Changes in China’s river water quality since 1980: management implications from sustainable development

Rivers provide critical ecosystem services and freshwater for drinking water, irrigation, aquaculture, navigation, and power generation, yet are increasingly degraded by anthropogenic activities and climate change. China’s rapid development since 1978 has imposed strong pressures on river water quality, contributing to national water shortages and nutrient-driven impairments with downstream impacts such as algal blooms and red tides. Despite notable national improvements since 2003 attributable to reduced discharges, a significant fraction of sites still fail to meet Class III standards. Key knowledge gaps persist due to limited long-term monitoring (pre-2003), scale-related uncertainties in identifying drivers (natural geography, socioeconomic indicators, land use, meteorology), and challenges translating science into management linked to SDGs. This study aims to reconstruct and predict spatiotemporal variations in riverine TN, TP, NH4-N, and CODmn from 1980–2050, quantify the relative roles of anthropogenic, climatic, and geographical drivers, and propose SDG-aligned management strategies.

Prior research has examined nutrient inputs to Chinese rivers from multiple sources and at multiple scales, nutrient cycling within river systems, and spatial patterns of impairment, as well as global threats to human water security and biodiversity. National assessments (2003–2017) documented widespread inland water quality improvements linked to reduced nutrient discharges. Nonetheless, uncertainties remain due to limited long-term monitoring data, difficulties attributing drivers across scales, and integration of scientific insights into policy. Studies highlight precipitation-driven variability in nutrient loading, the predominance of anthropogenic sources for N and P (e.g., point sources dominating TDP in the Yangtze), and the need to consider interactions among SDGs for effective pollution control.

Data: Monthly TN, TP, NH4-N, and CODmn from 613 monitoring sites across 10 major river basins (2003–2018) were obtained from the China National Environmental Monitoring Center. Ancillary datasets included geography (DEM-derived elevation/slope, 1 km), climate (CN05.1, 0.25°; temperature, precipitation, wind, extremes), soils (1:1,000,000 digital soil properties), land use (30 m), socioeconomic indicators (1 km), and net anthropogenic N and P inputs (NANI/NAPI) from reported discharge data and coefficients. Modeling: A stacking machine-learning framework combined three base learners—Random Forest (RF), Support Vector Machine (SVM), and k-Nearest Neighbors (KNN)—to simulate monthly concentrations for 1980–2018 and project decadal trends for 2020–2050 under SSP2-RCP4.5 and SSP5-RCP8.5. Variable screening used Maximum Information Coefficient (MIC<0.25), collinearity checks (Spearman R>0.8), and retraining/selection based on predictive relevance (Spearman R>0.4). Stacking employed a linear combination y = Σ wm fm with nonnegative weights summing to 1, estimated via quadratic programming using cross-validated predictions (leave-one-out style to avoid overfitting). Validation: Ten-fold cross validation and metrics including R2, RMSE, Nash–Sutcliffe efficiency (NSE), and MAE quantified performance; models reproduced spatial patterns and time series with p<0.01 and low bias. Attribution: Multiple linear regression (MLR) with annual means (1980–2018) per sub-watershed was fit separately by basin to quantify contributions of grouped drivers (anthropogenic, climatic, geographical). Standardized coefficients were converted to contribution percentages by the ratio of absolute coefficients to their sum. Scenarios: Interdecadal projections (2020–2050) under SSP2-RCP4.5 and SSP5-RCP8.5 evaluated potential future exceedances and basin-specific vulnerabilities.

- Model skill: Stacking models robustly reproduced measured TN, TP, NH4-N, and CODmn across basins (p<0.01) with favorable R2, RMSE, NSE, and MAE; ten-fold validation confirmed generalizability. - National trends (1980–2018): TN pollution increased overall. The proportion of sites with simulated TN<1.5 mg/L was ~20% in 1980 and remained ~18–20% through 2018, indicating a persistent high TN burden. NH4-N rose from 1980 to 2010, then declined (2010–2018). CODmn rose (1980–2000) then declined (2000–2018). TP rose (1980–2015) then declined (2015–2018). - Threshold exceedances: Across sites, 62.3% of samples had TN>1.5 mg/L. Simulated observations exceeding TN>1.0 mg/L, NH4-N>6.0 mg/L, and CODmn>60 mg/L were 13.3%, 16.3%, and 13.7%, respectively. - Basin differences: TN decreased notably in the Yangtze (n=41, p<0.05, r=−0.86) and Pearl (n=66, p<0.01, r=−0.97). Trends varied in other basins (e.g., Huaihe n=48, p<0.05, r=−0.21). Northwestern basins (e.g., Yellow, Huaihe) saw early increases in NH4-N; most regions showed recent declines in TN, NH4-N, and CODmn. - Future scenarios (2020–2050): Under SSP2-RCP4.5, proportions exceeding TN>1.5 mg/L, TP>0.2 mg/L, NH4-N>1.0 mg/L, and CODmn>6.0 mg/L are approximately 75.8%, 18.2%, 40.2%, and 17.6%, respectively. Under SSP5-RCP8.5, the corresponding proportions are 75.2%, 19.6%, 43.2%, and 17.5%. Nutrient pollution risks are projected to be significant, especially in the Yellow, Huaihe, and Haihe basins. - Driver attribution: Anthropogenic factors dominated variability in TN, TP, and NH4-N across basins, with higher contribution ratios than climatic and geographical drivers. CODmn drivers varied by basin; e.g., in the Northwest Inland River, geographical (41.60%) > anthropogenic (32.83%) > climatic (25.55%), while climatic drivers dominated in the Songhua, Yellow, Huaihe, Southwest, and Southeast rivers. Precipitation-related variables were strong predictors for TN, TP, and CODmn; temperature was generally weak. Among anthropogenic variables, higher percentages of urban area and farmland were positively associated with nutrient concentrations, while forest and grassland were negatively associated; population and NANI/NAPI were strong predictors in several basins. - Policy signals: COD decreased after the 2000 Total Amount of Pollutants Control Plan and the 2008 Water Pollution Prevention and Control Law; additional reductions in TP and NH4-N followed the 2015 Action Plan. Agricultural N and P loads declined from 2007–2017. However, TN showed limited improvement nationally, likely reflecting the lack of explicit TN control targets in the assessment system.

The study reconstructs four decades of riverine nutrient dynamics in China, overcoming monitoring gaps by leveraging a validated stacking machine-learning framework. The dominant role of anthropogenic drivers for TN, TP, and NH4-N indicates that management targeting land use, population pressures, and net anthropogenic nutrient inputs is essential to reduce riverine nutrient levels. Climatic and geographical influences, while secondary for most nutrients, are important for CODmn in several basins and modulate nutrient delivery via precipitation-driven hydrology and sediment processes. The results explain the mixed national trends: policy-driven improvements in CODmn, TP, and NH4-N contrast with persistent TN pollution, which lacks stringent control targets. Projections under SSPs highlight continued risks without strengthened measures, particularly in northern and north-central basins. Integrating water resources, water environment, aquatic ecology, and water security—aligned with SDG 6 and SDG 14—can decouple economic growth from nutrient pollution, emphasizing regionally tailored strategies and governance flexibility.

By combining 16 years of observations with an ensemble machine-learning approach, this work reconstructs and projects riverine TN, TP, NH4-N, and CODmn in 613 sub-watersheds across China (1980–2050). It shows nationwide improvements in TP, NH4-N, and CODmn linked to policy interventions, but persistent TN pollution likely due to the absence of explicit TN control targets. Anthropogenic drivers dominate nutrient variability, underscoring the need for targeted reductions in net nutrient inputs, better land-use planning, urban wastewater upgrades, and agricultural best management practices. Future work should integrate high-resolution data on pollution control infrastructure, hydrological regulation, and drainage networks; improve process understanding to complement data-driven models; and refine projections under evolving climate–socioeconomic scenarios. Policy should embed TN within control and assessment frameworks and adopt flexible, region-specific strategies to achieve SDG-aligned water quality improvements.

Key limitations include: (1) limited long-term monitoring prior to 2003, necessitating retrospective simulation; (2) the data-driven stacking approach has limited interpretability and deductive power relative to process-based models; (3) incomplete, heterogeneous, or low-resolution data for critical management variables (e.g., wastewater treatment capacity, drainage infrastructure, water conservancy operations) likely reduce explanatory power; (4) moderate explained variance for some targets in attribution models, especially NH4-N; and (5) uncertainties in future projections due to scenario inputs and potential unmodeled policy or technological changes.