Changes in global climate and frequent harmful human activities have caused water shortage, which restricts the midstream social development and leads to downstream eco-environment degradation in inland river basins of Northwest China [1
]. Runoff is generated mainly from cold mountainous regions, which significantly affects the midstream and downstream areas [3
]. Hydrological models are widely used for the integrated assessment and management of basin water resources. Precipitation is an important input for accurate hydrological simulation, and its numerical accuracy and detailed spatial distribution are necessary [5
]. However, precipitation gauge stations are scarce and unevenly distributed in the cold mountainous regions of Northwest China because of economy, terrain, transport, and technology limitations [7
]. Thus, these gauge stations barely represent the spatial heterogeneity of regional precipitation, therefore leading to high uncertainty in hydrological simulation and analysis. Alternatively, this data-scarce situation can be addressed by using high-resolution gridded precipitation.
Gridded precipitation based on gauge stations has been widely investigated and used to establish hydrological models. Version 6 of the global precipitation product developed by the Global Precipitation Climatology Centre has monthly resolutions of 0.5° from 1901 to 2010 [8
]. Li et al. [9
] and Huang et al. [10
] used the spline interpolation and trend surface methods to determine gridded precipitation in China. Yang et al. [11
] evaluated different gridded precipitations to establish a hydrological model of the Three Gorges Reservoir. Fuka et al. [12
] used National Centrer for Environmental Prediction Climate Forecast System Reanalysis data to construct a hydrological model for validating the accuracy of gridded precipitation. Previous studies also used sparse meteorological stations to construct gridded precipitation in China; these stations poorly represent the amount and spatial distribution of precipitation [13
]. By contrast, hydrological stations can provide gauged precipitation to complete precipitation data; the regional climate model (RCM) can also supply information on spatial distribution to correct gridded precipitation [14
]. Gridded precipitation data with daily resolutions of 3 km have been developed for the Heihe River Basin (HRB) through spatial interpolation of meteorological station, hydrological station, and Regional Integrated Environmental Model System (RIEMS) RCM simulation. This high-resolution gridded precipitation can fully depict spatial heterogeneity, which is preferred for hydrological simulation and analysis [15
]. This gridded precipitation exhibits certain credibility, but few researchers use it to analyse the climate and hydrology characteristics in the upper HRB.
The hydrological model has been increasingly used to analyse the hydrological process in the HRB; water shortage problems are typical of the inland river basins of Northwest China. Soil and water assessment tool (SWAT) is a physical, semi-distributed hydrological model that has a few advantages in predicting climate change effects on water-related and hydrological processes over a continuous time [17
]. The performance of this model relies on precipitation input parameters, namely accuracy and spatial distribution [18
]. So, many researchers have selected grid precipitation to drive hydrological models. Evans et al. [19
] used four RCMs coupled with a CMD-IHACRES hydrological model to compare the different results. Lakhatkia et al. [20
] coupled MM5 with a THM hydrological model to study water resources and hydrological process response to climate change scenarios. Zou et al. [21
] used RIEMS simulation as driving data for a SWAT model to improve monthly runoff simulation in the upper HRB. Most studies that directly input grid data into the hydrological model cannot maximise the precision of high-resolution data because most grids are ignored [22
]. The SWAT model employs precipitation data from only one station closest to the centroid of each sub-basin, which can be corrected by elevation band and lapse rate; thus, the current method of representing precipitation in the SWAT model is simple. Accordingly, sub-basin precipitation input data are inaccurately represented [23
]. The SWAT model applied to the upper HRB focuses on model modification and hydrological process responses to climate and land use change; thus far, few studies have optimised precipitation input parameters using gridded data [24
]. Therefore, a reasonable scale transformation from the grid to the sub-basin must be developed to maximise the precision of high-resolution gridded data. By this method one can overcome the model structure defects and improve model input parameters. The suitability of the hydrological model in data-scarce regions can be improved by scale transformation of high-resolution gridded precipitation. Moreover, the water balance components should be accurately described to provide a reference scheme for similar hydrological models when using high-resolution gridded climate data in data scarce regions. The spatial variability of water balance components can be quantified to assess and manage water resources in the upper HRB [29
]. Water balance components cannot be directly measured but can be calculated by the SWAT model. The SWAT model can simulate water balance components with high accuracy and detailed spatial distribution depending on the inputs of high-resolution gridded precipitation. Currently, water balance components are estimated at different scales, namely, global, regional, watershed, and ecosystem levels [30
]. Thus, the spatial distribution, change trends, and internal relationship of water balance components across different scales can further strengthen the understanding of the hydrological processes.
This study regarded the upper HRB as a case study, and gridded precipitation with 3 km resolutions was used to construct a SWAT model. A scale transformation method was proposed to overcome the structure defects of the SWAT. The gridded precipitation was upscaled from the grid to the sub-basin scale to accurately represent sub-basin precipitation input data. The spatial distribution, change trends, and internal relationship of water balance components across different scales were analysed based on the model simulation. The main content includes: (1) assessing the quality of gridded precipitation data in the upper HRB; (2) conducting scale transformation by building virtual precipitation stations to transfer gridded data into a sub-basin average and calculating precipitation lapses rates on the sub-basin scale, thereby optimising the input parameters of precipitation; (3) assessing the performance of the SWAT model by comparing the monthly runoff simulation with observed data; and (4) analysing the spatial variability and change trend of water balance components on the sub-basin, landscape, and elevation band scales on the basis of the simulation results.
This study aimed to optimise the input parameters for hydrological simulation using high-resolution gridded precipitation and provide reference for water resource assessment and management in data-scarce regions. The hydrological simulation presents some uncertainties due to uncertainties in input data, model structure, model parameter, and validation data.
Precipitation is an important input for accurate hydrological simulation, and its numerical accuracy and detailed spatial distribution are necessary. The evaluation of gridded precipitation in time series accuracy is used only for two experimental stations and with a short time series (2011–2014). The evaluation of description capability is concentrated on the overall change trend. This data presents high spatial heterogeneity when compared with RIEMS RCM simulation and China National gridded product with a 3 km and 0.25° resolutions, respectively [15
]. The description capability of the gridded data is highly reliable. However, these evaluations are insufficient in demonstrating the precision of gridded precipitation because of lack of validation data. Scale transformation is proposed by building virtual precipitation stations and calculating precipitation lapse rate at sub-basin scale, thereby upscaling the gridded data from the grid to sub-basin scale. To some extent, these methods can be used to optimise the precipitation input parameters for the SWAT model effectively and maximise the horizontal and vertical distribution precisions of the high-resolution gridded precipitation. However, the 1113 grids were converted into 97 virtual stations at the sub-basin scale to simplify the spatial distribution of precipitation. Thus, the setting of sub-basin drainage threshold area is significant to scale transformation. The optimal sub-basin drainage threshold area of the sub-basin division based on basin climate and terrain, the division into the sub-basin with a larger number and the building of virtual station with high density are necessary. Previous studies showed that precipitation and elevation can be best described by log-linear or exponential functions [66
]. In the present study, linear regression functions were selected because their precipitation lapse rate was considered the mean annual value on the sub-basin scale in the SWAT model. Although this method simplifies the vertical variability of precipitation with elevation, a linear regression function is suitable for calculating the precipitation lapse rate for the SWAT model. For model climate forcing, only precipitation inputs use high-resolution gridded data; the temperature, wind speed, solar ration and relative humidity still use gauged data, which are scarce and unevenly distributed. The high-resolution gridded data of other climate elements should be applied in the SWAT model.
The upper HRB is a typical high cold mountainous region. The process of glacier and permafrost are not considered by the SWAT model. This situation will increase the uncertainty of hydrological simulation. Considering that the glacier area and runoff contribution are low and the glacier area slightly changed in recent years [67
], the uncertainty of ignoring the glacier melting runoff have controlled within a reasonable range. In the gentle-elevation catchment, climate uncertainty is relatively low and probably few gauged stations can give reasonable model performance [68
]. Precipitation in the upper HRB exhibits a large spatiotemporal variability because of convection in mountainous terrain [50
]. Given the complicated mountainous terrain, the high-resolution gridded precipitation is selected for hydrological simulation and analysis; the gridded precipitation was upscaled from the grid to the sub-basin scale. Especially in high mountain regions, this scale transformation method can improve the representation of sub-basin precipitation input data and reducing uncertainty.
After the analysis of parameter sensitivity by SWAT-CUP, the 10 most sensitive parameters were achieved. Furthermore, the parameters with high sensitivity were adjusted to achieve optimal simulation results. The range of parameter calibration was controlled within in ±20%. The daily precipitation event has great uncertainty and randomness. This study concentrated on monthly runoff simulation and annual scale analysis to reduce the uncertainty caused by daily precipitation. The SWAT model achieves excellent monthly runoff simulation on the large-scale and long-term series, which is sufficient to support the study on the water balance component characteristics on the mean annual scale. This result can provide a credible reference for basin water resource assessment and management. Moreover, most hydrological models are simulated monthly runoff in this study area. The current research can be compared with the previous study. However, the monthly simulation barely reflects the superiority of the gridded precipitation in time series and spatial distribution. Thus, the water balance component characteristics on the daily and small catchment scale should be further investigated.
The SWAT model is widely used in the upper HRB to study hydrological processes, in which NS are usually higher than 0.85 at the monthly scale [7
]. Compared with previous studies that used gauged precipitation, the monthly simulation accuracy derived in the present study has yet to be improved. However, the daily simulation is improved significantly. The dynamic change process of daily runoff response to precipitation events is more obvious. Considering the daily precipitation with a high uncertainty, the daily runoff simulation was assessed in typical normal and high flow year and exhibit high representativeness. Daily simulation is slightly better than monthly simulation based on the model results. Yin et al. [29
] and Zou et al. [21
] directly inputted gauged precipitation and grid precipitation, respectively, into the SWAT model in the upper HRB. This study used high-resolution gridded precipitation and conducted a scale transformation, which are significantly superior to a few gauged stations and directly input grid data. The precision of model simulation is significantly improved when compared with using grid precipitation as driving data (NS: 0.73). After input precipitation parameters with a high spatial heterogeneity and spatial representation, the spatial distribution of water balance components is more detailed and reasonable and its spatial continuity is better compared with the hydrological simulation based on gauged data and directly input grid data. The model calibration not only relies on hydrographs but also refers to basin features, such as base flow coefficient, evapotranspiration and snow melting runoff. Although the statistical evaluation criteria of simulation are not perfect, the hydrological process and distribution of water balance components are reasonable.
The 15-year simulation present a certain limitation in analysing the change trend of water balance components. In this region, the meteorology and hydrology studies are plentiful and mature in the historical period. On the basis of previous studies [27
], water balance components were analysed on the period of recent years; such study is uncommon. The underlying surface data used by the SWAT model are released in recent years; thus, these data are credible for meteorology and hydrology changing trend analysis in recent years. The precipitation lapse rate, water yield, soil water content and evapotranspiration for data validation lack gauged data that match with the resolution of simulation; thus, the superiority of this study is uncertain. These factors influence the accuracy of the model simulation.
In summary, the uncertainty of scale transformation, model parameter and validation data increase the uncertainty of hydrological simulation. Future studies should focus on these limitations in investigating the SWAT model driven by high-resolution gridded data and in reducing the uncertainty of hydrological simulation.
This study considered the upper HRB as a case study, and daily gridded precipitation data with 3 km resolution were selected as forcing data for the SWAT model. Gridded precipitation was subjected to quality assessment and exhibited high time series accuracy and spatial description capability. The scale transformation of gridded precipitation was proposed by building virtual precipitation and calculating precipitation lapse rate on the sub-basin scale. The precision of gridded precipitation in spatial distributions is maximised, and the input precipitation parameters of the SWAT model are optimised. The SWAT model exhibits a good monthly runoff simulation compared with the observed data from 2000 to 2014. The statistical analyses show that the R2 is higher than 0.71, NS is higher than 0.76, and PBIAS is controlled within ±15%. The base flow coefficient, snow melt runoff, and potential evapotranspiration simulated by the model are consistent with those of previous studies.
The spatial variability of water balance components was analysed on sub-basin, elevation band and landscape scales. The landscape of meadow and sparse vegetation and the band of 3500–4500 m are major water yield region. At the sub-basin scale, the spatial distributions of the water yield and evapotranspiration are consistent with that of precipitation and decrease from the southeastern to the northwestern areas; the spatial distribution of soil water content is similar to that of the western and eastern areas because of the landscape and elevation band effect; the precipitation and evapotranspiration in the entire basin present a slightly increasing trend, whereas the water yield and soil water content present a slightly decreasing trend. The spatial distribution, change trend, and internal relationship of water balance components across different scales can further strengthen the understanding of the hydrological processes and provide references for the assessment and management of water resources in data-scarce regions.