High-resolution gridded dataset of sectoral water pollution discharges in China from 2007 to 2022

Abstract

High-resolution datasets of anthropogenic water pollution discharges are essential for characterizing pollution dynamics and informing water quality management. However, China’s pollution source data remain limited to provincial scales and decadal censuses, constraining spatiotemporal analyses and policy evaluation. We present a High-resolution Sectoral Water Pollution Discharge Dataset for mainland China (2007–2022), providing annual data at 30 arc-second (approximately 1 km at the equator) resolution. By integrating pollution source statistics with geospatial data through a top-down downscaling framework, we allocated provincial discharges to grid cells. The dataset quantifies gridded anthropogenic discharge measured by chemical oxygen demand (COD) and ammonium nitrogen (NH₃-N) from five sectors: urban residential, rural residential, industrial, crop farming, and livestock farming. Validation was performed by comparing city-level aggregated estimates against official census records from 73 cities, demonstrating strong agreement (R² > 0.6) for both pollutants across all sectors. This dataset enables identification of fine-scale pollution hotspots within river basins that were previously obscured by provincial-scale data, thereby supporting the implementation of targeted pollution control strategies.

Background & Summary

Anthropogenic water pollution discharges significantly influence biogeochemical cycles and human health on a global scale^1,2. China, experiencing rapid urbanization and industrialization alongside intensive agricultural activities, contributed over 30% of global nutrient discharges during the first decade of the 21^st century^3–7. Although the Chinese government has prioritized pollution control through stringent measures that have yielded preliminary success over the past two decades^8–11, including an approximately 15% decline in anthropogenic discharge between 2010 and 2020^12,13, river water quality deterioration and lake eutrophication driven by anthropogenic pressure remain considerable challenges^14–19. Therefore, assessing long-term water pollution source patterns in China can effectively enhance understanding of spatial heterogeneity and temporal dynamics of anthropogenic disturbances^5,12,20,21, help evaluate the effectiveness of implemented policies, and accurately identify priority regions for future control measures.

China’s National Pollution Source Census provides crucial baseline data for analyzing pollution from various sources (Fig. 1), but the resource-intensive processes of data collection, calibration, and compilation require substantial labor and financial investment. Consequently, this dataset has significant limitations, with restricted temporal coverage (only 2007 and 2017 data are available) and primarily provincial-level resolution, hindering the capture of spatiotemporal variations in water pollution discharges driven by diverse anthropogenic activities. This inadequacy fails to meet the requirements of global-scale models (e.g., Global NEWS-2²²) and national-scale models (e.g., MARINA 1.0²³), which operate at grid or watershed scales rather than administrative scales. As a result, many attribution studies of water pollution levels still rely on proxy data such as nighttime lights, GDP, fertilizer application, livestock farm counts, and wastewater discharge volumes to represent overall anthropogenic disturbances within specific sectors^8,24–27. However, these proxies cannot directly capture the actual pollution loads. Therefore, the coarse spatial and temporal resolution of available data is insufficient for a comprehensive assessment of pollution-induced environmental impacts, necessitating the development of refined anthropogenic discharge datasets with enhanced spatiotemporal resolution.

Fig. 1. Provincial-level changes in anthropogenic pollution discharges across five sectors between the 2007 and 2017 National Pollution Source Censuses.

Maps show the spatial distributions of relative reduction for (a) COD and (b) NH₃-N from different sectors. Bar plots indicate total discharges and sectoral contributions for each pollutant in both census years.

Although previous studies have estimated nutrient discharges based on census data, they primarily employed bottom-up coefficient-based calculations^12,28–38. These approaches typically yield results at administrative or coarse grid scales¹², creating a mismatch with environmental data (e.g., meteorological and water quality metrics) scales³⁹, and inadequately supporting precise watershed-scale research^22,23,40. The few studies addressing fine-scale pollution sources are constrained by extensive spatiotemporal gaps in required source data, limiting outputs to specific years¹³ and regions^30–35 or covering only select pollution sources^36–38 (e.g., crop farming or residential sources), thereby restricting a comprehensive understanding of China’s anthropogenic pollutant discharge patterns. Another commonly used gridding approach is the downscaling method, which has been widely adopted for sectoral water use datasets^41–43 and can effectively combine survey data accuracy with high spatiotemporal resolution gridded socioeconomic variables. However, its application to pollution source gridding remains limited, with existing studies providing insufficient resolution and timeliness for current research needs⁹.

To overcome these limitations and provide spatially explicit support for China’s water pollution source control efforts, this study presents a top-down approach for tracking long-term anthropogenic pollutant discharge. To ensure consistency with data from the censuses and environmental yearbooks spanning these sixteen years, while making our dataset comparable to existing research^9,12,13, we select COD and NH₃-N as indicators. COD (measured using the potassium dichromate method) quantifies the oxygen demand imposed by organic pollutants and thus serves as a standard indicator of organic pollution load. NH₃-N poses direct toxicity to aquatic organisms and drives eutrophication in freshwater systems. Together, these two pollutants represent the dominant dimensions of China’s water pollution challenges: organic contamination and nutrient overloading. By utilizing a spatiotemporal dynamic parameter system comprising remote sensing-derived land use data, nighttime light data, Point of Interest (POI) data, and socio-economic statistical data, we develop a downscaling algorithm to construct a high spatial resolution (30 arc-second) dataset for water pollution sources in China. We quantify gridded estimates from five major sources: urban residential, rural residential, industrial, crop farming, and livestock farming, covering both indicators.

This high-resolution annual dataset quantifies the total amount of pollutants from various anthropogenic sources prior to transport processes, representing the potential pollution pressure on surface water bodies. Therefore, it will enhance the reliability of water pollution studies from multiple perspectives. First, it reveals spatial distributions and temporal characteristics of different pollutants and anthropogenic sources, enabling decision-makers to identify hotspot areas and implement targeted pollution control measures. Second, it serves as input for hydrological models simulating pollutant transport and deposition in water bodies, supporting environmental impact and human health risk assessments. Additionally, when combined with meteorological and other global datasets, it facilitates factor analysis of water quality changes, supporting discharge reduction target setting and future water quality predictions.

Methods

The dataset was constructed using a top-down algorithm following the schematic in Fig. 2. We used China’s anthropogenic pollutant discharge inventory from the decadal pollution source censuses, which provides provincial and sectoral discharge data, as the baseline. Annual anthropogenic pollution discharge statistics from yearbooks (2007–2022) were then temporally adjusted to align with census data magnitudes. Finally, provincial-level anthropogenic pollution discharges were spatially distributed to 30 arc-second grids using socioeconomic variables to produce sectoral distribution maps.

Fig. 2.

Schematic framework of the top-down downscaling methodology.

Data collection

Data used in this study comprise three main categories: pollution source statistics, multi-source land-use data, and ancillary socioeconomic data, as detailed in Tables 1 and 2.

Table 1.

Sources of data for identifying the pollution source grids for each sector.

Land-use type	Data name	Data source	Spatial resolution	Temporal range	Sector	Interpolation method
Industrial area	Non-residential built-up surface maps	GHSL: Global built-up surface 1975-2030 (P2023A)66 (https://data.jrc.ec.europa.eu/dataset/9f06f36f-4b11-47ec-abb0-4f8b7b1d72ea)	100 m * 100 m	2005, 2010, 2015, and 2020	Industrial sector	For years before 2020, linear interpolation was applied. For years 2021 and 2022, exponential extrapolation was applied.
Urban area	Human urban settlement maps	Global urban and rural settlement (GURS) dataset from 2000 to 202055 (https://zenodo.org/records/11160893)		2000, 2005, 2010, 2015, and 2020	Urban residential sector
	National land use maps (urban category)	China’s National Land Use and Cover Change (CNLUCC) dataset54 (https://www.resdc.cn/DOI/DOI.aspx?DOIID = 54)
Rural area	Human rural settlement maps	GURS dataset			Rural residential sector Livestock farming.
	National land use maps (rural category)	CNLUCC dataset
Cropland area	Cropland extent maps	Extended Data on China’s 30-m Annual Cropland Dataset for 1990–2023 (CACD-v1)63 (https://zenodo.org/records/16927779)	30 m * 30 m	2007–2022	Crop farming Livestock farming

Table 2.

Sources of proxy variables to generate the weight factors for each sector.

Sector	Data name	Data source	Spatio-temporal scale	Description of data usage
Residential	Population census data	National Population Census of China75	County-level, from 2000 to 2020 in 10 year intervals	For urban population, distribute the urban population census data to grids by nighttime light intensity and urban area. For rural population, distribute the rural population census data to grids by hamlet POI density and rural area.
	Nighttime light data	An extended time series of global NPP-VIIRS-like nighttime light data58	Gridded data at 15 arc second resolution, annual
	Hamlet density	Gaode Map API service62	Point data, from 2012 to 2020 in 2 year intervals
	Wastewater treatment ratio and pollutant removal rate	China Urban-Rural Construction Statistical Yearbook47, Dataset of the First and Second Chinese Pollution Source Census44,45	Provincial, annual	Convert the gridded population data to wastewater discharge equivalent as weights.
	Daily residential water consumption (per capita)		City-level, annual
Industrial	Industrial GDP	Global Sectoral GDP map at 30” resolution (SectGDP30) v2.076	Gridded data at 30 arc second resolution, 2010, 2015 and 2020	Combine gridded industrial GDP and area data to generate weights.
Livestock farming	Livestock distribution maps	Annual global gridded livestock mapping77	Gridded data at 5 km resolution, annual	Convert all livestock to swine equivalent, then distribute it to grids by rural and cropland area as weights.
Crop farming	Cropping intensity	ACIA500: a 500 m annual cropping intensity dataset for monsoon Asia based on MODIS data64	Gridded data at 500 m resolution, annual	Generate the harvest area from original cropland area.
	Nitrogen fertilizer application	China Statistical Yearbook and Provincial Statistical Yearbooks65	Provincial, annual	Distribute the fertilizer data to grids by harvest area as weights.

Adjusting China’s yearly anthropogenic discharge statistics

To address substantial inconsistencies between census data^44,45 and yearbook⁴⁶ statistics and obtain more accurate annual time series of sectoral anthropogenic discharges, we implemented two adjustment steps:

(1) Estimation for the rural residential sector: Since no data sources other than the Second National Pollution Source Census provided provincial-level discharge data for the rural residential sector, we estimated missing rural residential discharge (${\mathit{RD}}_{\mathit{year},i}$) for province i and year from 2007 to 2022 excluding 2017 by establishing relationships among rural population (${\mathit{RPOP}}_{\mathit{year},i}$), daily per capita water consumption (${\mathit{WC}}_{\mathit{year},i}$), wastewater treatment ratios ($T{R}_{\mathit{year},i}$) from statistical yearbook⁴⁷ and pollutant concentration in rural wastewater (${\mathit{PC}}_{p,i}$), pollutant removal efficiencies ($R{E}_{p,i}$) for pollutant p provided by the Second Census, as shown in Eq. (1):

$${\mathit{RD}}_{\mathit{year},i}={\mathit{RPOP}}_{\mathit{year},i}\times {\mathit{WC}}_{\mathit{year},i}\times 365\times {\mathit{PC}}_{p,i}\times (1-T{R}_{\mathit{year},i}\times R{E}_{p,i})$$ 1

This approach is grounded in the emission coefficient framework used in the Second Census, and it avoids the oversimplification of assuming constant per capita discharge rates across all provinces.

(2) Annual statistics adjustment for other sectors: We adjusted anthropogenic discharges recorded in China’s annual Environmental Statistics Yearbooks (2007–2022) based on both census results. The National Pollution Source Censuses, conducted every ten years, investigate nearly 6 million pollution sources through field surveys and direct measurements. They provide the most accurate estimates of sectoral anthropogenic discharges at the provincial level. In non-census years, the Ministry of Ecology and Environment conducts sampling surveys of all pollution sources except the rural residential sector, with results published in annual Environmental Statistics Yearbooks. Methodological differences between comprehensive censuses (near-complete enumeration) and sampling surveys (targeted subsets) result in magnitude discrepancies. However, yearbook statistics effectively capture inter-annual variations in pollution trends. A comparison between the original statistics in the 2016–2019 yearbooks and the republished versions based on Second Census results shows that annual reduction rates remained consistent while absolute values differed. Here, we present an adjustment framework as follows, which leverages the accuracy of census data for absolute magnitudes while preserving the temporal dynamics captured by yearbook statistics.

First, the adjustment of national annual total anthropogenic discharge between censuses employs a proportional benchmarking approach⁴⁸ commonly used by statistical agencies (e.g., United Nations Population Division, U.S. Census Bureau) to calibrate survey data to census benchmarks. We calculated the relative reduction rate between adjacent years from yearbook statistics, then normalized these rates to derive adjustment weights that sum to unity over the inter-census period, following Eqs. (2–4):

$$r_{\mathit{year}}=\ln {\mathit{TYS}}_{\mathit{year}-1}-\ln {\mathit{TYS}}_{\mathit{year}}$$ 2

$$w_{\mathit{year}}=\frac{r_{\mathit{year}}}{\sum _{k=2008}^{2017}{r}_k}$$ 3

$${TD}_{year}={TD}_{2007}-({TD}_{2007}-{TD}_{2017})\times \sum _{k=2008}^{year}{w}_k$$ 4

where ${\mathit{TYS}}_{\mathit{year}}$ and ${\mathit{TD}}_{\mathit{year}}$ represent yearbook statistics and adjusted discharge totals for year from 2007 to 2017, respectively; $r_{\mathit{year}}$ denotes the relative reduction rate between adjacent years, and $w_{\mathit{year}}$ represents the reduction rate weight. This approach redistributes the absolute difference between census results proportionally according to the temporal variation pattern in yearbook data, thereby preserving inter-annual dynamics while ensuring consistency with census benchmarks. For post-2017 years, we applied yearbook change rates to Second Census data through sequential annual extrapolation. Following national-level adjustment, we disaggregated the adjusted sectoral totals to the provincial scale by applying the proportional contributions of each province and sector from the original yearbook data. This approach ensures that provincial and sectoral distributions preserve the spatial patterns documented in annual statistics. The resulting province-sector-year-specific discharge estimates serve as inputs for the spatial downscaling procedure described in the next section. Figure 3 shows the resulting year-over-year change rates for adjusted sectoral anthropogenic discharges.

Fig. 3. National decrease rates of sectoral anthropogenic pollution discharges directly obtained from annual statistical yearbooks.

Numeric values indicate annual reduction rate, and color intensity denotes the magnitude of rate within each sector.

Spatial allocation of provincial anthropogenic discharges

To downscale anthropogenic pollution discharges from provincial administrative regions and sectors across China, we first obtained the minimum bounding rectangle for mainland China and divided it into 7,356 × 4,249 regular grid cells at 30 arc-second resolution (approximately 1 km at the equator) in WGS84 coordinate system as our basic mapping units.

The following describes the identification of the pollution source grids from land-use data (Table 1) and the selection (detailed in Supplementary Text 1) and processing of proxy variables (Table 2) to generate spatial downscaling weights for each sector:

Residential sector. Over the past 20 years, substantial expansion and upgrading of urban and rural wastewater infrastructure have significantly reduced anthropogenic discharges from the residential sector^47,49–53. Most pollution from the urban residential sector is discharged through wastewater treatment plants (WWTPs). However, for the rural residential sector, the proportion of untreated wastewater remains considerably higher than in urban areas, particularly in areas lacking sanitary facilities, where human waste is discharged directly into nearby water bodies through surface runoff. Given the close relationship between domestic wastewater discharges and residential pollution sources, we used wastewater equivalents generated from a combination of population and correction factors related to wastewater treatment status as the allocation factor, implementing weight factor preparation through three steps:

(1) Settlement boundary generation: We generated basic regions of human settlements using China’s National Land Use and Cover Change (CNLUCC) dataset⁵⁴ and the Global Urban and Rural Settlement⁵⁵ (GURS) dataset, as both datasets distinguish between urban and rural areas that are typically classified only as impervious surfaces in most high-resolution land cover products. Urban and rural classifications were extracted from both datasets at 100 m resolution and aggregated to match our mapping units. Since both datasets are provided at 5-year intervals from 2000 to 2020, settlement areas for intermediate years within each interval were estimated using linear interpolation. For post-2020 values, we applied exponential extrapolation using the 2015-2020 average relative change rate to maintain recent trends while avoiding zero values as indicated in Eqs. (5–6):

$$r=\frac{1}{5}\times (\mathrm{l}\mathrm{n}{y}_{2020}-\mathrm{l}\mathrm{n}{y}_{2015})$$ 5

$$y_t=y_{2020}\times e^{r(t-2020)},t=2021,2022$$ 6

where r represents the 2015-2020 average annual relative change rate and $y_t$ is the extrapolated estimate for post-2020 years. After obtaining continuous time series of urban and rural settlement areas from both data sources, we created composite maps by selecting the maximum settlement area value within each grid cell from the two datasets as the final settlement area.

(2) Population gridding: Common gridded population datasets lack urban-rural population differentiation⁵⁶, inadequately supporting our requirements. Following established population downscaling approaches⁵⁷, we used nighttime light data as weighting factors to allocate urban population data within target cells containing urban settlements. For data continuity, we used a 20-year nighttime light dataset with extended coverage by integrating DMSP-OLS and NPP-VIIRS data⁵⁸. Nighttime light data were averaged and aggregated to mapping unit grids. We implemented adjustments when converting raw nighttime light data to weighting factors: we extracted the 99th percentile of brightness values in 2020 from residential community POI-located grids in China’s three megacities (Beijing, Shanghai, and Shenzhen) as a threshold to correct anomalously high values⁵⁹. Recognizing that the highest brightness areas represent commercial rather than residential zones, we applied logarithmic smoothing^60,61 to reduce extreme value influence on weights using the following Eq. (7):

$${NTL}_j^{\mathrm{\prime }}=\{\begin{array}{c}\mathrm{l}\mathrm{n}{NTL}_j/\mathrm{l}\mathrm{n}NT{L}_{threshold}\times NT{L}_{threshold},{NTL}_j>NT{L}_{threshold}\\ {NTL}_j,{NTL}_j\le NT{L}_{threshold}\end{array}$$ 7

where ${\mathit{NTL}}_j$ is the original nighttime light value for grid cell j, ${\mathit{NTL}}_j^{\prime }$ is the corrected value, and $\mathit{NT}{L}_{\mathit{threshold}}$ is the threshold value.

We then conducted a calibration experiment to identify the optimal combination of nighttime light intensity and urban settlement area as weighting factors for urban population downscaling. Different coefficient combinations were tested by downscaling city-level population data to grid cells, then aggregating the gridded estimates to the county level and validating against county-level census population data (detailed in Supplementary Text 2). The optimal weighting scheme achieved high validation accuracy (R² > 0.79 across all years) and was applied to downscale county-level population data to urban grids across all years.

In rural areas, where numerous grids lack corresponding light intensity values, we downscaled population data using the density of hamlet POIs⁶² and rural settlement area as weighting factors. Calibration similar to that for urban population downscaling was performed (detailed in Supplementary Text 2). The optimal combination of these two variables which achieved consistently high validation accuracy (R² > 0.88 across all years) was applied to downscale county-level rural population data to grids across all years.

Based on the gridded distributions of population in 2000, 2010, and 2020 (Fig. 4a,b), we interpolated and extrapolated data for the remaining years to obtain complete urban and rural population data using the same method described previously for generating land-use data.

Fig. 4.

Spatial distribution of major proxy variables used to generate weight factors for each sector in 2020: (a) urban population, (b) rural population, (c) harvested area, (d) industrial GDP, and (e) livestock distribution in pig units. Spatial distributions of major proxy variables used to generate weight factors for each sector in 2020. The panels display the gridded patterns for (a) urban population, (b) rural population, (c) harvested area, (d) industrial GDP, and (e) livestock distribution.

(3) Conversion to wastewater equivalents: To explicitly account for persistent urban-rural disparities in treatment coverage, we introduced correction factors to convert raw population data to residential wastewater equivalents separately for urban and rural areas. For urban areas, we obtained annual city-level data on per capita water consumption, wastewater treatment ratios, as well as national data on pollutant-specific removal efficiencies. For rural areas, we collected annual provincial data on per capita water consumption and treatment rates, and obtained removal efficiencies from the Second Census dataset. We then applied Eq. (8) to calculate the wastewater equivalent as the residential weight factor ${\omega }_{i,j}$ for grid cell j of city/province i:

$${\omega }_{i,j}={\mathit{POP}}_{j,i}\times {\mathit{WC}}_i\times (1-T{R}_i\times R{E}_{p,i})$$ 8

where ${\mathit{POP}}_{j,i}$ is the gridded population, ${\mathit{WC}}_i$ is the daily per capita water consumption, $T{R}_i$ is the wastewater treatment ratio, $R{E}_{p,i}$ is pollutant-specific removal efficiency, and p denotes the pollutant type.

Crop farming sector. As a typical non-point source pollution type, crop farming water pollution primarily originates from nutrient losses from agricultural fields, which are closely tied to fertilizer application. Therefore, pollution discharges in this category were spatialized using nitrogen fertilizer application data. We extracted farmland area based on high-resolution cropland extent map series provided by the CACD dataset⁶³ to better reflect fine-scale temporal changes in cultivation regions. We then introduced a cropping intensity dataset⁶⁴ to reflect the planting intensity of different agricultural regions. Cropping intensity quantifies the number of crop cycles per year (e.g., single cropping vs. double or triple cropping), thereby capturing the temporal intensification of agricultural activities. The final harvested area factor was calculated by multiplying the cropland extent by the corresponding grid-level cropping intensity:

$$H{A}_j=C{A}_j\times C{I}_j$$ 9

where $H{A}_j$ is the harvested area, $C{A}_j$ is the cropland area, and $C{I}_j$ is the cropping intensity for grid cell j. When grid-level cropping intensity data was unavailable, we used the provincial average for that year as a substitute.

The harvested area (Fig. 4c) was then multiplied by regional nitrogen fertilizer application intensity⁶⁵ to obtain the final crop farming pollution allocation weighting factor:

$${\mathit{NF}}_{i,j}=H{A}_{i,j}\times N{F}_i/{\sum }_{j=1}^{n}H{A}_{i,j}$$ 10

where ${\mathit{NF}}_{i,j}$ represents the nitrogen fertilizer application for grid cell j in province i, $N{F}_i$ is the provincial nitrogen fertilizer application, and n is the total number of grid cells within that province.

Industrial sector. Almost all industrial pollution is discharged directly through industrial WWTPs to surface water. Given that obtaining the locations of industrial enterprises and enterprise-level discharge data in non-census years is extremely difficult, we developed a pollution-equivalent approach combining multiple proxy variables to generate allocation weight factors for the industrial sector.

First, we extracted industrial area distributions from the Global Human Settlement Layer (GHSL)⁶⁶, which provides non-residential (NRES) land-use data at five-year intervals from 2005 to 2020. To validate GHSL’s temporal representativeness, we compared it against multi-year industrial enterprise POI data obtained from the Gaode Map API service. Results show that GHSL NRES areas consistently capture over 70% of industrial enterprise locations across all validation years (Supplementary Table 4), demonstrating that GHSL effectively tracks changes in spatial patterns of industrial enterprises. Data for the remaining years were interpolated and extrapolated using the same method applied in settlement boundary generation.

Second, following previous studies^12,43, we introduced industrial GDP (Fig. 4d) as a key allocation factor. Correlation analysis revealed that industrial GDP demonstrated strong and increasingly significant correlations with industrial discharge at the provincial scale over time, with coefficients rising from approximately 0.5 to above 0.7 (Supplementary Figure 1). To address the nonlinear relationships between industrial GDP and discharge, we tested multiple transformation forms (log-scale, power-scale, and root-scale) for both industrial GDP and industrial area. For each year from 2007 to 2022, we fitted zero-intercept linear regression models using data from all provinces, selecting the optimal transformation combinations that achieved the highest average R² values across all years (detailed in Supplementary Text 3). These year-specific models were then used to convert gridded industrial GDP and industrial area into industrial pollution equivalents, which serve as weighting factors for spatial allocation.

Livestock farming sector. Discharge from livestock farming, a significant agricultural non-point source pollution contributor^67–69, originates from animal waste generated at large-scale concentrated farms and rural household scattered farming operations. To address the spatial distribution of livestock pollution, we utilized long-term annual gridded livestock mapping data covering eight species as the source data for generating weights. The original 5 km resolution gridded livestock distribution data were first matched with our mapping units, we then used rural settlement and cropland area as weighting factors to downscale the distribution of each species to our approximately 1 km resolution grids. By applying pig unit conversion factors for these species^69,70, the final distribution of livestock weight factors was measured in standardized pig units (Fig. 4e).

Finally, annual sectoral anthropogenic discharges were gridded using a top-down approach. The converted socio-economic pixel values were normalized by respective provincial totals to derive spatial downscaling weights ${\omega }_{i,j}$:

$${\omega }_{i,j}=\frac{x_{i,j}}{\sum _{j=1}^n{x}_{i,j}}$$ 11

where $x_{i,j}$ is the pixel value of a socio-economic variable in grid cell j within province i, and n is the total number of pixels within that province.

Through raster calculations between spatial downscaling weight factors and provincial panel pollution source data, provincial annual pollution discharges were allocated to grid cells, producing 30 arc-second sectoral anthropogenic discharge spatial distribution maps:

$${{\mathit{PD}}_{i,j}=\omega }_{i,j}\times P{S}_i$$ 12

where ${\omega }_{i,j}$ is the spatial downscaling weight factor for grid cell j in province i, $P{S}_i$ is the sectoral pollution source data for province i and $P{D}_{i,j}$ is the sectoral pollution discharge for grid cell j in province i.

Data Records

The High-resolution Sectoral Water Pollution Source Dataset⁷¹ is publicly available on Zenodo (https://zenodo.org/records/16930435). This dataset provides annual water pollution discharge data for China from 2007 to 2022 at 30 arc-second spatial resolution, covering two typical water pollutants (COD and NH₃-N) across five key sectors: crop farming, livestock farming, industrial, urban residential, and rural residential.

Data are stored in GeoTIFF format with the following specifications: WGS84 coordinate reference system (EPSG:4326), 32-bit floating-point (float32) data type, LZW (Lempel-Ziv-Welch) lossless compression, values measured in kilograms per year, and no-data values set to −99.0. The dataset contains one file per sector per year per pollutant. For example, “chn_1km_pol_Urban_NHN_2007.tif” contains urban residential NH₃-N discharge for 2007.

Figures 5 and 6 show the spatial distributions of gridded anthropogenic discharges measured by COD and NH₃-N from different sectors across China in 2007, 2015, and 2022, respectively. These maps illustrate the spatiotemporal patterns captured in our dataset, demonstrating shifts in pollution source distribution over time. The density plots on the right show the changes in the distribution of pixel values for corresponding sectors across these three years, revealing temporal trends in anthropogenic discharge intensity.

Fig. 5. Spatial distributions of gridded anthropogenic discharges measured by COD from different sectors across China in 2007, 2015 and 2022.

The density plots on the right show the changes in the distribution of pixel values for corresponding sectors across these three years. Data are displayed at 2.5 arcminute resolution for visualization clarity.

Fig. 6. Spatial distributions of gridded anthropogenic discharges measured by NH₃-N from different sectors across China in 2007, 2015 and 2022.

The density plots on the right show the changes in the distribution of pixel values for corresponding sectors across these three years.

Technical Validation

City-level comparison with census data

We compared mapping results with published city-level census data for 73 cities from the reports of the Second National Pollution Source Census. Spearman correlation coefficients (r) and linear fit results (slope and R²) were calculated by comparing aggregated city-level pollution discharges against census data for each sector (Fig. 7). R² values between total estimates and census data reached 0.64 and 0.84 for COD and NH₃-N, respectively. Notably, crop farming and residential sectors demonstrated strong performance with R² approaching or exceeding 0.7. The industrial sector showed the next best fit after these two sectors, with correlation coefficients reaching 0.8 and R² values of approximately 0.6. Livestock farming sector estimates showed high correlation (~0.7) and generally aligned with the 1:1 line, although a few cities exhibited larger deviations, resulting in relatively lower R² values.

Fig. 7. Comparison of sectoral anthropogenic discharges at city level between census data and corresponding aggregated estimates.

(a) Distribution of cities used for validation. (b) and (c) Results for COD and NH₃-N, respectively. Blue solid lines represent linear regression fits between estimated and census discharges for each sector and red dashed line indicates the 1:1 reference line for comparison. Statistical notations shown in each subplot include r (Spearman correlation coefficient), R² (coefficient of determination from linear regression), and the slope of the fitted line.

Annual trends in anthropogenic pollution discharges and connection with changes in water quality

While most pollution source studies lack publicly available data for direct comparison, they consistently address interannual discharge variations. As illustrated in Fig. 8, long-term sectoral pollution discharge trends assessed using the Mann-Kendall test show clear and sustained decreases (P < 0.01). Further analysis based on Sen’s slope estimator indicated that reduction rates for livestock farming, residential, and industrial sectors exceeded 200 Gg/year for COD and 20 Gg/year for NH₃-N, while crop farming discharge reductions remained substantially lower than other sectors (6 Gg/year for NH₃-N). These sectoral pollution discharge trends align closely with research from Ma et al.⁹. Regarding relative contributions, livestock sector discharges increasingly dominated total discharges measured by COD, while crop farming and residential sectors showed increasing contributions to ammonia nitrogen discharges, consistent with previous findings^5,68,69 on how different sectors respond to environmental policies.

Fig. 8. Annual variations of estimated sectoral anthropogenic discharges (2007–2022).

(a) Sectoral trends: Solid lines fit local trends using locally weighted scatterplot smoothing (LOESS) with 0.8 bandwidth; shaded areas represent 95% confidence intervals of fitted results. Each line shows one sector’s discharge trend over time. (b) National total trends: Dashed lines show changes in national total pollution discharges across all sectors, with contributions from individual sectors shown in stacked areas. Colors correspond to sectors shown in (a).

Additionally, recognizing that in-stream pollutant concentrations are jointly determined by upstream anthropogenic discharges and watershed characteristics (including hydrological processes, landscape retention, and biogeochemical transformations), examining relationships between our estimated anthropogenic discharge data and observed water quality changes provides an additional perspective on dataset applicability and reliability. We selected 148 stations (Fig. 9a) with complete annual water quality series from an extensive spatiotemporal water quality dataset covering four decades (1980–2022) in China⁷², calculating overall water quality trends from 2007 to 2022 (Fig. 9b) using three different time series of annual average concentrations: Q1 (the first quartile of the national assemblage of site-level annual measurements), Q2 (the median), and Q3 (the third quartile). Based on upstream catchment boundaries with these stations as outlets, we extracted corresponding watershed attributes (including average slope and drainage density calculated as the total length of channels in the watershed divided by its total area), meteorological data (annual averages of temperature (°C) and precipitation (mm) extracted from the ERA5-Land monthly averaged dataset⁷³) and estimated total pollution discharges within each watershed to form independent variable datasets. Using gradient boosting models to simulate relationships between water quality changes and these variables, SHAP analysis quantified anthropogenic impact contributions (Fig. 9c). Results show that among anthropogenic factors contributing to China’s continuous water quality improvement, control of point source pollution (urban residential and industrial sectors) was the dominant contributor, accounting for more than twice the contribution of non-point source sectors. This aligns with most attribution studies of China’s water quality changes^8–10, indicating effective point source pollution control with future improvements requiring enhanced non-point source pollution management.

Fig. 9. Connection between pollution sources and water quality in mainland China.

(a) Spatial distribution of 148 water quality monitoring sites. (b) Country-level trends in annual mean observed concentrations of COD and NH₃-N from 2007 to 2022. Solid lines represent estimated linear trends in concentrations. (c) Relative contributions of different factors to the interannual variances of water quality. Pie plots show the percentages of relative impact for three kinds of variables.

Limitations and prospects

While this dataset represents a significant advancement in spatiotemporal resolution for China’s water pollution sources, several limitations should be considered when interpreting and applying these data.

(1) Uncertainties in industrial sector estimates. Although R² values were approximately 0.6, city-level validation indicated that our industrial sector estimates tend to be slightly overestimated. This bias likely stems from the inherent limitation of our weight generation method: since industrial discharges involve over 60 industrial subcategories with vastly different emission coefficients, our models using industrial GDP and area can only capture approximately 60% of the variability in industrial discharges at the provincial scale. This variability may be further amplified during the downscaling process, as different cities have different representative industrial subcategories, and the nonlinear relationships between GDP and discharges may also vary. Obtaining city-level industrial discharge data (rather than only provincial-level data) would enable fitting nonlinear relationships between industrial GDP, area, and discharge at the city scale, better capturing local variations in industrial structure.

(2) Sectoral allocation assumptions. For the industrial sector, while GHSL non-residential areas capture over 70% of industrial enterprises, this incomplete coverage may contribute to spatial allocation errors, particularly for dispersed small-scale industries not well represented in GHSL. Industrial source grid identification could be further improved by integrating high-resolution building identification from remote sensing imagery. For the livestock sector, although household low-intensity livestock farms account for over 70% of production in most Chinese provinces according to China Animal Husbandry and Veterinary Yearbooks⁴⁴ and previous studies, the distinction between point and non-point source pollution within the livestock sector remains partially unresolved in our dataset. While we incorporate detailed livestock distribution data, we may underestimate pollution concentration around large-scale intensive farming operations. This approach distributes pollution more evenly across rural areas than may actually occur, potentially missing critical hotspots near concentrated animal feeding operations. Future improvements could integrate satellite remote sensing data to identify and map large-scale intensive farming operations. By distinguishing between point-source intensive farms and diffuse household farms and applying differentiated emission coefficients, the spatial accuracy of livestock estimates could be substantially enhanced.

(3) Limited pollutant coverage. The dataset includes only COD and NH₃-N, omitting other critical pollutants such as total phosphorus (TP) and total nitrogen (TN) that are essential for comprehensive eutrophication assessment. This limitation stems from inconsistent reporting of these pollutants across provinces and years in the source statistics. Future work should incorporate TN and TP to provide comprehensive nutrient budgets, pending improved consistency in provincial-level reporting of these pollutants.

(4) Temporal resolution could be further enhanced. While our dataset provides annual resolution, users requiring seasonal or monthly data could build upon our annual gridded estimates through temporal disaggregation approaches. For the residential sector, monthly patterns could be derived by incorporating temperature-dependent water consumption variations and seasonal migration patterns. For the agricultural sector, annual fertilizer applications could be temporally distributed based on crop calendars, planting and harvesting schedules, and precipitation patterns, as nutrient losses are highly dependent on rainfall timing relative to fertilization events.

Usage Notes

In this study, “anthropogenic pollution discharge” refers to the total amount of pollutants generated from human activities that have the potential to enter surface water bodies. This includes: (1) direct discharge through wastewater from residential and industrial sectors, which enters water bodies via sewage systems or direct release to rivers, lakes, and reservoirs; and (2) indirect discharge from rural residential areas without sanitary facilities, crop farming, and livestock farming, where pollutants (in the form of human and animal waste and fertilizer residues) remain on land surfaces and reach water bodies through surface runoff during precipitation events. Our dataset quantifies pollutant loads at the source prior to transport processes, representing the potential pollution pressure on aquatic ecosystems rather than actual in-stream concentrations. The actual amount of pollutants reaching and remaining in specific water bodies depends on transport processes, retention in soils and landscapes, and in-stream attenuation. Quantifying these factors requires coupling with hydrological and biogeochemical models.

It should be noted that the environmental impacts of COD and NH₃-N are scale- and context-dependent. While our dataset quantifies total discharge amounts, actual ecological harm depends on receiving water body characteristics, background water quality conditions, and ecosystem sensitivity. Users should interpret discharge values in conjunction with local environmental conditions and water quality standards when assessing pollution impacts or prioritizing control measures.

Author contributions

T.M. and Z.Y. conceptualized and designed this study. Z.Y. collected the raw data and wrote the codes used to generate the dataset and drafted the manuscript. All authors contributed to the writing and revisions of the manuscript.

Data availability

The High-resolution Sectoral Water Pollution Discharge Dataset⁷¹ generated in this study has been deposited to Zenodo (https://zenodo.org/records/16930435). The dataset is provided in TIFF format. Additional details on the dataset contents and variables are described in the Data Records section.

Code availability

The source data and codes used for generating the study’s dataset are available on Zenodo⁷⁴ (https://zenodo.org/records/16925777).

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41597-026-06595-8.