatological factors could significantly improve the accuracy of precipitation estimates in mountainous regions [94,95].

Finally, the changing performance characteristics observed in some products over time highlight the importance of continuous monitoring and validation of precipitation datasets. Establishing robust quality control procedures and developing methods to detect and correct for temporal inconsistencies in long-term records will be crucial for maintaining the reliability of climate datasets used in hydrological applications. Future research should consider incorporating daily-scale spatial evaluation approaches, particularly the comprehensive framework developed by Shaowei et al. [39], which demonstrated that spatial evaluation metrics can effectively quantify precipitation distribution accuracy. Their methodology provides valuable insights for developing multiple temporal-scale and spatial-scale evaluation frameworks that can bridge more comprehensive assessments in complex mountainous regions.

The findings from this study contribute to the growing body of knowledge on precipitation product evaluation in complex terrain and provide practical guidance for users in the Himalayan region. As climate change continues to alter precipitation patterns and extreme event frequencies, the development of more accurate and reliable precipitation products becomes increasingly critical for sustainable water resource management and disaster risk reduction in this vulnerable region.

5. Conclusions

This study systematically evaluated the performance of four reanalysis precipitation products (ERA5-Land, MSWEP, CMA, and TPMFD) in the lower Great Bend region of the Yarlung Zangbo River during 2014–2022. Based on newly acquired ground station observational data, this study analyses their accuracy across temporal scales and ability to characterize precipitation’s spatial heterogeneity. The main conclusions are as follows:

(1)

The lower Great Bend region of the Yarlung Zangbo River exhibits significant seasonal precipitation patterns, with the flood season accounting for 62.09% of annual precipitation. TPMFD consistently demonstrates superior performance, maintaining the lowest RMSE values (annual: 663.70 mm/yr; monthly: 113.62 mm/mon), balanced relative bias (annual: 8.80%; monthly: 2.62%), and stable correlation coefficients (annual: 0.77; monthly: 0.70). ERA5-Land exhibits systematic overestimation (annual: 101.40%; monthly: 80.01%), with larger northeastern errors. CMA shows good temporal correlation but significant systematic underestimation (annual: −54.51%; monthly: −57.64%). MSWEP performs moderately, with systematic underestimation, but shows negative annual temporal correlation with observations.

(2)

Monthly analysis reveals distinct seasonal variations in product performance. All products show larger estimation errors during the monsoon season (June–September), when precipitation is highest. Correlation coefficients peak during summer months for most products, with spatial errors typically larger in western regions, where complex terrain influences precipitation patterns.

(3)

Precipitation-geography analysis reveals that slope, aspect, and longitude significantly influence flood-season precipitation, with temporal variations throughout the season and different dominant factors across months. TPMFD best captures relationships between precipitation and geographical factors (monthly Spearman = 0.87), while CMA demonstrates better accuracy in overall spatial patterns (lowest OLS MSE) but poorer seasonal variation representation (Spearman = 0.35). ERA5-Land and MSWEP show limitations in accurately representing precipitation’s spatial heterogeneity.

Overall, TPMFD emerges as the most reliable product for precipitation estimation in the study area, offering balanced performance across temporal scales and accurate representation of spatial heterogeneity. However, each product demonstrates unique strengths and weaknesses relevant to specific applications in this region.