Rivers are vital for freshwater resources, but face degradation from human activities and climate change. China's rivers have experienced significant water quality impairment since 1978 due to rapid economic development, with nutrient pollution a major contributor. While improvements were observed from 2003-2017 due to reduced nutrient discharge, challenges remain. Existing research has quantified N and P inputs and spatial water quality patterns, but gaps exist in understanding long-term spatiotemporal variation and underlying mechanisms. This study addresses these gaps by using a long-term dataset and sophisticated machine learning models to analyze water quality changes and identify driving forces to inform future water quality management and achieve sustainable development goals (SDGs). The lack of long-term, regularly informative monitoring data, limitations in identifying driving mechanisms due to spatial and temporal scale resolution of models, and the difficulty in bridging scientific research and management applications were identified as key limitations of prior research.

Several studies have examined water quality patterns in China's rivers and their drivers, including quantifying nutrient inputs from various sources, nutrient cycling in river systems, and spatial water quality patterns. However, these studies have limitations, including a lack of long-term data, challenges in identifying driving mechanisms due to limitations in spatial and temporal resolution, and the difficulty of bridging scientific research and management applications. This study aims to address these gaps by utilizing a comprehensive 16-year dataset and advanced machine-learning techniques to provide a more comprehensive understanding of the spatiotemporal dynamics of water quality in China's rivers.

This study used a 16-year (2003-2018) monthly dataset from 613 river water quality monitoring sites, combined with watershed characteristics (longitude, latitude, land use, anthropogenic N/P inputs, soil properties), to build stacking machine-learning models. These models integrated random forest (RF), support vector machine (SVM), and k-nearest neighbors (KNN) to improve prediction accuracy and stability. The models simulated and predicted annual and monthly variations in river water quality from 1980-2018 and projected decadal trends (2020-2050) under two future scenarios (SSP2-RCP4.5 and SSP5-RCP8.5). Multiple linear regression (MLR) models and correlation analyses quantified the relationships between anthropogenic, climatic, and geographical factors and changes in TN, TP, NH₄-N, and CODₘ concentrations. The Maximum Information Coefficient (MIC) was used for variable selection, addressing collinearity and ensuring predictive accuracy. A quadratic programming algorithm optimized model weights within the stacking framework. Model performance was evaluated using R², RMSE, NSE, and MAE. The MLR models assessed the contribution of anthropogenic, climatic, and geographic factors to nutrient variability in each of the 10 river basins separately. Data sources included the China National Environmental Monitoring Center, Resource and Environment Science and Data Center, Institute of Soil Science (Chinese Academy of Sciences), Institute of Geographic Sciences and Natural Resources Research, and China Statistical Yearbooks.

The machine-learning models accurately recreated TN, TP, NH₄-N, and CODₘ concentrations. TN concentrations showed an overall increase from 1980-2018, while TP, NH₄-N, and CODₘ showed initial increases followed by decreases in later years. These trends reflect the impact of water quality management policies. Spatial variations in nutrient concentrations existed among the 10 river basins. Anthropogenic factors were the primary drivers of TN, TP, and NH₄-N variations, exceeding the influence of climate and geographical factors. Specifically, the percentage of farmland, urban area, population, and anthropogenic N and P inputs were strong predictors. Forestland and grassland showed a negative relationship with nutrient concentrations. Precipitation was a strong predictor for TN, TP, and CODₘ concentrations. The study highlighted significant spatial variations in river water quality, particularly in eastern China. Projections under future scenarios indicate a continued influence of human activities and climate change on river nutrient concentrations.

The findings highlight the significant impact of anthropogenic activities on river water quality in China. While improvements have been observed in some parameters due to government policies and regulations, the persistent increase in TN concentrations underscores the need for strengthened TN control targets and assessment systems. The study's attribution analysis confirms that anthropogenic factors outweigh climatic and geographic influences on nutrient levels, emphasizing the crucial role of land-use change, urbanization, and agricultural practices. The model projections highlight the importance of sustainable water management practices to mitigate the projected impacts of future human activities and climate change. The variations in the relative contributions of anthropogenic, climatic, and geographic factors across different river basins underscore the need for region-specific water quality management strategies.

This study provides a comprehensive assessment of changes in China's river water quality since 1980, utilizing advanced machine learning and statistical techniques. The findings highlight the dominance of anthropogenic factors in driving nutrient pollution and the need for integrated water resource management strategies aligned with the SDGs. Future research could focus on refining the models by incorporating additional data, improving the representation of complex interactions between various factors, and further exploring the effectiveness of different water quality management strategies at various scales. This would help inform sustainable and effective water resource management policies for China and other developing countries.

The study relies on available data, which might not fully capture the complexity of all factors influencing water quality. The models, while robust, have limitations in fully explaining all observed variations. Further research is needed to incorporate more comprehensive data on pollution control measures, such as drainage systems and sewage treatment plant capacities, due to data collection challenges. Also, the model's interpretability is limited. Future research can enhance the model's explanatory power, and uncertainty associated with future scenarios needs further evaluation.