A Long-Term Simulation of Land Surface Conditions at High Resolution over Continental China

Peng Ji aKey Laboratory of Hydrometeorological Disaster Mechanism and Warning of Ministry of Water Resources/Collaborative Innovation Center on Forecast and Evaluation of Meteorological Disasters, Nanjing University of Information Science and Technology, Nanjing, China
bSchool of Hydrology and Water Resources, Nanjing University of Information Science and Technology, Nanjing, China

Search for other papers by Peng Ji in
Current site
Google Scholar
PubMed
Close
,
Xing Yuan aKey Laboratory of Hydrometeorological Disaster Mechanism and Warning of Ministry of Water Resources/Collaborative Innovation Center on Forecast and Evaluation of Meteorological Disasters, Nanjing University of Information Science and Technology, Nanjing, China
bSchool of Hydrology and Water Resources, Nanjing University of Information Science and Technology, Nanjing, China

Search for other papers by Xing Yuan in
Current site
Google Scholar
PubMed
Close
https://orcid.org/0000-0001-6983-7368
,
Chunxiang Shi cNational Meteorological Information Center, China Meteorological Administration, Beijing, China

Search for other papers by Chunxiang Shi in
Current site
Google Scholar
PubMed
Close
,
Lipeng Jiang dCMA Earth System Modeling and Prediction Centre (CEMC), China Meteorological Administration, Beijing, China

Search for other papers by Lipeng Jiang in
Current site
Google Scholar
PubMed
Close
,
Guoqing Wang eState Key Laboratory of Hydrology-Water Resources and Hydraulic Engineering, Nanjing Hydraulic Research Institute, Nanjing, China

Search for other papers by Guoqing Wang in
Current site
Google Scholar
PubMed
Close
, and
Kun Yang fMinistry of Education Key Laboratory for Earth System Modeling, Department of Earth System Science, Tsinghua University, Beijing, China

Search for other papers by Kun Yang in
Current site
Google Scholar
PubMed
Close
Free access

Abstract

With the improvement of meteorological forcings and surface parameters, high-resolution land surface modeling is expected to provide locally relevant information. Yet, its added value over the state-of-the-art global reanalysis products requires long-term evaluations over large areas, given uneven climate warming and significant land cover change. Here, the Conjunctive Surface–Subsurface Process version 2 (CSSPv2) model, with a reasonable representation of runoff generation, subgrid soil moisture variability and urban dynamics, is calibrated and used to perform a 6-km resolution simulation over China during 1979–2017. Evaluations against observations at thousands of stations and several satellite-based products show that the CSSPv2 has 67%, 29%, and 15% lower simulation errors for snow depth, evapotranspiration (ET), and surface and root-zone soil moisture, respectively, than nine global products. The median Kling–Gupta efficiency of the streamflow for 83 river basins is 0.66 after bulk calibrations, which is 0.38 higher than that of global datasets. The CSSPv2 also accurately simulates urban heat islands (UHIs) and the patterns and magnitudes of long-term snow depth, ET, and soil moisture trends. However, the global products do not detect UHIs and overestimate the trends (or show opposite trends) of snow depth and ET. Sensitivity experiments with coarse-resolution forcings and surface parameters reveal that advanced model physics and high-resolution surface parameters are vital for improved simulations of snow depth, ET, soil moisture, and UHIs, whereas high-resolution meteorological forcings are critical for modeling long-term trends. Our research emphasizes the substantial added value of long-term high-resolution land surface modeling to present global products at continental scales.

Significance Statement

Highly heterogeneous changes of terrestrial water and energy require kilometer-scale land surface information for the adaptation. High-resolution land surface modeling has been regarded as a promising approach to provide locally relevant information, but most applications are limited to a small region or a short period. By performing sets of 6-km resolution simulations over China during 1979–2017 with the Conjunctive Surface–Subsurface Process version 2 land model, here we show that high-resolution modeling has 15%–67% lower simulation errors of snow depth, streamflow, evapotranspiration, and soil moisture than nine global products, and the improvement is mainly attributed to the advances in model physical parameterizations and high-resolution surface parameters. Our results emphasize the great added value of kilometer-scale land surface modeling at continental scales.

© 2023 American Meteorological Society. For information regarding reuse of this content and general copyright information, consult the AMS Copyright Policy (www.ametsoc.org/PUBSReuseLicenses).

Corresponding author: Xing Yuan, xyuan@nuist.edu.cn

Abstract

With the improvement of meteorological forcings and surface parameters, high-resolution land surface modeling is expected to provide locally relevant information. Yet, its added value over the state-of-the-art global reanalysis products requires long-term evaluations over large areas, given uneven climate warming and significant land cover change. Here, the Conjunctive Surface–Subsurface Process version 2 (CSSPv2) model, with a reasonable representation of runoff generation, subgrid soil moisture variability and urban dynamics, is calibrated and used to perform a 6-km resolution simulation over China during 1979–2017. Evaluations against observations at thousands of stations and several satellite-based products show that the CSSPv2 has 67%, 29%, and 15% lower simulation errors for snow depth, evapotranspiration (ET), and surface and root-zone soil moisture, respectively, than nine global products. The median Kling–Gupta efficiency of the streamflow for 83 river basins is 0.66 after bulk calibrations, which is 0.38 higher than that of global datasets. The CSSPv2 also accurately simulates urban heat islands (UHIs) and the patterns and magnitudes of long-term snow depth, ET, and soil moisture trends. However, the global products do not detect UHIs and overestimate the trends (or show opposite trends) of snow depth and ET. Sensitivity experiments with coarse-resolution forcings and surface parameters reveal that advanced model physics and high-resolution surface parameters are vital for improved simulations of snow depth, ET, soil moisture, and UHIs, whereas high-resolution meteorological forcings are critical for modeling long-term trends. Our research emphasizes the substantial added value of long-term high-resolution land surface modeling to present global products at continental scales.

Significance Statement

Highly heterogeneous changes of terrestrial water and energy require kilometer-scale land surface information for the adaptation. High-resolution land surface modeling has been regarded as a promising approach to provide locally relevant information, but most applications are limited to a small region or a short period. By performing sets of 6-km resolution simulations over China during 1979–2017 with the Conjunctive Surface–Subsurface Process version 2 land model, here we show that high-resolution modeling has 15%–67% lower simulation errors of snow depth, streamflow, evapotranspiration, and soil moisture than nine global products, and the improvement is mainly attributed to the advances in model physical parameterizations and high-resolution surface parameters. Our results emphasize the great added value of kilometer-scale land surface modeling at continental scales.

© 2023 American Meteorological Society. For information regarding reuse of this content and general copyright information, consult the AMS Copyright Policy (www.ametsoc.org/PUBSReuseLicenses).

Corresponding author: Xing Yuan, xyuan@nuist.edu.cn

1. Introduction

Land represents a bridge between the biosphere, atmosphere, cryosphere, and hydrosphere and plays a vital role in the Earth system (Anderson and Radić 2020; Bastos et al. 2020; Seneviratne et al. 2010; You et al. 2020). Against the background of climate warming and land cover change (e.g., urbanization), hydrological and thermal variables over land surfaces are exhibiting significant changes, especially at local scales (Gudmundsson et al. 2021; Santamouris 2020; Yao et al. 2019), considerably impacting weather, crop yields, water resources, and ecosystem sustainability (Bellucci et al. 2015; Liang and Yuan 2021; Koster et al. 2004; Sehgal et al. 2018; Vicente-Serrano et al. 2013). Thus, accurate estimations of high-resolution land surface water and energy states are increasingly in high demand to understand the drivers of environmental changes and manage risks associated with food, water, and energy security.

Land surface models (LSMs), which describe the water, energy, and carbon cycles between the land surface and atmosphere, are efficient tools for providing land surface information (Fisher and Koven 2020). Decades of efforts have improved the hydrological and thermal parameterizations of LSMs, including the incorporation of runoff schemes suitable for mountainous regions (Clark et al. 2015; Li et al. 2011; Yuan et al. 2018; Zheng et al. 2017), consideration of subsurface lateral flows critical to soil moisture variation in regions with shallow groundwater (Ji et al. 2017; Maxwell and Condon 2016; Niu et al. 2014; Xie et al. 2020; Zeng et al. 2016), and the development of urban canopy modules describing radiative, hydrological, and thermal processes over different urban surfaces (e.g., walls, roofs, and roads) (Chen et al. 2011; Ji et al. 2021; Li et al. 2016; McNorton et al. 2021; Oleson et al. 2008). Moreover, the spatial resolution of meteorological forcings and surface parameters (e.g., soil texture, vegetation, and urban areas) has increased from 0.25°–1° to 0.000 27° (about 30 m)–0.125° over the continental scales in recent years (Gong et al. 2019; He et al. 2020; Shangguan et al. 2014; Shi et al. 2011). These achievements in model physics, high-resolution forcings, and surface parameters provide unprecedented opportunities for high-resolution (and even hyperresolution) land surface modeling, enabling the application of LSMs at kilometer resolution at national, continental, and even global scales (Bierkens et al. 2015; Wood et al. 2011).

However, LSMs are typically applied at coarse resolutions for global and continental simulations. For example, land data assimilation systems (LDASs) and atmospheric reanalysis usually use LSMs at 0.25°–1° resolution for simulations (Xia et al. 2019). Several well-known LSMs [e.g., the Community Land Model (CLM)] have been widely used at 0.5°–1° at continental to global scales (Lawrence et al. 2019; Toure et al. 2016; Xie et al. 2020) and at 0.1°–0.25° at local scales (Deng et al. 2020; Ma et al. 2021). Sensitivity experiments have shown that LSMs cannot take full advantage of the high-resolution forcings and parameters at coarse resolutions (Ji et al. 2017; Singh et al. 2015), resulting in uncertainty in simulating terrestrial water and energy cycles. Recent works have applied LSMs at high-resolution to represent land surface heterogeneity and key physical processes (Cheng et al. 2021; Ji et al. 2017; Naz et al. 2020; Singh et al. 2015; Tran et al. 2022; Hu. Zheng et al. 2019a). For example, Singh et al. (2015) performed a 1-km simulation over the southwestern United States during 2000–05 and obtained a significantly lower error than at 100- and 25-km resolutions. Naz et al. (2020) used the CLM3.5 LSM and an ensemble Kalman filter to produce a 3-km resolution daily soil moisture dataset over Europe during 2000–15. Tran et al. (2022) produced a 1-km resolution hydrological simulation product over the Upper Colorado River basin from 1983 to 2019 using the ParFlow-CLM LSM. However, these studies are either limited to a small region or a short period. High-resolution applications of LSM over large areas during a long period (e.g., >30 years) are required to obtain a comprehensive understanding of the added value of high-resolution land surface modeling.

In addition, the relative importance of high-resolution factors, including meteorological forcings, surface parameters, and LSMs, is being debated because the results vary by variables and study regions (Alavi et al. 2016; Ji et al. 2017; Rouf et al. 2021; Zeng et al. 2021). Previous research focused on one or two factors for one specific variable (e.g., soil moisture or soil temperature) in small catchments (Deng et al. 2020). Thus, further studies are needed to obtain more insights into the relative contributions of different high-resolution factors to improve the simulation accuracy of LSMs with multiple water and energy variables, especially over large regions (e.g., national and continental scales).

Using the Conjunctive Surface–Subsurface Process version 2 (CSSPv2) model that has advanced hydrological parameterizations (Choi et al. 2007; Choi and Liang 2010; Ji et al. 2021; Liang et al. 2012; Yuan and Liang 2011; Yuan et al. 2018) and wide applications in terrestrial hydrological and thermal changes (Ji and Yuan 2018, Ji et al. 2020a,b,c; J. Liu et al. 2022; Zhang et al. 2022), the objectives of this study are to (i) perform high-resolution (6 km) land surface simulations over continental China during 1979–2017 using numerous high-resolution (30 m–0.1°) meteorological forcings and surface parameters; (ii) comprehensively evaluate the simulated land surface variables critical for the energy and water budget and quantify the improvement in the high-resolution simulation accuracy over state-of-the-art global datasets; (iii) evaluate the performance improvement of the CSSPv2 by distinguishing the contributions from meteorological forcings, land surface parameters [e.g., soils, leaf area index (LAI)], and model physics. Observations from 83 streamflow stations on major rivers in China, 626 snow depth stations, 15 flux stations, more than 1000 soil moisture stations, and the recently released 8-km FLUXCOM evapotranspiration (ET) dataset and the 6-km MODIS land surface temperature (LST) dataset, are used to evaluate the CSSPv2 and nine global products comprehensively. The results provide a thorough insight into the model performance in representing the climatology, the variation and long-term trends of terrestrial water and energy variables. The results will shed light on high-resolution land surface modeling and provide an assessment of global products by comparing them with the CSSPv2 high-resolution simulation.

2. Model, data, and methods

a. The Conjunctive Surface–Subsurface Process version 2 model

The CSSPv2 is based on the Common Land Model (CoLM) (Dai et al. 2003) and inherits its advantages in simulating terrestrial water and energy processes (e.g., a five-layer snow module and a two-big-leaf model). The five-layer snow module changes the layer thickness dynamically according to the snow depth to better solve the snow melting, percolation, refreezing, and compaction. The snow albedo is parameterized according to the snow age, grain size, solar zenith angle, and fresh snow amount (Dai et al. 2003). The two-big-leaf model that considers sunlit and shaded fractions of the canopy is used to calculate canopy radiation transfer, photosynthesis, stomatal conductance, leaf temperature, and surface energy fluxes (Dai et al. 2004). The momentum, sensible heat flux, and latent heat flux are derived based on the Monin–Obukhov similarity theory.

The CSSPv2 has a larger soil layer depth (5.67 m) than the CoLM (3.43 m) to simulate the hydrological processes in the critical zone (e.g., 0–5 m) (Kollet and Maxwell 2008). The CSSPv2 solves soil moisture using the three-dimensional volume-averaged soil moisture transport (VAST) model, which decomposes the three-dimensional Richards’s equation into volume-averaged and perturbed terms and parameterizes the perturbed soil moisture terms based on subgrid topographic variability (Choi et al. 2007). The lateral soil water fluxes derived from the three-dimensional VAST model have a vital influence on soil moisture heterogeneity over humid mountainous regions at 1-km or finer spatial resolutions (Ji et al. 2017), which is a key process for hyper-resolution land surface modeling. The variable infiltration capacity (VIC) model (Liang et al. 1994) was incorporated into the CSSPv2 to model runoff (Yuan et al. 2018). This storage-based runoff scheme outperforms the groundwater-based runoff schemes over seasonally frozen and mountainous regions (Li et al. 2011; Yuan et al. 2018; Zheng et al. 2017).

Due to the crucial impacts of urbanization, a single-layer urban canopy model with a similar structure to that in the CLM model was included in the CSSPv2 (Ji et al. 2021). The model grid is divided into nonurban and urban tiles, and each tile has its own energy and water balance. Urbanization is also considered by changing the urban fraction each year, and the energy and water balance are maintained at the transition between urban and nonurban tiles following Li et al. (2016).

The CSSPv2 has been used in numerous studies, including the assessment of long-term terrestrial water budget changes in a headwaters region (Yuan et al. 2018), understanding the drivers of ground temperature warming (Ji et al. 2020a) and flash drought intensification (Zhang et al. 2022), detection, attribution, and projection of extreme streamflow (Ji and Yuan 2018; Ji et al. 2020b,c), coupling with weather forecast and deep learning models for streamflow prediction in a cascade reservoir watershed (J. Liu et al. 2022), and the assessment of urban heat islands (UHIs) in the Beijing metropolitan area (Ji et al. 2021). Zeng et al. (2021) found that CSSPv2 simulated soil moisture exhibited higher correlation coefficients and lower root-mean-square errors over China than other soil moisture products, and further sensitivity experiments revealed that the improvement was attributed to the model structure and parameterizations. High-resolution modeling over China because of its advantages in modeling soil moisture and other water and energy variables (Yuan et al. 2018).

b. Study area and validation data

This research evaluates the CSSPv2 over continental China, which has diverse landscapes with mountains and plateaus in the west and plains in the east (Fig. 1a). The country was divided into nine subregions (e.g., East, South, Southwest, Northeast, North, Xinjiang, Northwest, Inner Mongolia, and Tibet) according to climatic characteristics and conventions in geographical division (Wu et al. 2021) to conduct regional evaluations. The total urban area in China was 2.1 × 105 km2 in 2017 (Fig. 1b). The three urban agglomerations, i.e., Beijing–Tianjin–Hebei, the Yangtze River delta, and the Pearl River delta are shown in Fig. 1b.

Fig. 1.
Fig. 1.

The study region and station locations. (a) Topography of China. (b) Urban areas in 2017 (red grids). The blue lines are the boundaries of three metropolitan areas. (c) Locations of stations with monthly streamflow observations during 1980–89. (d) Locations of stations with monthly snow depth observations during 1980–2009. (e) Locations of stations with annual evapotranspiration (ET) measurements. (f) Locations of automatic soil moisture stations with daily records during 2013–17. The pink circles represent stations with valid records only at the surface (10-cm depth), and the blue circles are stations with surface and root zone (0–1-m depth) observations.

Citation: Journal of Hydrometeorology 24, 2; 10.1175/JHM-D-22-0135.1

Observations from thousands of stations and several satellite-based high-resolution products were used to evaluate the streamflow, snow depth, ET, surface and root-zone soil moisture, and LST simulations. The spatial and temporal resolutions of the validation datasets are listed in Table 1, and the station locations are given in Figs. 1c–f. All validation datasets underwent strict quality control during product generation (see Text S1 in the online supplemental material for detailed information). Note that the monthly streamflow at the stations on the Yellow River and Haihe River is the naturalized flow, which was estimated by adding human water interventions (e.g., reservoir operations) back to the observation. While the observed monthly streamflow of the Yangtze, Pearl, and Songliao Rivers was used due to the lack of naturalized streamflow data. In addition, nine global products were evaluated to compare them with the high-resolution CSSPv2 simulation with multisource observations. The details of the nine products are listed in Table 2.

Table 1

Validation data.

Table 1
Table 2

The global high-resolution and long-term products used to compare with the CSSPv2 high-resolution simulation. Note GLDASv2.0_Noah and GLDASv2.0_CLSM were used only when the GLDASv2.1_Noah and GLDASv2.2_CLSM did not have records.

Table 2

c. Calibration of runoff parameters in CSSPv2

Calibrating the runoff parameters is critical for streamflow simulations and is widely used in hydrological studies to achieve satisfactory results (Guo et al. 2019; Tsai et al. 2021). Here, three parameters (including the b factor, Ds,max, and Ds) for the VIC runoff scheme were calibrated. The b factor determines the shape of the surface infiltration curve and is used to calculate the surface saturated area and infiltration capacity. Therefore, b factor influences the Dunnian runoff (when precipitation occurs over the saturated area) and the Hortonian runoff (when the rainfall intensity exceeds the maximum infiltration rate). The Ds,max and Ds terms determine the shape of baseflow curve, where Ds,max is the maximum baseflow rate and Ds is the ratio of the maximum linear baseflow rate to the maximum baseflow rate. Following previous studies (Yuan et al. 2016; Zhang and Yuan 2020), the monthly runoff (R) of a catchment is estimated by
R=ΔQ/S,
where ΔQ is the streamflow difference between the gauge located at the basin outlet and the adjacent upstream gauge, and S is the catchment area between two gauges. The calibration was performed as follows: 1) a catchment was defined as one hydrologic response unit (HRU) (Chaney et al. 2016), whose meteorological forcings, soil texture, vegetation fraction, and LAI were derived from the regional averages and the dominant land use category in the catchment; 2) the target runoff for the HRU was calculated using Eq. (1); 3) 2000 simulations were performed for the HRU with different sets of runoff parameters that were generated using a uniform distribution; 4) the Kling–Gupta efficiency (KGE) between the simulated and estimated runoff was used as a cost function, and the shuffled complex evolution method developed at the University of Arizona (SCE-UA) (Duan et al. 1994) was used to obtain the optimal parameters.

The calibration was performed over 31 subbasins of the five large river basins in China using records from 31 streamflow stations located on the main streams (Fig. 1c). The remaining 52 stations were only used for the validation. The parameters proposed by Nijssen et al. (2001) were utilized for other basins without observations. We did not calibrate the parameters for each grid because it was too time-consuming, especially at high resolution.

d. Experimental design and evaluation metrics

1) The 6-km long-term simulation using CSSPv2

The CSSPv2 was applied at 6-km resolution over China with 830 × 700 grids generated from the Lambert conformal conic map projection centered at (36.7°N, 102.0°E). Simulations were conducted for two cycles from 1979 to 2017 at a half-hourly time step (hereafter referred to as OBS/CSSPv2). The initial conditions for the second cycle came from the land surface conditions at the end of the first cycle. The first cycle and the first year in the second cycle were taken as the spinup period, which is required for the water and energy cycles in the LSM to reach equilibrium (Alavi et al. 2016). The data during 1980–2017 in the second cycle was used for evaluations. The schematic diagram of the simulation procedure is shown in Fig. 2, and detailed information of high-resolution surface parameters and meteorological forcings are given below.

Fig. 2.
Fig. 2.

Diagram of the long-term and high-resolution land surface simulation using the CSSPv2 model. The meteorological forcing data includes precipitation (P), near-surface air temperature (T2m), near-surface humidity (Q2m), surface pressure (Surf Pres), surface wind speed (Surf Wind), surface downward longwave radiation (Surf LW), and surface downward shortwave radiation (Surf SW). The CLDAS, CN051, and CMFD products were interpolated to 6-km spatial resolution before producing the meteorological forcing data, therefore all the datasets in the left panel have the same spatial resolution.

Citation: Journal of Hydrometeorology 24, 2; 10.1175/JHM-D-22-0135.1

The meteorological forcing data, including near-surface air temperature, surface pressure, wind speed, humidity, and the shortwave and longwave radiation fluxes, were interpolated bilinearly from the 0.1° and 3-hourly China Meteorological Forcing Dataset (CMFD) (He et al. 2020). The bilinear interpolation method is a straightforward remapping algorithm widely used in generating meteorological forcings (Qian et al. 2006; Rouf et al. 2020; Wang et al. 2016). Bilinear interpolation and topographic bias correction of the temperature, humidity, and pressure data were performed as follows: 1) the dew temperature (Td) was derived according to the surface air temperature (T2m), surface air specific humidity (Q2m), and the surface pressure (Surf Pres); 2) the Td, T2m, and pressure were interpolated bilinearly to 6-km resolution; 3) the interpolated Td and T2m were elevation-corrected assuming a fixed lapse rate (6.5 K km−1); 4) the interpolated Surf Pres was elevation-corrected using the barometric height formula derived from hydrostatic approximation and the ideal gas law; 5) the bias-corrected Surf Pres, Td, and T2m were used to derive Q2m.

In contrast to the CMFD, which uses precipitation from about 700 meteorological stations, the CN05.1 precipitation dataset uses observations from more than 2000 meteorological stations (Wu and Gao 2013). However, the CN05.1 precipitation dataset has a relatively coarser (0.25°) spatial resolution and overestimates the precipitation magnitude as validated by independent observations (Yang et al. 2017). The China Meteorological Administration (CMA) Land Data Assimilation System (CLDAS) (Shi et al. 2011) precipitation data has a high spatial resolution (0.0625°) with observations from 30 000–60 000 regional meteorological stations (Sun et al. 2020). Although the CLDAS precipitation dataset performs better in estimating the precipitation magnitude than the CMFD and CN05.1 datasets (Yang et al. 2017), the short data period (2008–present) prevents its direct application in this research. Therefore, the CMFD, CLDAS, and CN05.1 were all employed to provide consistent, continuous and accurate precipitation forcing. 1) The CMFD, CN05.1, and CLDAS daily precipitation data were interpolated to the 6-km resolution using the conservation interpolation method. 2) The CN05.1 precipitation dataset during 2008–17 was corrected using quantile-mapping (Wood et al. 2002), which adjusts the cumulative distribution function (CDF) from the CN05.1 to the CLDAS precipitation dataset. 3) The CN05.1 precipitation dataset during 1979–2007 was corrected using equidistant CDF to account for distribution changes from 1979–2007 to 2008–17 (Li et al. 2010). 4) All bias corrections were performed at the monthly time scale, and the daily precipitation was adjusted according to the bias-corrected monthly precipitation. 5) The 3-hourly variation from the CMFD was used to downscale the corrected daily CN05.1 precipitation to 3-hourly data. This bias correction method preserves the long-term precipitation trend (not shown), which is consistent with previous research (Yuan et al. 2019).

Finally, the 3-hourly T2m, Q2m, Surf Pres, Surf Wind, and Surf LW were linearly interpolated to hourly time step. The 3-hourly Surf SW was decomposed into hourly data according to the solar zenith angle. The 3-hourly total precipitation was separated uniformly into hourly data. A similar temporal interpolation method was used by the CLM (Oleson et al. 2013).

Numerous state-of-the-art land surface datasets were used to provide high-resolution surface parameters, including the global soil texture dataset at 1 km resolution (Shangguan et al. 2014), the hydrologically conditioned Shuttle Radar Topography Mission (SRTM) elevation dataset at 3-arc-s resolution (Lehner et al. 2008), the GLASS monthly LAI product at 0.05° resolution during 1981–99 retrieved from AVHRR (Liang et al. 2021), the reprocessed MODIS version 6 monthly LAI dataset at 0.05° resolution during 2000–17 (Yuan et al. 2011), and the annual dataset of urban areas in China at 30 m resolution during 1978–2017 (Gong et al. 2019). Note that vegetation greening/withering was considered by updating the observed monthly dynamic LAI. The monthly LAI during 1979/80 was the same as that in 1981 due to the lack of data. Urbanization was represented in the simulation by changing the urban fraction of each grid every year.

2) Sensitivity experiments and contribution separation method

Three sensitivity experiments (OBS/CSSPv2_NoCalib, ERA5Forc/CSSPv2, and ERA5SurfForc/CSSPv2) were performed to provide insights into the improvement of OBS/CSSPv2 in terms of the model, forcings, and surface data. The OBS/CSSPv2_NoCalib was the same as OBS/CSSPv2 but used the default runoff parameters developed over the headwaters region (Yuan et al. 2018). The ERA5Forc/CSSPv2 experiment was the same as OBS/CSSPv2 except for the ERA5 meteorological forcings. The ERA5SurfForc/CSSPv2 experiment was the same as OBS/CSSPv2 but used surface data (soil property, vegetation cover, land-use category, and LAI) and meteorological forcings from ERA5. Thus, the difference between ERA5 and ERA5SurfForc/CSSPv2 was the influence of the LSM (Table 3).

Table 3

Experimental design. The observed meteorological forcings are derived from the CMFD, CLDASv2.0, and CN05.1 datasets, while observed surface parameters come from different high-resolution datasets.

Table 3

Following Zeng et al. (2021), the influences of LSM, surface parameters, and meteorological data on the evaluation metric (e.g., KGE that will be introduced in the next section) are calculated as
ΔKGEmodel=KGEERA5SurfForc/CSSPv2KGEERA5,
ΔKGEsurface=KGEERA5Forc/CSSPv2KGEERA5SurfForc/CSSPv2,
ΔKGEforcing=KGEOBS/CSSPv2KGEERA5Forc/CSSPv2.
The contributions (CTs) of the model, surface parameters, and forcing for a grid (or a station) where OBS/CSSPv2 shows a higher KGE than ERA5 are calculated as
CTmodel=max(0,ΔKGEmodel)max(0,ΔKGEmodel)+max(0,ΔKGEsurface)+max(0,ΔKGEforcing),
CTsurface=max(0,ΔKGEsurface)max(0,ΔKGEmodel)+max(0,ΔKGEsurface)+max(0,ΔKGEforcing),
CTforcing=max(0,ΔKGEforcing)max(0,ΔKGEmodel)+max(0,ΔKGEsurface)+max(0,ΔKGEforcing).
When all factors lead to a positive ΔKGE, Eqs. (5)(7) provide the relative contributions of each factor. When one factor has a negative ΔKGE, for example, the surface parameters, Eq. (6) provides a zero value of CTsurface, indicating that using high-resolution surface data does not provide a positive contribution. For the model and forcing factors that have positive influences, Eqs. (5) and (7) guarantee their relative importance. When only one factor shows a positive KGE, 100% of the improvement is attributed to the factor.

It is worth noting that this contribution method can be applied to any dataset and the OBS/CSSPv2. We chose the ERA5 for the analysis because it had the best performance among the nine global datasets (section 3).

3) Evaluation metrics

The metrics used for the quantitative evaluation of the products include the correlation coefficient (CC), root-mean-square error (RMSE), unbiased RMSE (unRMSE), and KGE (Gupta et al. 2009). The unRMSE does not have systematic errors and is typically used for soil moisture evaluation. The KGE is widely used to evaluate streamflow and is calculated as
KGE=1(CC1)2+(μsμo1)2+(σsσo1)2,
where μs and μo are the mean simulated and observed values, respectively, during the evaluation period. The σs and σo terms are the simulated and observed standard deviations, respectively. The KGE ranges from negative infinity to 1. Streamflow simulations are regarded as satisfactory when the KGE is larger than 0.5 (Moriasi et al. 2007). In addition, the KGE was chosen to compare different products because it is an integrated measure of the correlation, bias, and variance.

3. Evaluation of CSSPv2 simulation and its superiority against nine global products

a. Streamflow

The CSSPv2 simulated streamflow shows a good performance after bulk calibrations (Fig. 3a). Figures 3b–e show the observed and simulated monthly streamflow at four stations, suggesting that the OBS/CSSPv2 captures the monthly streamflow variations well. The KGE is larger than 0.5 at most stations (Fig. 3a), indicating the simulation accuracy is satisfactory. Figure S1 in the online supplemental material exhibits the observed and simulated annual cycle of monthly streamflow at 30 stations used in the calibration. In general, the calibration improves the simulation of the annual cycle.

Fig. 3.
Fig. 3.

Evaluations of monthly streamflow. (a) Kling–Gupta efficiencies (KGEs) between the CSSPv2 simulation and the observations in 83 catchments during 1980–88. The circles represent the stations used in the calibration, and the triangles are stations used for the evaluation. (b)–(e) Observed and simulated monthly streamflow during 1980–88 at four stations.

Citation: Journal of Hydrometeorology 24, 2; 10.1175/JHM-D-22-0135.1

Figure 4 shows the evaluation of annual mean streamflow in 83 catchments. The OBS/CSSPv2_NoCalib (pink dots in Fig. 4a) underestimates the streamflow by 37% (the regression slope is 0.63 between the observation and OBS/CSSPv2_NoCalib), and the median value of the KGE in 83 catchments is 0.43. The global atmospheric and land reanalysis datasets have much lower performances than the OBS/CSSPv2_NoCalib because the streamflow is underestimated or overestimated by up to 69% and 52%, respectively (the same is found for the seasonal cycle in Fig. S1), with median KGEs being 0.1–0.39. The water budget-based TerraClimate product (orange dots and box), however, shows slightly higher KGEs than the OBS/CSSPv2_NoCalib. After calibration, the OBS/CSSPv2 only has a 15% underestimation of streamflow and has KGEs larger than 0.5 in about 75% of the catchments. The median value increases to 0.66, indicating the considerable added value of the bulk calibration procedure. In general, the OBS/CSSPv2 has a 48%–78% lower magnitude of the absolute bias and a 0.17–0.54 higher median KGE (mean changes of the median KGE is 0.38) than the global products.

Fig. 4.
Fig. 4.

(a) Comparison of simulated and observed multiyear mean monthly streamflow during 1980–89 in 83 catchments and (b) boxplot of the Kling–Gupta efficiency (KGE) between simulated and observed monthly streamflow during 1980–89. The boxes represent the 25th and 75th percentiles, the line is the median value, and the whiskers (dashed vertical lines) represent 99% of the data. The OBS/CSSPv2_NoCalib denotes the CSSPv2 simulation using the default runoff parameters in Yuan et al. (2018).

Citation: Journal of Hydrometeorology 24, 2; 10.1175/JHM-D-22-0135.1

b. Snow

Figure 5a shows that the OBS/CSSPv2 has an R2 value of 0.71 for simulating winter snow depth, although underestimations occur when the snow depth exceeds 100 mm. Figures 5b–d show that MERRA2 and GLDASv2.2_CLSM perform worst, with low CCs and high RMSEs. Although ERA5_Land has a higher CC than ERA5, it still has a large positive bias and high RMSE. GLDASv2.1_Noah and FLDASv1.0_Noah have smaller RMSEs (12–17 mm) than ERA5 and ERA5_Land, but their CCs are only 0.4–0.42. The OBS/CSSPv2 has the highest CC (0.69) and the lowest RMSE (9.6 mm), indicating it is superior for representing the snow magnitude and variability. Compared with the global products, mean increases of medium CC in OBS/CSSPv2 are 0.11–0.49, and the mean relative decreases of RMSE are 67%. The RMSE of snow depth in the OBS/CSSPv2 is also lower than the values obtained from remote sensing retrievals (e.g., AMSR2 and FY3B) and machine learning products, whose RMSEs range from 18 to 36 mm (Jiang et al. 2014; Yang et al. 2021; Zhang et al. 2017). In addition, the OBS/CSSPv2 proves superior for simulating snow depth in other seasons (Fig. S2).

Fig. 5.
Fig. 5.

Evaluation of the mean snow depth in winter (December–February). (a) Scatterplot between the in situ observed and CSSPv2 simulated snow depth. The climatology during the winter season of 1980–2009 is compared. The red line shows the linear regression result. (b) Bar plot of the mean correlation coefficient (CC) between the observed and simulated snow depth in winter from different products during 1980–2009. The error bars show one standard deviation. (c),(d) As in (b), but depict the bias and root-mean-square error (RMSE), respectively.

Citation: Journal of Hydrometeorology 24, 2; 10.1175/JHM-D-22-0135.1

Figure 6 shows the observed and simulated linear trends of winter snow depth during 1980–2009. Similar to previous research (Peng et al. 2010; Zhang et al. 2020), stations with significant increasing trends are located north of 40°N (e.g., Xinjiang, Inner Mongolia, and Northeast China), whereas a significant decreasing trend occurs in the eastern part of Tibet (Fig. 6a). The OBS/CSSPv2 captures the trend well, and the spatial correlation between the simulation and observation is 0.55 (p value = 0). However, the global products have difficulties in reproducing the changing pattern or overestimate the trends by more than 2–4 times. For example, the ERA5_Land and ERA5 show large decreasing trends in Northeast China and the northern part of Xinjiang. The GLDASv2.0_CLSM and MERRA2 do not depict the increasing trend in Northeast China and overestimate the decreasing trend by 2–4 times in Tibet. The GLDASv2.0_Noah shows a much larger increasing trend for nearly the entire country. The spatial correlations are low or negative between these global products and the observation (from −0.26 to 0.16), except for the GLDASv2.0/Noah (0.47).

Fig. 6.
Fig. 6.

Linear trends (mm decade−1) of (a) the observed winter snow depth during 1980–2009 and (b)–(h) the results of different products. The solid circles indicate the trend is significant at the 95% significance level.

Citation: Journal of Hydrometeorology 24, 2; 10.1175/JHM-D-22-0135.1

c. Evapotranspiration

1) Evaluation against station observations

Figure 7 shows that OBS/CSSPv2 has a similar performance as GLEAM v3.5 for the simulation of annual ET, with a slightly lower RMSE. The ET from the remaining seven global products shows a wide dispersion around the dashed line, and the RMSE is 25%–99% larger than that of the OBS/CSSPv2.

Fig. 7.
Fig. 7.

Comparison of annual evapotranspiration (ET) derived from different products with in situ measurements. The R2 values and root-mean-square errors (RMSEs) are shown in the lower-right corners, and the relative RMSEs compared to the multiyear averaged ET are shown in parentheses. The model grids containing the observation stations were selected for the comparison.

Citation: Journal of Hydrometeorology 24, 2; 10.1175/JHM-D-22-0135.1

Figure 8 shows the Taylor diagrams of the monthly ET at different stations. All products have larger variability than the observation at the DHS evergreen forest station, and the CC is 0.8–0.9. The GLEAMv3.5 is the closest to the observations, whereas the OBS/CSSPv2 performs slightly worse than the GLEAMv3.5 and GLDASv2.1_Noah, with a larger RMSE and standard deviation. The GLEAMv3.5 still provides the closest simulations to the observations at the QYZ evergreen needleleaf forest station, followed by the OBS/CSSPv2 and GLDASv2.1_Noah. However, the OBS/CSSPv2 exhibits superiority (a 64% lower RMSE) over the GLEAMv3.5 and other products at mixed forest, grassland, cropland, alpine meadow, and alpine desert stations (Figs. 8c–h). The OBS/CSSPv2 shows similar performance as the GLEAMv3.5 at urban stations, with 40%–80% lower RMSEs than the other products. The GLDASv2.1_Noah and FLDASv1.0_Noah exhibit low variability at the urban station because the high wilting point used in the Noah model in the urban grids prevents the evapotranspiration in urban areas (Ji et al. 2021).

Fig. 8.
Fig. 8.

Taylor diagrams of monthly evapotranspiration during 2003–10 obtained from station observations and different products. The observation period for the Beijing station is 2013–14. DHS, QYZ, CBS, HBGC, INM, YC, SETORS, QOMS, and Beijing are station names.

Citation: Journal of Hydrometeorology 24, 2; 10.1175/JHM-D-22-0135.1

2) Evaluation against FLUXCOM dataset

The FLUXCOM gridded ET dataset was used to evaluate the OBS/CSSPv2 and other datasets at the grid and regional scales. Figure 9a shows the spatial distribution of the mean annual ET during 2001–15 obtained from the FLUXCOM product, and the uncertainty is shown in Fig. 9b. Figure 9c shows that the OBS/CSSPv2 captures the decrease in ET from southern to northwestern China, with a spatial correlation of 0.87 (p value = 0). Figures 9d–l show the biases between FLUXCOM and different products. The OBS/CSSPv2 exhibits negative biases in the center of north China and the southwestern part of northwest China (from −13 to −12% relative to the regional mean ET) and positive biases in east China (3%–4%). The national mean ET bias is −13 mm yr−1, whereas the regional biases are from −61 to 25 mm yr−1 (from −13% to 4% relative to the regional mean). The GLEAMv3.5 ET shows a similar bias pattern as the OBS/CSSPv2, but the regional biases are larger (from −93 to 44 mm yr−1, from −16% to 6% relative to the regional mean). The ERA5_Land and ERA5 have positive biases for nearly the entire country, whereas the other products generally show large positive and negative biases to the south and north of 30°N, respectively (Figs. 9f–l).

Fig. 9.
Fig. 9.

Evaluation of the mean annual evapotranspiration (ET). (a) Mean annual ET during 2000–15 and (b) its uncertainty in FLUXCOM. (c) As in (a), but for the OBS/CSSPv2. (d)–(l) The biases between different products and FLUXCOM [e.g., ΔET(OBS/CSSPv2) = ETOBS/CSSPv2 − ETFLUXCOM], and the regional ΔET is shown in the subplot in the lower-left corner. Note, only the grids with biases larger or smaller than the FLUXCOM ET uncertainty ranges are shown.

Citation: Journal of Hydrometeorology 24, 2; 10.1175/JHM-D-22-0135.1

Figure 10 shows the KGEs for different datasets as validated by the FLUXCOM ET at the monthly scale during 2001–15. The KGE for the OBS/CSSPv2 exceeds 0.5 in most grids, and the spatial mean KGEs in the nine subregions are 0.59–0.86, indicating good performance in representing the magnitude and variability of monthly ET. The KGE of the OBS/CSSPv2 is 0.05–0.1 higher than that of the GLEAMv3.5 in north China, northwest China, and Tibet, and its performance is slightly better in the other regions. The regional mean KGE of the OBS/CSSPv2 is 0.13–0.39 higher than that of other products, and the largest difference is observed in Xinjiang; Tibet; and east, south, and southwest China. The time series of the monthly ET averaged over nine subregions (Fig. S3) shows that the global reanalysis datasets tend to overestimate ET during the growing season (April–September) in east, south, and southwest China, causing overestimations of the mean and variability of ET. In contrast, the global reanalysis datasets underestimate the ET significantly in the Xinjiang region (except for the ERA5 and ERA5_Land), which may be related to the large dry bias in root-zone soil moisture. In Tibet, the global datasets (e.g., ERA5, ERA5_Land, FLDASv1.0_Noah, and TerraClimate) show a delayed peak in monthly ET by 1–2 months, resulting in low CCs. In contrast, the OBS/CSSPv2 has a similar ET magnitude and variability as FLUXCOM and captures the annual cycle of ET.

Fig. 10.
Fig. 10.

Spatial distribution of the Kling–Gupta efficiency (KGE) between the monthly evapotranspiration derived from FLUXCOM and other products.

Citation: Journal of Hydrometeorology 24, 2; 10.1175/JHM-D-22-0135.1

Figure 11 gives the annual ET anomaly averaged over China during 2001–15. Similar to FLUXCOM, the OBS/CSSPv2 shows a slight but insignificant increasing trend of the national mean ET. However, the GLEAMv3.5, GLDASv2.1_Noah, FLDASv1_Noah, TerraClimate, and MERRA2 exhibit significant increasing ET trends, with 2–12 times higher rates of increase than FLUXCOM. In contrast, ERA5 and ERA5_Land show decreasing ET trends during 2001–15.

Fig. 11.
Fig. 11.

Annual evapotranspiration anomaly averaged over China during 2001–15 for different products. The linear trends (mm yr−1 decade−1) and the p value are provided.

Citation: Journal of Hydrometeorology 24, 2; 10.1175/JHM-D-22-0135.1

Figure 12 shows the linear trends of annual ET. The FLUXCOM shows three hotspots with significant ET changes, including the Hetao region, northeast China, and Tibet. These results are consistent with previous research using different datasets. For example, the water balance-based ET (Ma and Zhang 2022) and the Penman–Monteith-based ET (Song et al. 2017) showed a decreasing ET trend in Tibet during 2003–14 and 2000–10, whereas the MODIS-based ET and modeled ET showed increasing ET trends in Hetao and northeast China (Meng et al. 2020; Ha. Zheng et al. 2019b). The OBS/CSSPv2 and GLEAMv3.5 capture the significant changing patterns, but the GLEAMv3.5 shows a 2 times larger ET increasing trend over northern China. In addition, the GLEAMv3.5 shows a significantly increasing ET trend in southern China (e.g., south, east, and southwest China) while FLUXCOM and OBS/CSSPv2 do not. This can interpret the overestimated increasing trend of national mean ET in the GLEAMv3.5 dataset. The other products have difficulty in representing the ET changing patterns or magnitudes. For example, the ERA5_Land, GLDASv2.2_CLSM, and ERA5 show opposite trends in the Hetao region, the GLDASv2.1_Noah shows a significant increasing ET trend in nearly the entire country, whereas the MERRA2 shows a 2–3 times larger increasing trend than FLUXCOM in northeast China. Thus, the OBS/CSSPv2 exhibits significant advantages over current high-resolution global land reanalysis datasets in capturing the ET magnitude, variability, and trend over China.

Fig. 12.
Fig. 12.

Linear trends (mm yr−1 decade−1) of annual evapotranspiration during 2001–15 for different products. The black dots represent the 95% significance level. The boxes indicate three hot spots showing significant ET changes.

Citation: Journal of Hydrometeorology 24, 2; 10.1175/JHM-D-22-0135.1

d. Soil moisture

Figure 13a shows that the observed surface soil moisture is higher in south and east China (0.3–0.45 m3 m−3) and lower in North China (<0.15 m3 m−3). The OBS/CSSPv2 captures this pattern, but underestimates the soil moisture in Xinjiang and northeast China (from −0.03 to −0.05 m3 m−3) and overestimates it in central and eastern China (from 0.03 to 0.05 m3 m−3) (Figs. 13b,c). The other products generally show a wet bias, except in the Xinjiang region, and the magnitude of the bias is larger than that of the OBS/CSSPv2, especially for the ESA CCI, ERA5_Land, FLDASv1.0_Noah, and ERA5. The national mean absolute bias is 0.057 m3 m−3 for the OBS/CSSPv2 and 0.063–0.11 m3 m−3 for the other products. According to the regional averages (Fig. S4), the OBS/CSSPv2 has a significantly lower wet bias than the global products in east, south, and southwest China, where the vegetation cover is dense. The evaluation results are similar for the root-zone soil moisture (Fig. S5), and the absolute bias of the OBS/CSSPv2 is 18% lower compared to other products. It is possible that these biases were caused by other processes (e.g., imperfect model physics) because we masked out the soil moisture stations that might have been influenced by irrigation using the irrigation map created by Xiang et al. (2020) (see Text S1 for detailed information).

Fig. 13.
Fig. 13.

(a) Observed and (b) CSSPv2 simulated mean surface soil moisture in the growing season (April–September) during 2013–17. (c)–(k) The biases between the observations and different products [e.g., ΔSM(OBS/CSSPv2) = SMOBS/CSSPv2 − SMOBS].

Citation: Journal of Hydrometeorology 24, 2; 10.1175/JHM-D-22-0135.1

Figure 14 shows the soil moisture evaluation results based on different metrics at daily and monthly time scales. Both the automatic observed and the gravity-based soil moisture observations were used to calculate the monthly statistics. The OBS/CSSPv2 has the highest KGE and CC and the lowest RMSE and unRMSE at both depths and time scales. The median KGE of the OBS/CSSPv2 is 0.25 at the surface layer and daily scale, slightly larger than that of the ERA5 product and 0.03–0.16 larger than that of the other datasets. The KGEs for the daily root-zone SM are smaller than that of the surface SM for all products, but the OBS/CSSPv2 still has the highest KGE (0.23) that is 0.05–0.12 larger than other products (Fig. 14). At the monthly scale, the median KGE is 0.16–0.24 for the OBS/CSSPv2 and 0.04–0.20 for the global products. The OBS/CSSPv2 shows better performance for simulating the root-zone SM than the ERA5 and ERA5_Land, which have relatively higher KGEs for the surface layer soil moisture. Compared with the satellite-retrieved ESA CCI product, the OBS/CSSPv2 has a larger CC, resulting in a higher KGE. The regional metrics (Fig. S6) demonstrate a significantly higher performance of the OBS/CSSPv2 than the ESA CCI in east, south, and southwest China. The reason is that the soil moisture derived from the ESA CCI has high uncertainty in regions with dense vegetation due to the physical limitations of microwave remote sensing (Dorigo et al. 2017; Pan et al. 2014). Generally, the CC of the OBS/CSSPv2 is 0.11 higher and the mean RMSE and unRMSE are 15% and 17% lower than the metrics of the global products.

Fig. 14.
Fig. 14.

Evaluation of different products against observed surface (10 cm) and root zone (0–100 cm) soil moisture at (left) daily and (right) monthly scales. The evaluation metrics include the Kling–Gupta efficiency (KGE), correlation coefficient (CC), root-mean-square-error (RMSE), and unbiased root-mean-square-error (unRMSE). The boxes represent the 25th and 75th percentiles, the line is the median value, and the whiskers (dashed vertical lines) represent 99% of the data.

Citation: Journal of Hydrometeorology 24, 2; 10.1175/JHM-D-22-0135.1

The observed and simulated linear trends of the surface soil moisture during 1992–2013 are represented in Fig. S7. Although the results of all products show smaller trends than the observations, the OBS/CSSPv2, ESA CCI, and GLEAMv3.5 capture the increasing trends in north China and the eastern part of northwest China. The other products generally show drying (ERA5_Land, GLDASv2.2_CLSM, and ERA5) or wetting (FLDASv1.0_Noah and MERRA2) soil moisture patterns without the regional contrast trends that are shown in observations. Jia et al. (2018) suggested that precipitation is the primary factor responsible for the soil moisture trends. Therefore, the high-resolution precipitation data, which fuse numerous station observations and have a high quality (both in accuracy and long-term trend), appear to be the main reason for the OBS/CSSPv2 to reproduce the surface soil moisture trend.

e. Land surface temperature

Figures 15a and 15b show that the OBS/CSSPv2 LST has similar spatial distributions with MYDv21 LST, and the KGE exceeds 0.7 in most areas. Small KGEs occur in regions with complex terrain, such as the periphery of the Tibetan Plateau, where satellite retrievals may have high uncertainty. To further scrutinize the improvement of OBS/CSSPv2, Figs. 15d–i present the difference in the KGE between the OBS/CSSPv2 and different global products. A positive ΔKGE indicates the OBS/CSSPv2 has better performance and vice versa. Positive values occur mainly over Tibet, Xinjiang, northeast China, and some parts of the southern and eastern regions. The boxplots of the ΔKGE for the urban grids (urban fraction > 0) are shown in the lower-left corner. The following is observed. First, the ΔKGE is not significantly positive or negative when the urban fraction is less than 15%. However, when the urban fraction is larger than 15%, about 75% of the grids show a positive ΔKGE, indicating the robustness of the OBS/CSSPv2 to represent urban LST. Second, although the GLDASv2.1_Noah and FLDASv1.0_Noah consider urban grids using a bulk scheme (e.g., changing the hydrological and thermal properties), the OBS/CSSPv2 has a larger KGE in urban grids than these two products. Possible reasons are the fine spatial resolution and the use of an urban canopy model in the OBS/CSSPv2. Third, the mean median ΔKGE values are 0.019, 0.03, 0.035, and 0.037 for grids with urban fractions of 0%–5%, 5%–15%, 15%–30%, and >30%, respectively (the values are calculated from all global products), suggesting that the improvement of the OBS/CSSPv2 LST over the other products increases with an increase in the urban fraction.

Fig. 15.
Fig. 15.

Mean land surface temperature (LST) during 2003–11 for the (a) MYDv21 satellite observations and (b) OBS/CSSPv2 simulation. (c) Kling–Gupta efficiency (KGE) of the CSSPv2-simulated LST and the MYDv21 LST. The KGE was calculated based on the monthly LST during 2003–11. (d)–(i) Difference in KGE between the OBS/CSSPv2 and other products [e.g., ΔKGE(ERA5_Land) = KGEOBS/CSSPv2 − KGEERA5_Land]. The ΔKGEs for the urban grids were separated into four groups according to the urban fraction, and the results are shown in the lower-left corners of (d)–(i).

Citation: Journal of Hydrometeorology 24, 2; 10.1175/JHM-D-22-0135.1

Another advantage of high-resolution land surface modeling is its capability to capture UHIs (Ji et al. 2021). Figure 16 shows the mean LST maps of three metropolitan areas. Several UHIs are observed in the MYDv21 LST map because they have warmer LST than the surrounding rural grids (Figs. 16a,i,q). The OBS/CSSPv2 LST captures the UHIs quite well, and the spatial CCs between the OBS/CSSPv2 and MYDv21 are 0.81–0.91. The GLDASv2.1_Noah, FLDASv1.0_Noah, and FLDASv1.0_Noah also depict UHIs due to the bulk urban representation in the Noah model. The other products do not detect UHIs and have lower spatial CCs with the MYDv21. The spatial CC of OBS/CSSPv2 is slightly higher than the FLDASv1.0_Noah in the Beijing–Tian–Heibei and Yangtze River delta regions, while it is 0.11 higher than the other products. These results demonstrate the superiority of the OBS/CSSPv2 LST in urban areas.

Fig. 16.
Fig. 16.

Spatial distributions of the mean annual land surface temperature during 2003–11 over three metropolitan areas for different products. The spatial correlations between the satellite retrieval (MYDv21) product and the other datasets are shown in the lower-right corners.

Citation: Journal of Hydrometeorology 24, 2; 10.1175/JHM-D-22-0135.1

4. Attribution for the CSSPv2 improvement

Figure 17 shows the composite color map of the relative contributions of meteorological forcings, surface data, and CSSPv2 LSM. The contributions of each factor were calculated based on the results of the sensitivity experiments (Table 3) and Eqs. (1)(6), and the color in each location on the red–green–blue map was then determined.

Fig. 17.
Fig. 17.

Color composite map of the relative contributions of three factors (i.e., meteorological forcings, surface parameters, and model physics) to the higher KGE for the OBS/CSSPv2 than the ERA5 for monthly streamflow (Streamflow), snow depth (Snow), evapotranspiration (ET), root-zone soil moisture (Rootzone SM), and land surface temperature (LST). The red color indicates that the higher KGE is attributed to the fine-resolution surface parameters (including soil texture, vegetation cover, LAI, and land use category). Similarly, green and blue colors indicate that the meteorological forcing and land surface model are dominant factors, respectively.

Citation: Journal of Hydrometeorology 24, 2; 10.1175/JHM-D-22-0135.1

a. Streamflow

Meteorological forcing is the dominant factor contributing to the improved performance of the OBS/CSSPv2 for streamflow simulation in most catchments (Fig. 17a). Previous research has revealed the large dependence of runoff parameters on precipitation data (Guo and Su 2019; Sirisena et al. 2018; Strauch et al. 2012). Thus, it is not surprising to obtain lower streamflow simulation performance when the high-resolution forcings are replaced by the ERA5 forcings. The surface parameters are dominant in a few catchments (red areas in Fig. 17a) or have a similar importance as the meteorological forcings (light-red areas in Fig. 17a). The surface parameters influence the streamflow simulation by changing the soil hydraulic properties (e.g., the hydraulic conductivity, matric potential, and porosity), affecting the infiltration speed and runoff generation.

b. Snow depth

Overestimation of snow depth by reanalysis products has been noticed in previous studies over the Tibetan Plateau (Orsolini et al. 2019) and the Tianshan Mountains (Li et al. 2022). Orsolini et al. (2019) suggested that the overestimation of ERA5 was directly related to excessive snowfall in meteorological forcings. Here we find similar results, as the CSSPv2 model, when driven by the ERA5 meteorological forcings (ERA5Forc/CSSPv2), shows a clear positive bias of the snow depth in northeast China, Tibet, and the western part of northwest China (Figs. S8a,b), suggesting that high-resolution and high-quality meteorological forcings are critical for modeling snow depth in these regions.

Although Yang et al. (2020) found that using more realistic vegetation parameters could improve the snow simulation, especially in forested regions, our research shows minor differences between the ERA5Forc/CSSPv2 and ERA5ForcSurf/CSSPv2 results. A possible reason is that the CMA snow depth stations are predominantly located in regions with barren land or low herbaceous plants. Although using the same coarse-resolution forcings and surface parameters from ERA5, the ERA5ForcSurf/CSSPv2 simulation exhibits less snow depth overestimation than the ERA5 reanalysis in central Tibet, east China, and north China (Figs. S8c,d). Studies have shown that a snow albedo scheme that only considers snow age and/or snow cover (e.g., the scheme in the HTESSEL model) tends to overestimate snow albedo, decreasing the absorption of incoming solar radiation, slowing the snow melting rate, and thus overestimating snow depth (L. Liu et al. 2022; Malik et al. 2014). In contrast, a snow albedo scheme that considers multiple snow properties (e.g., the scheme in the CSSPv2 model) is found to outperform the simple scheme in simulating snow albedo and other snow-related variables (Liu et al. 2019). This may be the reason for less overestimation of snow depth by the ERA5FocSurf/CSSPv2. The results in Fig. 17b, i.e., the LSM (blue) and meteorological forcings (green) are attributed as dominant factors, are consistent with those shown in Fig. S8. The mean contributions of the model physics, meteorological forcings, and surface parameters at the national scale are 69.4%, 26.9%, and 3.7%, respectively.

c. Evapotranspiration

The OBS/CSSPv2 has significantly higher performance for simulating ET than the global land/atmosphere reanalysis datasets, especially in Xinjiang; Tibet; and east, south, and southwest China (section 3c). Figure 18 shows the seasonal cycle of regional ET during 2001–15 simulated by CSSPv2 with different meteorological forcings and surface parameters. The difference between ERA5 and ERA5SurfForc/CSSPv2 shows that using the CSSPv2 model significantly decreases the ET overestimation (from the blue line to the pink line in Figs. 18a,b). The fine-resolution surface parameters further improve ET simulation in nearly all subregions (ERA5SurfForc/CSSPv2 versus ERA5Forc/CSSPv2), indicating the importance of surface parameters for ET simulation. However, the high-resolution meteorological forcings have relatively smaller influence on the ET simulation than the LSM and surface parameters (ERA5Forc/CSSPv2 versus OBS/CSSPv2). The results in Fig. 17c, where red and blue colors dominate all the subregions except some parts of the northeast China, are consistent with those in Fig. 18. We also calculated the mean regional contributions by averaging the statistics of each grid. It is found that the CSSPv2 model and the surface parameters have similar contributions (about 30%–40%) in different regions, whereas meteorological forcings have relatively smaller (around 20%–30%) contributions.

Fig. 18.
Fig. 18.

Contribution separation for the improved evapotranspiration (ET) simulations in nine subregions in China. The ERA5SurfForc/CSSPv2 is the same as the OBS/CSSPv2, but the meteorological forcing and surface parameters are the same as ERA5. ERA5Forc/CSSPv2 is the same as OBS/CSSPv2, but the meteorological forcing is from ERA5.

Citation: Journal of Hydrometeorology 24, 2; 10.1175/JHM-D-22-0135.1

The reason for the influences of the meteorological forcing and surface parameters is that high-resolution datasets provide more detailed information. However, the CSSPv2 model can also reduce the positive ET biases of ERA5 with the same coarse-resolution meteorological forcings and surface parameters. Bonan et al. (2011) suggested that considering the colimitation of photosynthesis (i.e., colimitation among Rubisco-, light-, and export-limited rates) reduces the photosynthesis rate, resulting in smaller stomatal conductance and a lower ET. The CSSPv2 model considers the colimitation of photosynthesis in its photosynthesis-stomatal conductance model (Dai et al. 2004), whereas the HTESSEL model used by ERA5 does not. Thus, the less ET overestimation of the ERA5SurfForc/CSSPv2 than the ERA5 can be explained by differences in the physiological parameterizations. However, different parameterizations for the same vegetation type (e.g., roughness, vegetation height, leaf structure), different canopy radiative schemes and different soil water modules may also have impacts.

d. Soil moisture

Figure 19 shows the regional mean root-zone soil moisture from different experiments. Generally, the ERA5SurfForc/CSSPv2 has higher performance than the ERA5 in north, southwest, south, and east China, indicating that the CSSPv2 LSM is advantageous in over these regions. Different from the HTESSEL model used by ERA5 reanalysis dataset, CSSPv2 uses the VAST model which considers the influence of subgrid topography on soil moisture. In addition, the finer soil layer discretization in CSSPv2 LSM can provide more vertical variability of soil hydraulic properties and more realistic soil moisture variation, improving the soil moisture simulation results (Dantec-Nédélec et al. 2017; Shellito et al. 2020). The ERA5Forc/CSSPv2 (with high-resolution surface parameters) has a smaller bias than the ERA5SurfForc/CSSPv2 (with ERA5 surface parameters) in Xinjiang and in south, east and southwest China, demonstrating the added value of using high-resolution surface parameters. Although high-resolution forcings also result in a lower SM bias in southwest, northwest, south, and east China (OBS/CSSPv2 versus ERA5Forc/CSSPv2), their effects are much smaller than that of the CSSPv2 LSM and surface parameters over south and east China (Fig. 19). Thus, the LSM and surface parameters are responsible for the higher KGEs in east, south, north, and northeast China (Fig. 17d), with the mean contribution being 38%–51% and 30%–36%, respectively. In contrast, the meteorological forcings play an equally important or even dominant role in southwest China and northwest China, where most reanalysis meteorological forcing products have large uncertainty (Fig. 17d; Yang et al. 2017).

Fig. 19.
Fig. 19.

Observed and simulated regional daily root-zone (0–1 m) soil moisture averaged over 2013–17 for different experiments. The ERA5SurfForc/CSSPv2 is the same as the OBS/CSSPv2, but used the meteorological forcing and surface parameters from ERA5 dataset. The ERA5Forc/CSSPv2 is the same as OBS/CSSPv2, but used the ERA5 meteorological forcing. Only stations where the OBS/CSSPv2 has higher KGE values than the ERA5 were used.

Citation: Journal of Hydrometeorology 24, 2; 10.1175/JHM-D-22-0135.1

e. Land surface temperature

Figure 17e illustrates that the improvement in the OBS/CSSPv2 simulated LST is attributed to the meteorological forcings and CSSPv2 LSM. Meteorological forcings have a dominant contribution in the northern part of Xinjiang and Tibet, where few observations are used in the global reanalysis products, such as ERA5. The LSM has a dominant influence in the south and northeast China with dense vegetation. The two-big-leaf model in CSSPv2 considers radiation and thermal processes for sunlit and shaded fractions of the canopy, which may lead to a better representation of the canopy radiative process and improve the LST simulation. The higher KGEs in urban grids are attributed to the urban canopy model in the CSSPv2 LSM (Ji et al. 2021).

f. Long-term trends

Recent studies have shown that, changes in snow depths, ET, and soil moisture are dominated by climate change (e.g., changes in temperature and precipitation), although some ecological factors (e.g., vegetation greening/withering and land use change) may have influences (Jia et al. 2018; Ma and Zhang 2022; Meng et al. 2020; Zhang et al. 2020). Therefore, the improved simulation of long-term trends of snow, ET, and surface soil moisture in OBS/CSSPv2 experiment is explained by the meteorological forcings here.

Figure 20 shows the linear trends of precipitation during different seasons and periods. Similar to Zhang et al. (2020), the increasing snow depth in the OBS/CSSPv2 experiment in Xinjiang and northeast and northwest China from 1980 to 2009 is consistent with the significant increase in winter precipitation (Figs. 20a and 6a). However, the precipitation in ERA5 dataset shows insignificant and smaller trends (Fig. 20b). Therefore, the ERA5 reanalysis product cannot reproduce the observed trend of winter snow depth. Similarly, the increasing surface soil moisture in north China and some parts of northeast China derived from OBS/CSSPv2 is directly related to the increased precipitation in CN051 (Fig. S4 and Fig. 20c), whereas the ERA5 generally shows decreasing precipitation trends in eastern China. The significant increase in ET over the Hetao region and northeast China (Fig. 12b) is also consistent with the increased precipitation (Fig. 20e). The ERA5, however, does not depict the increasing precipitation trend in the Hetao region, resulting in an insignificant increasing or decreasing ET trend (Figs. 20f and 12i). Thus, when the CSSPv2 is driven by the ERA5 meteorological forcing (e.g., ERA5Forc/CSSPv2), the snow depth, soil moisture and ET show drying patterns similar to those of the ERA5 dataset (not shown).

Fig. 20.
Fig. 20.

Comparison of precipitation (Preci) trends between the bias-corrected CN051 and the ERA5 datasets. (a),(b) Spatial distributions of linear trends of winter (December–February) precipitation during 1980–2009 derived from CN051 and ERA5 datasets. (c)–(f) As in (a) and (b), but for the growing season (April–September) precipitation during 1992–2013 and annual mean precipitation during 2001–15, respectively.

Citation: Journal of Hydrometeorology 24, 2; 10.1175/JHM-D-22-0135.1

5. Conclusions and discussion

This research used the CSSPv2 LSM, which has advanced hydrological parameterizations and considers urban dynamics, to perform a high-resolution (6 km) and long-term (1979–2017) land surface modeling over continental China. Main contributing factors to the improved performance of CSSPv2’s high-resolution simulation as compared with state-of-the-art global products were also analyzed.

The CSSPv2 LSM is proved an efficient tool for simulating the terrestrial hydrological and thermal processes over China using observations from thousands of in situ stations and high-resolution satellite-based products. The RMSEs of winter snow depth during 1980–2009, annual ET during 2003–10, and daily surface and root-zone soil moisture during 2013–17 for OBS/CSSPv2 high-resolution simulation were 16 mm, 153.57 mm, 0.07 m3 m−3, and 0.06 m3 m−3, respectively, and all the metrics were calculated based on observations over entire China. Compared with the nine global products, including ERA5-Land, ESA CCI v06.1, FLDASv1.0, and GLEAMv3.5, CSSPv2’s high-resolution simulation decreased the RMSEs of snow depth, ET, and soil moisture by 67%, 29%, and 12%–19%, respectively. The high-resolution simulation also captured changing patterns of snow, ET and soil moisture with the reference data, whereas the other global reanalysis products did not detect significant changes in these parameters. Moreover, the global reanalysis datasets overestimated the trend of ET and snow depth by more than 2 times, especially at local scales. The calibration resulted in a higher median KGE (0.66) of monthly streamflow in 83 catchments simulated by the CSSPv2. This value was 0.17–0.54 higher than that of the global land reanalysis products. The urban module in the CSSPv2 LSM enabled the high-resolution simulation to capture UHIs in metropolitan areas with a high (0.81–0.91) spatial correlation coefficient.

By conducting another two simulations with coarse resolution forcings and surface parameters from the ERA5 dataset, we further provided insights in the main reasons for the improvement in the CSSPv2 simulations. Although the high-resolution meteorological forcings contributed substantially to improving the simulation results in western China, where few observations can be used in the reanalysis datasets, the CSSPv2 LSM model was a main contributing factor to the improved simulations of ET, root-zone soil moisture, snow depth, and LST in eastern China where the vegetation is dense and urbanization is significant. Possible reasons are the advanced physical parameterizations in the CSSPv2 LSM, including the colimitation of photosynthesis, the volume-averaged soil water transport module, the sophisticated two-big-leaf model, the five-layer snow model with an albedo scheme that considers multiple snow properties, and the urban canopy module.

The high precision of the CSSPv2’s high-resolution simulation in representing terrestrial hydrological variables shows promise for its further applications in China to provide locally relevant information (e.g., drought monitoring systems, hydrological influences of vegetation greening, and long-term terrestrial hydrological changes). The superiority of the high-resolution simulation over the state-of-the-art global reanalysis products emphasizes the necessity of performing high-resolution land surface modeling to investigate local water and energy processes. The dominant contribution of advanced model parameterizations to the improved simulation as discussed in section 4 can shed lights on LSM development. For example, future improvements of the HTESSEL model should focus on snow albedo parameterizations, colimitation of photosynthesis, soil moisture subgrid variations, and urban dynamics. It should be noted that, we used a 6-km resolution for the CSSPv2 model due to the limited data availability from the spatial resolution of meteorological forcings, although the CSSPv2 model can be applied at 1-km or higher (e.g., 100 m) resolutions (Ji et al. 2017). Continued efforts should be devoted to developing high-resolution meteorological forcings and investigating the performance of the CSSPv2 model at 1-km (or higher) spatial resolution.

This research has some uncertainties and limitations. First, the influence of water management (e.g., irrigation and reservoir operations) was not considered, causing some uncertainties during the streamflow evaluations. However, the influence of reservoir operations may not have impacted the significance of our results because the evaluation period (1980–89) was before the rapid construction of reservoirs (Song et al. 2021; Zhang et al. 2008). A new version of the CSSP model is being developed, by considering human water regulations, which is expected to improve the representation of water and energy cycle over the regions with heavy human activities. Second, Rouf et al. (2020) suggested that a dynamic lapse rate leads to more accurate results than a fixed lapse rate during temperature downscaling. We performed the temperature and dewpoint downscaling with a dynamic lapse rate and repeated the simulation during 2000–17. Minor differences are found in hydrological variables (e.g., soil moisture, not shown), which is consistent with the recent research (Rouf et al. 2021). For the land surface temperature, the change of KGE after using a dynamic lapse rate is highly heterogeneous and the national mean of ΔKGE is only 0.002 (Fig. S9). Therefore, uncertainties caused by using a fixed lapse rate have limited influences on our results. Finally, we focused on the positive effects of high-resolution forcings, surface parameters and LSMs during the attribution. However, the KGE may decrease after using these high-resolution factors, which may be related to data uncertainties and model deficits. Future studies should use ensemble simulations with different combinations of forcings, parameters, and models to investigate this issue comprehensively.

Since there is no public high-resolution and long-term regional land reanalysis dataset in China, we released the OBS/CSSPv2 high-resolution simulation on the National Tibetan Plateau/Third Pole Environment Data Center to facilitate research on terrestrial water and energy changes in China. The dataset will be updated upon with the meteorological forcing datasets and the surface parameters (e.g., urban areas).

Acknowledgments.

This work was supported by National Key R&D Program of China (2018YFA0606002), National Natural Science Foundation of China (41875105), Natural Science Foundation of Jiangsu Province for Distinguished Young Scholars (BK20211540), Natural Science Foundation of Jiangsu Province (BK20220460), and the Startup Foundation for Introducing Talent of NUIST.

Data availability statement.

The CSSPv2 high-resolution simulation result is available in the National Tibetan Plateau/Third Pole Environment Data Center (https://doi.org/10.11888/Terre.tpdc.271898). FLUXCOM ET is available at the Max-Planck-Institute for Biogeochemistry (https://www.bgc-jena.mpg.de/geodb/). GLEAM SM and ET are from the https://www.gleam.eu/. ESA CCI SM is available at https://www.esa-soilmoisture-cci.org/. The MYDv6.1 LST is from the Land Processes Distributed Active Archive Center (LPDAAC) (https://lpdaac.usgs.gov/products/myd21c3v061/). ERA5_Land and ERA5 are available at https://cds.climate.copernicus.eu/, while GLDASv2.0_Noah, GLDASv2.1_Noah, GLDASv2.0_CLSM, GLDASv2.2_CLSM, FLDASv1_Noah, and FLDASv1.0 are all from the https://disc.gsfc.nasa.gov/datasets/. TerraClimate is available at https://climate.northwestknowledge.net/TERRACLIMATE/index_directDownloads.php.