Introduction

Human water use significantly influences the natural water cycle within the broader context of global change1,2,3. Over the past century, global water use has surged from approximately 500 km3 to 4000 km3 due to increasing urbanization, rising food demand, and socio-economic development4,5. The increasing water demand, coupled with pronounced temporal and spatial mismatches in water resource availability, has exacerbated global water scarcity, posing a significant barrier to achieving the United Nations Sustainable Development Goals (SDGs)6,7. In China, water use has steadily risen in recent decades, with over 40% of cities experiencing varying levels of water shortages8,9. Seasonal and spatial disparities between water supply and demand have further intensified summer water scarcity, especially in economically developed urban areas and arid regions10,11,12. Accurately identifying trends in water use is essential for informing effective water resource management policies and addressing water scarcity. Developing high-quality water use datasets is crucial for advancing the understanding of water resource utilization patterns.

Current water use data are mainly derived from surveys and hydrological model outputs, or generated as gridded datasets via downscaling. Survey-based water use information, collected by organizations via field surveys and data integration, accurately reflects regional water usage patterns6,13,14. The Food and Agriculture Organization (FAO) established the AQUASTAT global water and agriculture information system to collect national-level agricultural, domestic, and industrial water use data every five years13. Since 1997, China’s Ministry of Water Resources has published annual water resources bulletin based on administrative divisions, encompassing data on domestic, production, and ecological water use. However, the processes of data collection, calibration, and compilation for these statistical surveys are resource-intensive, requiring substantial labour and financial investment. This limitation leads to datasets with limited spatial coverage and mostly annual resolution, making it difficult to capture seasonal variations in water use driven by climate and socio-economic factors13,15,16. Consequently, there is a critical need for high spatial resolution, temporally detailed water use datasets to enhance our understanding of water resource dynamics and to support effective water management strategies.

Gridded water use data products derived from model simulations and spatiotemporal downscaling methods offer high temporal continuity and finer spatial coverage, making them widely applicable in water resource management research. Table 1 summarizes representative gridded water products, including their generation methods, temporal and spatial resolutions, and sector distinctions14,17,18,19,20,21,22. Early developments of global hydrological models (GHMs) and land surface models (LSMs), such as WaterGAP and PCR-GLOBWB, integrated human water use into hydrological cycle simulations at the grid level by considering regional development levels and meteorological variables17,18,20,21,22,23,24. For example, the PCR-GLOBWB v2.0 model simulates irrigation water use by parameterizing the feedback between irrigation and soil moisture, while modeling domestic and industrial water use based on population and GDP distributions, with adjustments for predefined technological advancement coefficients22. The accuracy of these simulations relies heavily on the choice of parameterization schemes for human water use activities, which are often subjectively determined. Another issue is that structural differences among models can lead to significant variability in simulated water use characteristics across different GHMs, making it challenging to generate consistent data products.

Table 1 Comparison of representative characteristics of gridded water use datasets.
Full size table

Downscaling methods effectively combine the accuracy of survey-based data with the high temporal and spatial resolution of model outputs. By modeling the relationship between survey-based water use data and gridded socio-economic variables, these methods disaggregate water use into grid-based datasets. At the global scale, Huang et al. 2018 utilized GHMs and global socio-economic datasets to downscale national water use statistics released by the FAO to a 0.5° × 0.5° grid. This effort produced a global multi-sector gridded water use dataset for the period 1971–2010, which has been widely adopted in water resources research14. At the national scale, Ma et al. 2020 developed a 0.25° × 0.25° grid water use dataset for China by incorporating multiple socio-economic datasets, which enabled a more detailed assessment of monthly water scarcity conditions12.

The limited transparency of water use data among Chinese enterprises, coupled with the lack of detailed geospatial data on industrial water use, has led previous studies to assume a direct correlation between macro socio-economic indicators and the spatial distribution of industrial water use14,16. While data such as GDP reflect the overall economic value generated by industrial activities within a region, they do not directly capture the processes involved in industrial water use. In contrast, company-specific output data, which are typically collected for individual enterprises, are more directly linked to industrial water withdrawal than GDP data15. For irrigation water withdrawal, some studies have employed outputs from GHMs to spatially downscale national irrigation statistics, thereby accounting for the complexity of irrigation processes14,25,26. However, GHMs often rely on complex sub-models to parameterize variables such as crop types, growing seasons, agricultural management practices, and irrigated areas. This introduces significant variability in results, depending on the specific parameterization schemes used, which in turn generates considerable uncertainty in the resulting gridded irrigation water data27,28. Furthermore, the challenge of obtaining detailed data on agricultural management and crop coefficients limits the application of GHMs in data-scarce regions. Given that the spatial distribution of irrigation water is primarily determined by the extent of irrigated cropland, alternative models utilizing more readily available cropland area maps could be considered to more efficiently generate the spatial distribution of irrigation water29.

Water use exhibits significant inter-annual variability and intra-annual fluctuations, driven by seasonal factors such as climatic conditions and variations in electricity demand6. For instance, irrigation water use tends to surge during heatwaves, while prolonged extreme events, such as droughts, can result in resource depletion and a subsequent reduction in water use30,31,32. However, most existing water use data products overlook seasonal variations, typically focusing only on inter-annual changes in sectoral water use or assuming a uniform distribution of water demand throughout the year17,18,20,21. A limited number of studies have addressed seasonal variations in specific sectoral water use12,33,34. For example, estimates for thermal power cooling water use integrate the effects of temperature variations on electricity demand, alongside inter-monthly heating and cooling requirements33. Monthly irrigation water estimates are typically derived from crop-specific irrigation requirements, which vary seasonally12,34. However, aside from relatively detailed consideration of irrigation water use, existing water use datasets still lack comprehensive representation of seasonal variations across all sectors (irrigation, industrial, and domestic), limiting our understanding of the overall seasonality of water demand.

To address these gaps, this study employs remote sensing-derived land use data, population density data, reanalysis meteorological data, and geospatial data on thermal power plants and industrial enterprises to develop a downscaling algorithm. This algorithm utilizes multiple covariates to downscale provincial water use statistics for China from 1965 to 2022, thereby refining the spatial resolution and seasonal distribution of water use data. The result is a 0.1° monthly gridded water use dataset (HSWUD), which enhances the understanding of the impacts of human water use on the natural water cycle in the context of a changing environment. This dataset and its associated methodology will contribute to more refined water resources management, supporting decision-making in response to climate-induced challenges.

Data

This study utilizes a range of data sources, including panel water use statistics, socio-economic indicators, and meteorological reanalysis data. Specifically, the dataset includes annual sectoral water use, daily meteorological variables, land use and population density maps, as well as micro-level survey data from industrial enterprises and power plants (Fig. 1).

Fig. 1
figure 1

Spatial distribution of spatiotemporal downscaling covariates (a) Proportion of irrigated land in China in 2000 (b) Annual electricity generation of thermal power plants in China in 2017 (c) Distribution of manufacturing enterprises in 2007 (d) Population density distribution in China in 2000 (e) Daily average temperature from 1965 to 2022 (f) Monthly average potential evapotranspiration from 1965 to 2022 (g) Monthly average precipitation from 1965 to 2022.

Full size image

Statistic water use data

Water use in this study refers to the total gross volume of water withdrawn by different sectors, including conveyance losses during distribution. It is classified into three categories: agricultural water, industrial water, and domestic water (Ministry of Water Resources of the People’s Republic of China, Code of practice for water resources bulletin, 13.060.10, https://openstd.samr.gov.cn/bzgk/gb/newGbInfo?hcno=13D6942C70B4661CD83B2B3F1C87772B, (last access:5 November 2024), 2009). Since the driving factors for thermal power and manufacturing water withdrawal differ, industrial water withdrawal is further disaggregated into thermal power and manufacturing components33. These four sectors together account for 90% of China’s total water use from 1965 to 2013, effectively representing overall water use patterns9. The provincial water use data from 1965 to 2013 comes from the high-resolution water use survey dataset published by Zhou et al. (National Long-term Water Use Dataset of China, NLWUD). This dataset integrates results from China’s second water resource assessment starting in 2002 (reconstructed water use data from 1965 to 2000) and the provincial water resource bulletins from 2001 to 2013, covering 341 prefecture-level cities in China and encompassing three major categories (agricultural, industrial, and domestic water use) and 16 sub-sectors, which represent 93% of the total surface water and groundwater abstraction in China over the past few decades9. Since its release, this dataset has been widely used in studies related to water resources utilization35. Provincial surveyed-based water use data of 2014 to 2022 were sourced from China Water Resources Bulletin(http://www.mwr.gov.cn/sj/tjgb/szygb/, last access: 5 November 2024) and China Statistical Yearbook(https://www.stats.gov.cn/sj/ndsj/, last access: 5 November 2024).

Note that the “agricultural water use” category in the China Water Resources Bulletin includes not only irrigation, but also livestock and other agricultural uses, and thus cannot be directly equated with irrigation water use. For the period 2014–2022, we estimated provincial-level irrigation water use by multiplying the actual irrigated area (from the China Water Statistical Yearbook) by the irrigation water use per area of actual irrigated land (reported in the China Water Resources Bulletin).Detailed manufacturing water use data for 31 sub-sectors at the provincial level were obtained from the 2008 China Economic Census Yearbook (https://www.stats.gov.cn/sj/pcsj/jjpc/2jp/left.htm, last access: 1 April 2025). The data are based on a survey of enterprises above a designated production threshold (annual output > CNY 5 million).

Socio-economic and meteorological data

The spatiotemporal downscaling covariates involved in this study are mainly divided into socio-economic data and meteorological data, as listed in Table 2:

Table 2 Data sources of spatiotemporal downscaling covariates.
Full size table

Annual thermal power generation data were obtained from the GPPD by the World Resources Institute, which integrates government statistics and remote sensing sources. This database includes over 28700 power plants and covers more than 80% of the global installed capacity36. Since this study only considers thermal power generation, it retains data for 1123 thermal power plants, which account for 75.32% of the national electricity generation for that year. Irrigated cropland map in China at 250 m resolution was derived from the CIrrMap250 dataset, which integrates multiple data sources using a semi-automated machine learning approach and offers relatively high accuracy compared to existing datasets37. Annual total output dataset from industrial enterprises comes from CIED, which is collected by the National Bureau of Statistics and includes over 310000 industrial enterprises from 38 subsectors. Due to several changes in statistical criteria during the reporting period of this database, this study uses data from 2007 as the spatial reference variable for downscaling manufacture water use38. The CIED lacks detailed geographic coordinates and contains missing values. To enable spatial analysis, we conducted the following preprocessing: (1) removed enterprises marked as “not opened”; (2) used the GeoPy library in Python 3.11 to geocode enterprise names. The final dataset includes 28989 enterprises, covering 86.41% of total industrial output for that year. In addition, we used monthly sales revenue of the manufacturing subsectors from the China Industrial Database at the provincial level as a proxy for intra-annual variation in manufacturing water use.

Methods

The generation of the dataset involves two main steps: (1) Spatially distributing provincial-level annual water use to a 0.1° × 0.1° grid using socio-economic variables to produce annual distribution maps; and (2) Temporally disaggregating the annual maps into monthly grids using meteorological and monthly economic data. The flowchart is shown in Fig. 2.

Fig. 2
figure 2

The workflow for generating the HSWUD.

Full size image

Spatial allocation of provincial water withdrawal based on socio-economic variables

Annual sectoral water withdrawal is downscaled to grid cells based on the spatial distribution of socio-economic variables, enabling the conversion of provincial-level data from 1965 to 2022 into gridded datasets. For irrigation, domestic, and thermoelectric cooling water withdrawal, we adopted a proportional allocation approach following previous studies14,34. Considering the substantial changes in land use patterns in China after 200039—driven by rapid urbanization and agricultural policy shifts40—we used irrigated area maps updated every five years for the post-2000 period to better reflect temporal dynamics in irrigation water withdrawal allocation.

Using the GeoPandas package in Python 3.11, the pixel values of socio-economic variables were normalized by their respective provincial totals to derive spatial downscaling weights \({\alpha }_{i,j}\):

$${\alpha }_{i,j}=\frac{{x}_{i,j}}{\mathop{\sum }\limits_{j=1\,}^{n}{x}_{i,j}},$$
(1)

where \({x}_{i,j}\) is the pixel value of a socio-economic variable in pixel j within province i, and n is the total number of pixels withi-n that province.

By performing raster spatial calculations between the spatial downscaling weight factor and the panel water withdrawal data for the province to which the pixel belongs, the provincial annual water withdrawal is allocated to grid points, resulting in a 0.1° × 0.1° sectoral water withdrawal spatial distribution map:

$${{WU}}_{j}={\alpha }_{j}\times {{WW}}_{{\rm{i}}},$$
(2)

where \({\alpha }_{j}\) is the spatial downscaling weight factor for the j pixel, \({{WW}}_{{\rm{i}}}\) is the sector water withdrawal for province i and \({{WU}}_{j}\) is the sector water withdrawal for the j pixel.

For manufacturing water withdrawal, we applied a more detailed enterprise-level approach that accounts for differences in water use efficiency among industrial enterprises within each province15. For province i and manufacture subsector m, we calculated the water use efficiency \({WU}{E}_{i,m}\) as:

$${WU}{E}_{i,m}=\frac{{Outpu}{t}_{i,m}}{W{U}_{i,m}}$$
(3)

where \({Outpu}{t}_{i,m}\) is the manufacturing subsector m output value of province i and \(W{U}_{i,m}\) is the corresponding manufacturing water withdrawal.

Assuming a uniform water use efficiency within each province-subsector combination, the water withdrawal of an individual enterprise k (belongs to subsector m) \(W{U}_{i,k}\) was estimated as:

$$W{U}_{i,k}={WU}{E}_{i,m}\times {Outpu}{t}_{i,k}$$
(4)

where \({WU}{E}_{i,m}\) is the water use efficiency of subsector m; \({Outpu}{t}_{i,k}\) refers to the annual output of individual enterprise k in province i. These enterprise-level estimates were then aggregated from point to grid cells, and the sum across all 31 manufacturing subsectors yielded the gridded manufacturing water withdrawal map.

Temporal disaggregation of annual gridded water withdrawal based on meteorological data

Using multi-source data, each pixel value in the annual gridded water withdrawal map is allocated to a monthly time scale. For this temporal disaggregation, we assumed no seasonal variation in water use efficiency within a year. Thus, seasonal patterns in water withdrawal were approximated using proxy variables that reflect monthly activity levels for each sector:

$${WU}_{j.t}={{WU}}_{j}\times {{fraction}}_{t}$$
(5)

Where \({{WU}}_{j}\) is annual water withdrawal in pixel j, \({{fraction}}_{t}\) is the monthly proportion of each sector in month t.

For the irrigation sector, monthly water withdrawal was estimated based on crop-specific irrigation demand using a method adapted from the CROPWAT model. We first calculated daily net irrigation requirement (NR) at pixel level as the difference between crop evapotranspiration (computed as the product of potential evapotranspiration and crop coefficient) and effective precipitation:

$$N{R}_{j,d}={k}_{c,j,d}\times {{PET}}_{j,d}-{{PE}}_{j,d}$$
(6)

where \(N{R}_{j,d}\) s the daily net irrigation demand in grid cell j on day d;\(\,{k}_{c,j,d}\) is the crop coefficient for a specific crop in grid cell j and day d; \({{PET}}_{j,d}\) is the daily potential evapotranspiration; and \(P{E}_{d}\) is the daily effective precipitation.

Following the USDA’s guidelines for calculating effective precipitation, effective precipitation PE is computed based on daily precipitation41 \({P}_{d}\):

$${{PE}}_{d}=\left\{\begin{array}{c}{P}_{d}\times \frac{4.17-0.2\times {P}_{d}}{4.17},{P}_{d} < 8.3\\ 4.17+0.1\times {P}_{d},{P}_{d}\ge 8.3\end{array}\right.$$
(7)

We selected rice, wheat, and maize—three major crops in China—as representative crop types12, derived their crop coefficients and growing periods from MIRCA200042 and national standard for grade of agricultural drought (China Meteorological Administration, Grade of agricultural drought GB/T 32136-2015, https://www.cma.gov.cn/zfxxgk/gknr/flfgbz/bz/202209/t20220921_5098178.html, (last access:1 April 2025), 2009). The monthly irrigation demand was then obtained by aggregating the daily NR values within each month. Finally, the monthly proportion of annual irrigation demand was used to temporally allocate annual irrigation water withdrawal at the grid level.

For the thermal cooling sector, the temporal allocation coefficient of cooling water for each pixel was calculated based on the number of heating and cooling degree days for that year26,43:

$${TEL}{E}_{t}={p}_{b}\times \left({p}_{h}\frac{{HD}{D}_{t}}{\mathop{\sum }\limits_{t=1\,}^{12}{HD}{D}_{t}}+{p}_{c}\frac{{CD}{D}_{t}}{\mathop{\sum }\limits_{t=1}^{12}{CD}{D}_{t}}+{p}_{u}\times \frac{1}{12}\right)+\left.{p}_{{it}}\times \frac{1}{12}\right)$$
(8)

where \({HD}{D}_{t}\) is the number of heating degree days for the t month, and\(\,{CD}{D}_{t}\) is the number of cooling degree days for the t month. In this study, we estimated the parameter values using publicly available aggregated energy data. Where detailed data were unavailable, we applied reasonable assumptions based on relevant literature.

Specifically, we set: \({p}_{b}=0.4\) (proportion of electricity used in buildings) based on IEA sectoral electricity consumption data44, aggregating Residential, Commercial and Public Services, and Other Final Consumption according to the definition in the 2016 China Building Energy Consumption Report45.These categories accounted for approximately 32.5% of national electricity use between 1990 and 202046. The shares of heating, cooling, and other utilities (\({p}_{h},{p}_{c},{p}_{u})\) were each assumed to be 0.33 due to the lack of disaggregated end-use data.

According to the Design Code for residential buildings (GB50096-2011) (Ministry of Housing and Urban-Rural Development of the People’s Republic of China, Design Code for residential buildings, GB50096-2011, https://www.mohurd.gov.cn/zhengcefabu/index.html, (last access:5 November 2024), 2011), the indoor heating calculation temperature for bedrooms and living rooms in ordinary residences should not be lower than 18 °C, using 18 °C as the daily heating threshold. Based on the Design standard for energy efficiency of public buildings (GB50189-2005) (Ministry of Housing and Urban-Rural Development of the People’s Republic of China, Design standard for energy efficiency of public buildings, GB50189-2005, https://www.mohurd.gov.cn/zhengcefabu/index.html, (last access:5 November 2024), 2011), the standard design temperature for central air conditioning in summer is 26 °C to 28 °C, using 26 °C as the daily cooling threshold. During the study period, the number of monthly heating and cooling days at each grid point is calculated using the following formulas:

$${{HD}}_{d}^{t}=\left\{\begin{array}{cc}1, & {T}_{d}^{t} < 18^\circ {\rm{C}}\\ 0, & {T}_{d}^{t}\ge 18^\circ {\rm{C}}\end{array}\right.,$$
(9)
$${HD}{D}^{t}=\mathop{\sum }\limits_{d=1}^{m}{{HD}}_{d}^{t},$$
(10)

where \({T}_{d}^{t}\) is the daily average temperature on day d of month t, \({{HD}}_{d}^{t}\) is the heating day count for that day, and \({HD}{D}^{t}\) is the number of heating degree days for the t month.

$${{CD}}_{d}^{t}=\left\{\begin{array}{cc}0, & {T}_{d}^{t} < 26^\circ {\rm{C}}\\ 1, & {T}_{d}^{t}\ge 26^\circ {\rm{C}}\end{array},\right.$$
(11)
$${CD}{D}^{t}=\mathop{\sum }\limits_{d=1}^{m}{{CD}}_{d}^{t},$$
(12)

where \({{CD}}_{d}^{t}\) is the cooling day count for d day, and \({CD}{D}^{t}\) is the number of cooling degree days for the t month.

Based on the monthly temperature fluctuations throughout the year, the time allocation coefficient for domestic water withdrawal for each pixel is calculated annually12,47:

$${TDO}{M}_{t}=\frac{1}{12}\times \left(\frac{{T}_{t}-{T}_{{avg}}}{{T}_{\max }-{T}_{\min }}\times R+1\right),$$
(13)

where \({T}_{t}\) is the monthly average temperature for the t month; \({T}_{{avg}}\) is the average monthly temperature for that year; \({T}_{\max }\) is the maximum monthly average temperature for that year; \({T}_{\min }\) is the minimum monthly average temperature for that year; and \(R\) is a dimensionless parameter representing the amplitude of domestic water withdrawal between the coldest and warmest months. Based on previous studies, \(R\) is taken as 0.2 in this paper14,22.

The seasonal variation of manufacturing water withdrawal was approximated using sub-sector’s monthly sales revenue, calculated as the proportion of monthly revenue to its annual total:

$${{fraction}}_{m,t}=\frac{{{Sales}}_{m,t}}{\mathop{\sum }\limits_{t=1}^{12}{{Sales}}_{m,t}}$$
(14)

where \({{Sales}}_{m,t}\) is the month t sales revenue of each sub-sectors m within province; The resulting monthly revenue fractions \({{fraction}}_{m,t}\) was then used to distribute annual enterprise-level manufacturing water withdrawal to monthly scale. Previously estimated enterprise-level annual manufacturing water withdrawal was then disaggregated into monthly values by multiplying it with the corresponding monthly revenue fractions. Finally, the monthly water withdrawal of all enterprises was aggregated to 0.1° grid cells to produce gridded, monthly manufacturing water withdrawal data.

Data Records

The High-resolution Sectoral Water Use Dataset (HSWUD) is publicly available and can be accessed at https://doi.org/10.6084/m9.figshare.2761052448. This dataset provides monthly water withdrawal data for China from January 1965 to December 2022 at a spatial resolution of 0.1°, measured in units of 108 m3 per month. It encompasses four key sectors: irrigation, thermal power cooling, manufacturing, and domestic water withdrawal. The data are stored in NetCDF format, with one file dedicated to each sector for ease of use. For example, the file named “HSWUD_dom.nc” contains the domestic water withdrawal data. All files utilize the WGS84 coordinate reference system (EPSG:4326), ensuring compatibility with standard geographical information systems.

Technical Validation

Comparison in volume

The HSWUD dataset was compared and validated against previous gridded sectoral water withdrawal products. Specifically, we compared our estimates with datasets derived from three categories of sources: (1) remote sensing-based products49,50, (2) GHMs output18,51,52,53, and (3) downscaled from statistical data14,15. We derived a median ensemble of six GHMs outputs from ISIMIP3a project database (https://data.isimip.org/, last accessed: 1 April 2025), including CWatM, LPJmL5-7-10-fire, HydroPy, H08, WaterGAP2-2e and MIROC-INTEG-LAND53. These models were selected based on the availability of water withdrawal data under the hostsoc (historical) scenario, in which only these models provided sector-specific outputs. The validation was performed across multiple dimensions, including total volume, interannual trends, and seasonal patterns, for all sectors. During the validation process, the water use dataset covering 341 prefecture-level cities from Zhou9 was considered the ground truth. Although aggregated provincial data from Zhou was used as one of the source datasets in the spatial downscaling process, we did not generate the HSWUD product using prefecture-level city data9. Thus, using Zhou’s prefecture-level data for spatial validation is reasonable.

The results in Fig. 3a,b indicate that the HSWUD dataset outperforms the HUANG and ISIMIP3a datasets in capturing the volumes of total water withdrawal. During the period from 1971 to 2010, the correlation coefficient between the HSWUD and the statistical values was higher (R2 = 0.85), with a lower mean squared error (RMSE = 468 million m3). Similar results were observed in the analysis of results for specific sectors (Fig. 3c–f). When simulating sectoral water withdrawal at the prefecture-level city scale, HSWUD generally exhibited lower error values and narrower error distribution ranges. The superior performance of HSWUD over HUANG and ISIMIP3a likely stems from its use of high-resolution subnational data, allowing better capture of spatial heterogeneity in water withdrawal. In contrast, HUANG and ISIMIP3a rely partly or entirely on GHMs outputs and national statistics, which may lead to larger regional uncertainties14,22. Despite the overall agreement, Hou et al.’s dataset performed slightly better than HSWUD in simulating manufacturing and thermoelectric cooling water withdrawal (manufacturing RMSE: 1.73 for Hou vs. 1.97 for HSWUD; thermoelectric cooling RMSE: 0.21 for Hou vs. 0.52 for HSWUD). This may be attributed to the higher level of detail in Hou et al.’s modeling of industrial water withdrawal, particularly the finer classification of industrial sectors (38 sub-sectors).

Fig. 3
figure 3

Comparison of sectoral water withdrawal between statistics and corresponding values in this study. (a) Scatter plots of average total water withdrawal simulated by this study, Huang et al., and ISIMIP3a with statistical data for 341 prefectural-city from 1971–2010. The corresponding dotted lines of the same color represent the accuracy fitting lines of the three products against the statistical data. The grey solid line indicates the 1:1 accuracy fitting line, serving as a comparison of fitting accuracy. Box plots of prefectural-city RMSE for total water withdrawal from 1971–2010 (b), domestic water withdrawal from 2000–2022 (c), thermal cooling water withdrawal from 2000–2010 (d), irrigation water withdrawal from 2011–2015 (e), manufacture water withdrawal from 2000–2010 (f).

Full size image

Comparison in seasonal variation

Figure 4 illustrates the comparison of seasonal variation among multi-source datasets from 1971 to 2010, expressed as the proportion of monthly water withdrawal relative to annual totals. Most datasets agree that irrigation water withdrawal in China peaks during summer, likely due to high temperatures and transpiration rates in June and July, when rice in southern China enters the heading and grain-filling stages, leading to increased irrigation demand54. The low irrigation withdrawal in April and October in HSWUD originates from MIRCA2000 crop calendars42 and relatively high precipitation in southern China during these months. Our simpler FAO-based approach contrasts with the more complex process-based model used in GHMs like ISIMIP3a and HUANG, explaining their divergence.

Fig. 4
figure 4

Comparison of monthly variation in sectoral water withdrawal generated by this study and previous studies. (a) Line plots of multi-year average (1971–2010) irrigation water withdrawal seasonal variation for China mainland, domestic water withdrawal (b), thermal cooling water withdrawal (c), manufacture water withdrawal (d).

Full size image

In addition, most datasets also show relatively consistent results in describing the summer peaks of thermoelectric cooling water and domestic water withdrawal. This is consistent with the results of previous articles on modeling factors affecting domestic water withdrawal: high evaporation and high temperatures in summer may also lead to increased domestic water demand55. On the other hand, in addition to the high demand for cooling in summer, which leads to high use of thermoelectric cooling water, the need for centralized heating in winter in northern China also drives the thermoelectric cooling water use high from November to January of the following year56, with an average annual share of 24.0% (Huang)-25.0% (Hou) in all datasets. For manufacturing water withdrawal, our dataset shows a similar intra-annual trend to that of Hou et al., both exhibiting a general increase from the beginning to the end of the year. One key difference is that Hou’s dataset shows a distinct peak in June, which is not observed in our results. This discrepancy may be attributed to differences in the choice of temporal downscaling covariates: Hou et al. used industrial product output data, while we employed main sales revenue as the proxy.

Comparison in annual trends

In terms of annual trends, our results show a consistent increase in domestic water withdrawal, aligning with most existing datasets (Fig. 5). In contrast, manufacturing and thermoelectric cooling water withdrawal first increased and then slightly declined, a pattern that corresponds with previous studies evaluating the effectiveness of China’s water resource management policies57 and further supports the success of national water-saving initiatives. The decline in industrial water withdrawal around 2010 reflects the impacts of national water-saving policies, which are not accounted for in HUANG and GHM-based datasets, explaining their divergent trends. Moreover, the unrealistic continued increase in manufacturing water withdrawal projected by ISIMIP3a may stem from its reliance on global macroeconomic assumptions, which fail to accurately capture local dynamics58.

Fig. 5
figure 5

Comparison of sectoral water withdrawal annual trend generated by this study and previous studies. (a) Line plots of temporal evolution of annual irrigation water withdrawal for China mainland (2000–2022), domestic water withdrawal (b), manufacture water withdrawal (c), thermal cooling water withdrawal (d). Note that all values were normalized by their respective values in the year 2000. The comparison was limited to the past 20 years to account for the confidence levels of each dataset in capturing long-term water withdrawal trends.

Full size image

Over the past two decades, our dataset also captures a slight upward trend in irrigation water withdrawal from 2000 to 2022, consistent with earlier studies on irrigation trends in China59. This may reflect a balance between the expansion of water-saving technologies and changes in cropping structure32,59. Overall, above validation suggests that the HSWUD data can reasonably reflect the patterns of water withdrawal in China.

Spatial distribution of water withdrawal in China

From 1965 to 2022, China’s total water withdrawal exhibited significant spatial variation (Fig. 6). The East China region and the Pearl River Delta had the highest total water withdrawal, followed by the industrial and irrigated zones of the North China Plain. These findings are in accordance with those previously reported by Zhang et al., which constructed a city-level sectoral withdrawal datasets for China in 201560.The spatial pattern of water withdrawal is closely related to differences across sectors (Fig. 7). Irrigation water withdrawal was notably high in the middle and lower reaches of the Yangtze River, the Yellow River basin, and the oasis irrigated regions of Xinjiang. Due to the dense concerntration of industrial enterprises, manufacturing water withdrawal is relatively high in the Pearl River Delta, Yangtze River Delta, Chengdu Plain, and the cities of Beijing and Tianjin, accounting for 32.64% of the national manufacturing water withdrawal (1.45 billion m3). The total cooling water for thermal power is relatively low (5.50%) and spatially dispersed, with high-value areas concentrated along the Yangtze River in Jiangsu Province, while low-value areas are mainly located in the western provinces, closely related to local energy structures.

Fig. 6
figure 6

Spatial distribution of monthly total water withdrawal in China from 1965 to 2022. Pixels colored towards red indicate higher total water withdrawal, while those colored towards yellow indicate lower total water withdrawal. Areas with zero total water withdrawal are represented in grey.

Full size image
Fig. 7
figure 7

Spatial distribution of monthly sectoral water withdrawal in China from 1965 to 2022. (a) Irrigation water withdrawal (b) Cooling water for thermal power generation (c) Manufacturing water withdrawal (d) Domestic water withdrawal. Pixels colored towards red indicate higher total water withdrawal, while those colored towards yellow indicate lower total water use. Areas with zero total water withdrawal are represented in grey.

Full size image

Overall, these findings are in accordance with those previously study, the irrigation sector remains the largest consumer of water32, accounting for 71.35% of total water withdrawal, followed by manufacturing (12.03%), domestic water (11.12%), and cooling water for thermal power generation.

Long term changes in sectoral water withdrawal

In terms of long-term sectoral water withdrawal trends, domestic water withdrawal shows a clear and continuous increase, while other sectors exhibit varying degrees of fluctuation between 1965 and 2022 (Fig. 8). Domestic water withdrawal rose significantly from 15.7 billion m³ per month in 1965 to 77.7 billion m³ per month in 2022. Irrigation water withdrawal increased from 188.1 billion m³ per month in 1965 to a peak of 298.4 billion m³ per month in 1992, followed by relatively minor fluctuations over the past two decades, reaching 300.4 billion m³ per month in 2022. Thermoelectric cooling and manufacturing water withdrawal both showed upward trends from around 3 and 25 billion m³ per month in 1965, peaking at 65 and 30 billion m³ per month in 2008 and 2011, respectively, before entering a period of steady decline. These trends are consistent with the findings of Zhou et al., while our dataset further enhances temporal resolution by providing monthly-scale details9.

Fig. 8
figure 8

Time series of sectoral water withdrawal in China generated by HSWUD.

Full size image

Uncertainties

The spatial distribution of manufacturing water withdrawal in the HSWUD dataset is primarily derived from micro-level data of over 310000 industrial enterprises in 2007, combined with the water use efficiency of their corresponding manufacturing sub-sectors. One potential limitation lies in the spatial sampling of industrial enterprises, as the dataset includes only enterprises above a designated scale (annual output >5 million CNY), thus excluding smaller enterprises. While this may affect the comprehensiveness of spatial representation, data from the 2008 China Economic Census Yearbook indicate that these larger enterprises account for 93% of total industrial output and 85% of industrial water withdrawal, suggesting that the exclusion of smaller enterprises has a limited impact on the overall spatial pattern. Furthermore, spatial aggregation to 0.1° grid cells helps mitigate potential biases introduced by sampling limitations.

Another source of uncertainty is the assumption of uniform water use efficiency within each sub-sector at the provincial level, which may overlook intra-sectoral variability due to differences in technology and production processes61. However, empirical studies have shown that enterprises within the same industrial sub-sector tend to cluster geographically due to input-output linkages, labor market pooling, and knowledge spillovers62. In addition, research on China’s industrial water use efficiency has demonstrated strong spatial autocorrelation, indicating that regions with similar efficiency levels tend to be geographically concentrated63. These findings support the use of sub-sector-level efficiency to approximate spatial patterns of manufacturing water withdrawal. In future work, we plan to further improve our approach by incorporating more detailed, enterprise-specific water use efficiency data as they become publicly available.

Spatial downscaling of thermal power cooling water remains challenged by insufficient spatial representativeness of thermal power plants due to incomplete data. The precise locations of all thermal power plants in China are difficult to determine, with instances of plant closures, reconstructions, and new constructions, leading to uncertainties in both the number and location of power plants in the 2017 data from GPPD. This study identified 1123 thermal power plants, accounting for 75.3% of the national electricity generation in that year. Furthermore, we utilized provincial-level thermal power water withdrawal data (rather than the coarser spatial resolution of national-level data) and allocated it to a 0.1° grid, partially mitigating the uncertainties associated with spatial location.

Due to limitations in historical covariate data, we adopted a single reference year for spatial downscaling, which may not fully capture the spatiotemporal evolution of irrigated area and population density from 1965 to 2022, introducing some uncertainty over long time spans. However, aggregating Zhou’s prefecture-level data to the provincial scale preserved some degree of spatiotemporal variability was retained in the provincial water use estimates. Land use change in China was relatively limited before 2000 but accelerated thereafter due to ecological policies and rapid urbanization39,40. To account for this, we updated irrigated area data every five years after 2000 to balance consistency and temporal dynamics.

Meanwhile, the spatial distribution of population in China has remained largely stable over recent decades. Population density patterns in 1990, 2000, and 2020 are highly correlated (R2 > 0.92), indicating minimal shifts in the overall spatial structure (Fig. 9). Given the strong correlation between population distribution and water use64, we believe that intra-provincial water use patterns have likely remained relatively consistent over time.

Fig. 9
figure 9

Population density map in 1990(a), 2000(b), 2010(c), 2020(d).

Full size image

The spatial distribution of irrigation water is influenced by factors such as irrigation methods, cropping patterns, and cropping intensity. While omitting these variables introduces some uncertainty, incorporating them into a national, long-term dataset is challenging due to limited data availability and consistency. Moreover, the use of sparse or poorly validated sources may introduce further uncertainties65. As a practical alternative, many studies use irrigated area as the primary proxy for spatial allocation of irrigation water, balancing feasibility and spatial representativeness12,34.

To validate this approach, we analyzed city-level data from 2000 to 2020. For each province (excluding Beijing, Tianjin, Shanghai, Chongqing, and Tibet due to limited data), we calculated the shares of irrigated area and irrigation water withdrawal at the prefecture level and assessed their correlation. As shown in Fig. 10, panels (a)–(c) illustrate the average spatial distribution, while panels (d)–(f) show the temporal evolution of correlation coefficients. The results reveal a consistently strong positive relationship, with over 61% of provinces showing R² above 0.8, supporting the use of irrigated area as a reasonable proxy for sub-provincial irrigation water distribution.

Fig. 10
figure 10

Correlation analysis between the share of irrigation water withdrawal and the share of irrigated area across prefecture-level cities within each province. Panels (a–c) show the multi-year averages for eastern, central, and western provinces from 2000 to 2020, respectively. Panels (d–f) present the time series of correlation coefficients for the same regions over the same period.

Full size image

When allocating annual water withdrawal to the monthly scale, we adopted a simplified scheme: seasonal variations in irrigation, thermal power cooling, and domestic water withdrawal were estimated using monthly meteorological indicators, while manufacturing water withdrawal was based on monthly changes in sales revenue. This method does not account for the influence of seasonal water management strategies on monthly irrigation variation66. Additionally, due to the lack of observed monthly domestic water demand data in China, we selected a representative seasonal amplitude parameter (R) based on existing literature14. These parameters can be refined in future studies as high-resolution observational data become available.

We selected ERA5-Land reanalysis data for temporal downscaling based on its long temporal coverage and relatively high spatial resolution. While previous studies have demonstrated that ERA5-Land reasonably captures monthly variations in potential evapotranspiration67 and precipitation68 across China, it remains a global-scale reanalysis product and may introduce some uncertainties into our dataset. We advise users to take this limitation into account when applying the dataset in practice.