Human alterations to global riverine phosphorus fluxes to the ocean Authors: Guowangchen Liu, Dongfeng Li, Ting Zhang, Jim Best, James J. Elser, Gabriel M. Filippelli, Hanqin Tian, Yi Zhao, Shang Tian, Anmeng Sha, Shouliang Huo, Hanxiao Zhang, Wei Zhi, Lei Chen, Zhenyao Shen, Feng Zhou, Qiuwen Chen, Yong Liu, Jinren Ni Published: 2025 ABSTRACT Rivers regulate land-ocean total phosphorus (TP) fluxes critical to ecosystem health and food security, yet global dynamics remain poorly understood due to limited observations. Here, we develop a machine learning framework integrating multimodal data and 280,000 TP measurements to reconstruct TP flux patterns over 1980–2019 across 420 major rivers. Results reveal a deceptive global equilibrium. While TP flux declines in the Northern Hemisphere, driven by dam trapping, particularly in Western Europe (−16.2%) and Eastern Asia (−8.7%), it rises in the Southern Hemisphere due to increased fertilizer use and deforestation, especially in Southern Africa (+15%) and the Malay Archipelago (+20.3%). Notably, the number of small rivers with rising TP flux is nearly double that of large rivers. This growing TP export from small rivers and Southern Hemisphere basins may intensify eutrophication, expand hypoxic zones, and threaten fishery yields. Our findings highlight a shifting global phosphorus landscape and underscore the need for more targeted, sustainable phosphorus management strategies. INTRODUCTION Phosphorus is an essential element for living organisms, playing a crucial role in biological processes, agricultural productivity, and freshwater ecosystem health (1–3). For all life, phosphorus serves as a critical building block in the genetic materials DNA and RNA, lipids in cell membranes, and the energy carrier adenosine triphosphate (4). Phosphorus fertilizers are vital for crop growth and food production (5, 6), but their excessive use results in nutrient runoff and subsequent eutrophication of water bodies (7–9). This process not only degrades water quality but also leads to the production of algal toxins that bioaccumulate through the food chain, posing risks to freshwater ecosystems and human health (10, 11). Fundamental environmental and ecological processes in global river basins are increasingly affected by anthropogenic activities, such as pollution discharge, dam construction, and land-use change (3, 12, 13), which have profound implications for total phosphorus (TP) flux (7, 12). Pollution discharge, primarily through agricultural expansion and urbanization, often increases phosphorus input into rivers via runoff and soil erosion (14–16), as well as the contributions of human and animal wastes (17). Dam and reservoir construction can reduce TP flux substantially by trapping sediment-bound phosphorus (11). Land use changes, particularly deforestation and phosphate mining (18), exacerbate soil erosion and increase nutrient runoff, further influencing TP levels in river basins (19). Furthermore, climate change can exacerbate or mitigate phosphorus transport through alterations in precipitation patterns, river discharge, and temperature regimes (6). This combination of pollution discharge (20), dam construction (13), and changes in land use and climate presents escalating and unpredictable threats to riverine and marine ecosystems globally. However, estimating global river TP flux remains highly uncertain. Differences in assessment methodologies, including mechanistic models such as Global-NEWS (21) and IMAGE-GNM (22), as well as empirical approaches (23, 24), can lead to divergent estimates of TP flux. Moreover, the absence of long-term global TP flux monitoring data impedes systematic investigations into long-term trends and driving mechanisms. Although global water quality data have increased over recent decades (25, 26), identifying large-scale drivers are historically constrained by data that are scarce, inconsistent, and discontinuous. Recent advances in machine learning methodology integrating multimodal inputs offer potential to partially mitigate these data limitations (27). We capitalize on these advances to develop a robust machine learning framework integrating multimodal inputs with a global compilation of 280,000 in situ TP observations, to reconstruct four decades of changes in TP concentrations and fluxes in major rivers worldwide from 1980 to 2019. We select 420 major rivers that discharge into the oceans or inland seas, each with a drainage area exceeding 20,000 km2 and collectively supporting ~5.2 billion people. In this study, our definition of river-ocean TP flux focuses specifically on the phosphorus discharged at the outlet of the river, as it represents a net response to multiple human activities on land and directly affects marine ecosystems. This comprehensive analysis provides insights into the spatial and temporal variability of global river-ocean TP flux, highlighting the complex interactions between human activities and natural processes in shaping riverine phosphorus dynamics and thus their potential impacts on algal blooms, water quality, and fishery yields. We highlight a global shift in phosphorus flux estimation and propose sustainable and forward-looking phosphorus control strategies to help sustain global riverine and coastal ecosystems. RESULTS AND DISCUSSION Large rivers dominate global phosphorus flux We observe that the long-term average river-ocean TP flux over 1980–2019 exhibits considerable spatial heterogeneity across 420 major rivers, ranging from 27 tonnes year−1 (Deseado River, Argentina) to 640 × 103 tonnes year−1 (Amazon River), with a median value of 1092 tonnes year−1. Notably, 57 rivers display an annual TP flux exceeding 7192 tonnes (mean value), collectively accounting for 83% of the global TP flux (Fig. 1B). Regionally, South America contributes 35% of the global TP flux, followed by Africa at 21%, and then Asia, North America, and Europe/Eurasia contributing 18, 12, and 12%, respectively (Fig. 1, A and D). Global distribution of average TP flux from the world’s rivers over the period 1980–2019. (A and D) Pie charts represent the proportion of TP flux contributed by each continent, with the box plots illustrating the levels of TP flux in the top three highest phosphorus-discharging rivers in each continent. In the pie charts, orange, almond, and blue represent the top three rivers with the highest TP flux in descending order, while gray signifies the combined TP flux of all other rivers in the respective continent. The central label indicates the continent’s contribution to the global TP flux percentage. For the box plots, the colors of the top three TP flux rivers correspond to those shown in the pie chart. Each box spans the 25th and 75th percentiles, with a solid black line denoting the median TP flux and hollow points indicating the mean value. (B) Global map of average discharge-weighted TP flux for 420 major rivers shows global patterns in TP flux over the period 1980–2019. Symbols are scaled by average annual TP discharge, with red indicating rivers with higher mean TP flux and blue indicating those with lower values. (C) Latitudinal profiles illustrating the spatial distribution of TP flux. A bar is plotted for every 5° of latitude, with red and blue bars denoting higher and lower than the global average TP flux level, respectively. Gray dashed line represents the global mean TP flux. t yr−1, tonnes per year. The spatial variability in river-ocean phosphorus flux is driven largely by big rivers with high water discharge. In Europe/Eurasia, the Lena, Ob, and Volga rivers contribute a total of 48% of the TP flux, explaining why an elevated TP flux is observed near 70°N (Fig. 1B). In North America and Asia, the Mississippi and Yangtze Rivers play dominant roles contributing 16 and 32%, respectively, of their continent’s TP flux and thus driving the elevated flux near 30°N (Fig. 1A). Near the equator, the Amazon (South America) and Congo (Africa) rivers are responsible for 58 and 37% of the TP flux, respectively. The Paraná River, contributing 13% of South America’s TP flux, dominates the area around 40°S while, at the same latitude in Oceania, TP flux levels are minimally influenced because of the relatively smaller river discharges (Fig. 1C). Big rivers exert a markedly greater influence on TP flux in the Southern Hemisphere compared to those in the Northern Hemisphere. A closer examination of the top three rivers contributing to TP flux on each continent reveals that in North America, Europe/Eurasia, and Asia, these rivers account for 42, 48, and 34% of the respective continent’s TP flux, averaging 41%, all below 50% (Fig. 1A). Conversely, in South America, Africa, and Oceania, the top three rivers contribute significantly more—76, 53, and 74% of their continent’s TP flux (averaging 68%), respectively, which is considerably higher than in the Northern Hemisphere (Fig. 1D). While big rivers predominantly govern global TP flux discharges, the TP flux yield per unit area reveals an inverse pattern, with small rivers exhibiting higher yields. The median drainage area of the 420 rivers is 39,500 km2, with a mean of 170,000 km2. We classify these rivers into six categories on the basis of drainage area and calculate the mean TP flux yield per area for each category. Results indicate that basins with drainage areas less than 20,000 km2 show a mean yield of 1565 tonnes km−2 year−1, whereas basins with drainage areas between 50,000 and 100,000 km2 exhibit a mean yield of 329 tonnes km−2 year−1 (fig. S6). Hot spots of changes in global phosphorus flux We find that the global river-ocean TP flux exhibits a deceptive equilibrium (−1.56 ± 0.64%/10 years, P < 0.05) sustained through opposing regional trajectories: decreases in the Northern Hemisphere and increases in the Southern Hemisphere (Fig. 2A). The northern river-ocean TP flux exhibits an overall decline over the past four decades (−3.15 ± 0.19%/10 years, P < 0.01, Fig. 2E). In particular, substantial declines in TP flux occur between 25°N and 55°N (Fig. 2B). However, there has been a resurgence of global phosphorus flux at a rate of 3.24 ± 0.42% per decade since 2000 (P < 0.05), particularly in Pan-Arctic regions above 55°N. In the equatorial regions and the Southern Hemisphere, many of which are located in developing regions, the TP flux exhibits an overall increasing trend over 1980–2019 (+1.46 ± 0.55%/10 years, P < 0.01), particularly around 20°S, where it increases by 15% from 1980–1999 to 2000–2019. Spatial pattern of TP flux change rates before and after 2000. (A) The global map of average discharge-weighted TP flux change rates for major rivers shows the global patterns in TP flux change rates, comparing periods before and after the year 2000. The TP flux change rate is determined by subtracting the pre-2000 TP flux values from the post-2000 values and then dividing by the pre-2000 values. Symbols are scaled by average annual TP discharge. (B) Latitudinal profiles exhibiting the spatial distribution of rate of change in TP flux. A bar is plotted every 5°, with brown and green bars denoting higher and lower than the global average TP flux change level, respectively. (C) Density distribution of TP flux change rates across continents. The horizontal axis represents the percentage change in TP flux, with color-coded density curves indicating the distribution in each continent. (D) Rates of change in TP flux for representative regions comparing before and after the year 2000, with brown and green indicating regions where the TP flux change rate is greater and less than zero, respectively. For the box plots, each box spans the 25th and 75th percentiles, with a solid black line denoting the median TP flux change rate and hollow points indicating the mean value. “*” indicates P < 0.05, “**” indicates P < 0.01, and “n.s.” indicates a nonsignificant change rate (P > 0.05). (E) Mann–Kendall (M-K) analysis of temporal trends in TP flux. The green solid and dashed lines represent Northern Hemisphere TP flux change and 2-year moving average, respectively, and the brown solid and dashed lines represent Southern Hemisphere TP flux change and 2-year moving average, respectively. Mt yr−1, million tonnes per year; yr, years. Globally, ~20, 12, and 68% of rivers exhibit significantly (P < 0.05) increased, significantly decreased, and nonsignificant changes in TP flux trends from 1980–1999 to 2000–2019, respectively (Fig. 2A). To better investigate the variations in TP flux in different regions, we select eight representative regions (Fig. 2A). After 2000, TP flux declines in Western Europe (−16.19%, P < 0.01), continental US and Middle America (−5.63%, nonsignificant), and Eastern Asia (−8.66%, P < 0.01) but increases in the Pan-Arctic (+6.13%, P < 0.01), South and Southeast Asia (+6.16%, P < 0.05), Northern South America (+10.89%, nonsignificant), Southern Africa (+15.04%, P < 0.01), and the Malay Archipelago and Northern Oceania (+20.34%, P < 0.01) (Fig. 2D). Our analysis further reveals that the change trends of TP flux are related to basin size (28). Basins exhibiting a decreasing TP flux have a mean drainage area of 300,000 km2, more than twice the mean area of 120,000 km2 observed in basins with an increasing TP flux (P < 0.05) (fig. S7A). Also, distinct patterns emerge between big rivers (>170,000 km2) and small rivers (<170,000 km2) (fig. S7B). In particular, nearly half (~47%) of big rivers show a decline by −30 to 0% in TP flux from 1980–1999 to 2000–2019. Conversely, more small rivers exhibit an increasing trend in TP flux as compared to big rivers. Drivers of global phosphorus flux dynamics Agricultural development is one of the most important causes of global phosphorus release due to the extensive use and runoff of phosphorus fertilizers (fig. S12). Fertilizer use is the main source of global phosphorus overloading, especially in the Southern Hemisphere (Fig. 3A). The global mean phosphorus fertilizer use intensity is ~8.3 kg per ha, with a median of 7.0 kg per ha (Fig. 3A) (29). Spatially, regions with high phosphorus fertilizer use intensity are concentrated in the agricultural areas of Eastern Asia (28.7 kg per ha, 3.46 times the global level) and Western Europe (14.7 kg per ha, 1.77 times the global level). Over 1961–2019, there is a notable shift in phosphorus fertilizer use from Western Europe to regions such as Eastern Asia and Northern South America (fig. S8A). Improved agricultural practices and environmental regulations (30, 31), particularly in Europe, have led to a reduction in the quantity of phosphorus fertilizer used (fig. S8B). Despite reductions in current fertilizer use, accumulated phosphorus from historical applications, often referred to as the “phosphorus legacy,” continues to contribute to water body pollution for extended periods (32). Phosphorus fertilizer use and damming. (A) Global map of average phosphorus (P) fertilizer use for major river basins from 1980 to 2019. Symbols are scaled by basin P fertilizer annual use, with red indicating rivers with higher P fertilizer use and blue indicating those with lower values (the global mean P fertilizer use intensity is ~8.3 kg per ha, with a median of 7.0 kg per ha). t yr−1, tonnes per year. (B) Long-term changes in dam capacity ratio from 1920 to 2020 for different typical regions. The dam capacity ratio is calculated as the dam capacity divided by the basin area, and is expressed in million cubic meters per square kilometer (mcm km−2). (C) Relationship between dam capacity ratio (from the Global Reservoir and Dam Database) and proportion of fertilizer discharged (regionally averaged TP flux divided by the corresponding P fertilizer use) to the ocean for different regions. The points in different colors represent different regions. The red dashed line indicates the fitted correlation (r = −0.66, P = 0.07), and the light red-shaded area represents the 95% confidence interval. P Arctic, Pan-Arctic; W Europe, Western Europe; US & M America, continental US and Middle America; E Asia, Eastern Asia; S SE Asia, South and Southeast Asia; N S America, Northern South America; S Africa, Southern Africa; MA & N Oceania, Malay Archipelago and Northern Oceania. Phosphorus trapping by dams also contributes to the decreasing trend present for the Northern Hemisphere. From the 1950s to 2000, there was a surge in dam construction, with ~114,467 km3 of added dam capacity, accounting for 90% of all historical dam capacity (Fig. 3B). The dam capacity and dam capacity ratio (defined as dam capacity per unit basin area) vary across different regions, with the most notable increases observed in continental US and Middle America, Eastern Asia, and South and Southeast Asia, where the capacities increased by 2, 17, and 91 times, respectively. The dam capacity ratios in these regions also rose to 0.25, 0.17, and 0.17 km3 per km2. A negative correlation is observed between dam capacity and phosphorus fertilizer use across various regions (Fig. 3C). Although regions with increased dam capacity also have higher-than-average fertilizer use, these dams intercept much of the particulate phosphorus, preventing additional TP flux to the oceans, similar to the reduction in sediment load caused by damming (11, 33). Dethier et al. (34) revealed that dams control the global reduction in sediment load, with significant quantities of sediment retained in dams, particularly in the Northern Hemisphere. Intensive land use change, such as deforestation (35), is a primary driver of the increase in TP flux in the Malay Archipelago region, where the deforested area accounts for up to 30% (fig. S13) due to agricultural expansion (36) and urbanization (37). Palm oil plantations in Southeast Asia, which rapidly expanded over the past few decades, are well-documented contributors to increased nutrient runoff, including phosphorus, into rivers and coastal zones (38). Changes in temperature and precipitation drive the variability in TP flux in the Pan-Arctic. With increasing temperature and rainfall, the Pan-Arctic region is experiencing a rise in TP flux, a trend only slightly influenced by other human activities (fig. S9). Increased permafrost thaw across Siberia and North America releases previously frozen organic matter into the active hydrological system (39). Due to climate change, river runoff and sediment load are also increasing rapidly in the Pan-Arctic region, which may be the primary reason for the increase in phosphorus flux here (figs. S10 and S11). We speculate the significant increase in phosphorus flux in the Yukon and Mackenzie rivers correlates with rising temperatures and increased discharge over the past few decades (40). Global phosphate production is primarily concentrated in regions such as China, Morocco, and the US (fig. S5). Advances in chemical wastewater treatment technologies effectively remove phosphorus to levels below regulatory limits, mitigating environmental impacts and supporting sustainable phosphorus management. The broad implications of changing global phosphorus flux Changes in phosphorus input affect biological primary productivity, algal blooms, dead zones, as well as the carbon cycle (10, 41, 42). The levels of chlorophyll a (Chl-a) are the most intuitive indicator of the effects responsive to phosphorus inputs. We collected Chl-a concentrations from 420 river estuaries worldwide spanning the period from 1997 to 2019, based on NASA’s publicly accessible Ocean Color database. High Chl-a levels are primarily concentrated in Western Europe (4.84 mg m−3), Eastern Asia (3.61 mg m−3), and continental US and Middle America (5.80 mg m−3), exceeding global average levels (3.34 mg m−3) (Fig. 4). This is similar to the spatial distribution of TP flux (Fig. 1B) and phosphorus fertilizer use (Fig. 3A), corroborating the impact of river-ocean TP on coastal ecosystems. Furthermore, our data suggest a strong relationship between Chl-a and TP flux. For instance, the Zambezi River exhibits a significant correlation between Chl-a and TP flux [correlation coefficient (r) = 0.48, P < 0.05; Fig. 4F]. However, damming can affect the response of Chl-a to TP flux. In the Yangtze River (Fig. 4D), the correlation between Chl-a and TP flux is negligible (r = 0.04, P > 0.5), perhaps due to the operation of the many mega-dams in recent years. Overall, these patterns in riverine TP discharges appear linked to coastal algal blooms at the global scale (10). Global change in coastal Chl-a and in various regions. (A) Global average Chl-a from 1997 to 2019. This value represents the average Chl-a in marine areas within a 100-km radius centered on river estuaries. Two representative basins, the Yangtze and Zambezi rivers, are marked on the map. (B) Seasonal variations in average Chl-a across different regions. (C and E) True-color images of periods of algal blooms for the Yangtze and Zambezi rivers, based on Landsat satellite images. (D and F) Trends in TP flux (yellow lines) and mean Chl-a in their estuarine areas (green lines) for two rivers [(D) Yangtze and (F) Zambezi] from 1997 to 2019. Each solid point denotes the annual average TP flux and Chl-a, and the error bars show the SE. The light gray shading represents the impact of the period of construction of the Three Gorges Dam on the response of Chl-a to TP flux. t yr−1, tonnes per year; US & M America, continental US and Middle America; E Asia, Eastern Asia; MA & N Oceania, Malay Archipelago and Northern Oceania; N S America, Northern South America; P Arctic, Pan-Arctic; S SE Asia, South and Southeast Asia; S Africa, Southern Africa; W Europe, Western Europe. Excessive phosphorus in small rivers can affect the health of socio-ecological systems (Fig. 5). Increased phosphorus flux into coastal waters can lead to eutrophication, causing oxygen depletion in the water and leading to hypoxic or anoxic conditions that are detrimental to aquatic life (43) (violet rectangle on Fig. 5). For example, in August 2017, the hypoxic zone in the Gulf of Mexico was declared to be the largest ever measured (44). Dead zones in Chesapeake Bay entail massive die-offs of fish, crabs, and other aquatic organisms (45). Phosphorus-induced changes in coastal ecosystems can also influence greenhouse gas emissions, including carbon dioxide and methane (46). In the Baltic Sea, increased nutrient inputs drive elevated emissions of these gases, exacerbating climate change impacts (47). Meanwhile, increasing phosphorus discharge also leads to a series of socioeconomic impacts (48–50) (violet rectangle on Fig. 5). Declines in water quality and subsequent fish and shellfish mortality damage both commercial and recreational fisheries, resulting in economic losses (51–53), with the impact on fisheries being profound in the Gulf of Mexico dead zone (54). Phosphorus-driven algal blooms can produce toxins that contaminate seafood and water supplies (55), posing serious health risks to humans and animals (56). Deteriorating water quality and the formation of hypoxic dead zones also reduce the attractiveness of coastal areas for tourism, affecting local economies that rely on this sector. The loss to tourism-related businesses due to the 2018 Florida red tide bloom was estimated to be $2.7 billion USD, demonstrating that algae-driven impacts on tourism can cause multibillion-dollar disasters (57). Conceptual framework for a comparison of natural and anthropogenic processes of phosphorus generation. The left side illustrates the natural factors driving river-ocean phosphorus (P) flux, and biogeochemical cycling of phosphorus. The right side shows the anthropogenic factors driving phosphorus flux and the impact of algal blooms. Our findings suggest that current phosphorus management involves a spatial mismatch. Strong phosphorus control measures and practices are implemented in large river basins and developed regions, yet the impacts are manifested predominantly in small rivers and developing regions through eutrophication, hypoxia, and altered nutrient cycling. To address the issue of global TP pollution and the global imbalance in the shift of the river-ocean phosphorus flux, we identify several strategies toward sustainable and forward-looking phosphorus management. In heavily dammed river basins in developed regions, our global river-ocean TP assessment supports proactive reservoir sediment and phosphorus management strategies, such as periodic dredging or controlled sediment routing, and recovery of trapped phosphorus for reuse, thereby reducing legacy phosphorus release and preserving downstream water quality (58). For developing regions, we suggest adopting precision phosphorus fertilizer strategies including soil-test guided placement, right-rate application, and fertilizer deep placement to significantly reduce anthropogenic TP flux while sustaining crop yields and advancing sustainability (59). Furthermore, sustainable phosphorus management should also mitigate the persistent impacts of legacy phosphorus, enhance cleaner industrial production to contain waste discharge, restore riparian zones, and account for other drivers such as climate change and phosphate mining, which increasingly influence phosphorus transport and availability (60–62). Ultimately, successful global phosphorus management necessitates data-driven solutions and collaborative efforts between governments and local communities to ensure the long-term health of riverine and coastal ecosystems in a changing world. MATERIALS AND METHODS Data sources and processing TP data for riverine systems worldwide are sourced predominantly from the Global Riverine Water Quality Archive (GRQA) database (63). To ensure integrity of the dataset, TP values from the GRQA are cross-referenced meticulously with those documented in primary sources. Furthermore, an exhaustive literature review is conducted, using search terms such as “total phosphorus/TP,” “phosphorus/P concentration,” and “riverine total phosphorus/TP” across platforms including Web of Science, ResearchGate, and Google Scholar. This comprehensive search aims to integrate the most recent studies not captured by the GRQA. Given China’s pivotal role in global phosphorus flux, long-term TP concentration data are also integrated from the Chinese water quality dataset (64) and the China National Environmental Monitoring Center (65). Both the temporal and spatial availability of the measured data are critically assessed. Temporally, only datasets with monitoring periods exceeding 5 years are selected to mitigate the risk of inaccuracies and support reliable long-term trend simulations. Spatially, only regions within the same basin as the study site or within a 1° proximity to the basin boundary are included. This consideration aligns with the first law of geography, near things are more related than distant things (66), ensuring that the model’s precision is not compromised by greater distances between research and monitoring locations. Data from 280,331 TP samples are compiled, establishing a robust and comprehensive foundation for analysis (fig. S1). Multimodal environmental variables are derived from established databases. The variables used to project TP values in riverine systems encompass hydrometeorological data (precipitation, temperature, and runoff), basin attributes [Digital Elevation Model (DEM), slope, and soil type], and anthropogenic activities (dam construction, phosphorus fertilizer application, and land use/land cover) (67). Hydrometeorological data, which primarily define the fundamental driving forces within monitored catchment areas, are sourced predominantly from the European Centre for Medium-Range Weather Forecasts (ECMWF) Reanalysis v5 (ERA5)-Land, offering monthly scale data on 50 parameters, including precipitation, temperature, runoff, and evaporation. Basin attributes, which describe relatively stable characteristics of the catchment areas, are obtained from widely recognized databases in Google Earth Engine (GEE). Data on anthropogenic activities, reflecting the direct impact of human production on phosphorus flux, include dam construction data from the GRanD v1.3 database (68), phosphorus fertilizer use data from database (29), and land use/land cover change data from GEE. Global phosphorus simulation framework We develop a robust machine learning framework integrating the above multimodal inputs with a global compilation of 280,000 in situ TP observations. To build reliable models for predicting the coastal phosphorus values, we compare different machine learning approaches, including logistic regression, support vector machine (linear, polynomial, and radial basis function), stochastic gradient descent regression, k-nearest neighbor regression (uniform weight and distance weight), decision tree regression, random forest, gradient boosting tree, extra tree regression, multilayer perceptron regression, convolutional neural network, and extreme gradient boosting (69). Several classic machine learning methodologies are explored. Logistic regression, a statistical technique for binary classification, predicts the probability of a dichotomous outcome. Support vector machines are potent classifiers that determine the optimal hyperplane for separating data points across different kernels, such as linear, polynomial, and radial basis functions (70). Stochastic gradient descent regression is an iterative optimization method suited for large-scale and sparse datasets. K-nearest neighbor regression estimates the target value of a new data point by averaging the values of its k-nearest neighbors, using either uniform or distance-based weighting. Although these methods are widely used in fields such as geosciences, environmental studies, and hydrology, they exhibit limited efficacy in simulating complex global phosphorus patterns (71). Decision tree–based machine learning methods excel in capturing intricate nonlinear relationships and are extensively used in large-scale research. Decision tree regression models the target variable by learning simple decision rules derived from data features, effectively partitioning the data into branches (72). Random forests, an ensemble learning technique, construct multiple decision trees during training and yield the average prediction from these trees (73). Given the substantial dataset assembled in the present study, neural network-based methods are also used. Multilayer perceptron regression, a neural network comprising multiple layers of neurons, adeptly captures complex nonlinear relationships in the data (69). Convolutional neural networks, specifically designed for structured grid data like images, use convolutional layers to identify features. In comparison to the singular models, ensemble models (encompassing bagging, boosting, and stacking algorithms) enhance generalization by integrating multiple models, thereby reducing the overfitting risk associated with individual models. Boosting methods, such as gradient boosting trees, extra tree regression, and extreme gradient boosting, emerge as the preeminent tree-based ensemble learning algorithms due to their robust predictive power (74). Gradient boosting trees construct an ensemble of trees sequentially, with each tree correcting its predecessor’s errors to enhance accuracy. Extra tree regression, or extremely randomized trees, builds an ensemble of unpruned decision trees with randomly chosen splits to reduce variance (75). XGBoost, an optimized gradient boosting library, is renowned for its efficiency, flexibility, and speed, making it a preferred choice in machine learning competitions (76). The compiled datasets are randomly split, with 70% allocated for training and 30% for testing. Traditional evaluation methods may lack precision when assessing performance with limited test data, often due to the potential lack of representativeness. Cross-validation mitigates this by partitioning the data into multiple folds and iteratively training and validating the model across different folds, offering a more reliable and consistent performance evaluation while minimizing overfitting or underfitting risks (77). A 10-fold cross-validation approach is used to evaluate model performance (table S2). During this process, the training data are divided into 10 equal subsets; nine subsets are used for training, while the remaining subset is reserved for validation. To further reduce the uncertainty associated with stochastic sampling and to identify the most predictive models, we train the machine learning models 100 times, optimizing for the highest determination [coefficient of determination (R2), Nash-Sutcliffe efficiency (NSE), bias, and Metric Multidimensional Scaling Analysis (MdSA)]. Each run involves three iterations of tenfold cross-validation. In addition to ensuring an adequate sample size within the training dataset, the use of ensemble models, cross-validation methods, and appropriate feature selection is critical in mitigating overfitting by reducing model complexity and minimizing the impact of irrelevant or redundant features. For TP predictions, we develop several models on the basis of the compiled datasets, using the SciPy package in Python 3.11. Model performance is evaluated using the R2, NSE, and bias. High R2 values, coupled with low bias and error, indicate superior model performance. Among the models tested, all demonstrated good training performance, but not all models achieved acceptable testing performance. We selected the extra tree regression because of its superior testing performance. The high generalization capability of this algorithm is more robust when implemented globally, as compared to the other algorithms (fig. S2D and table S1). Global phosphorus flux prediction For coastal phosphorus simulation, the geographic coordinates of the prediction sites are delineated using the HydroBASINS database. The precise boundaries of river basins are determined through expert judgment, specifically using level 3 or 4 HydroBASINS. To minimize spatial heterogeneity and reduce the uncertainty in simulations caused by extensive catchment areas, hydrometeorological data near monitoring sites are selected as representative watershed characteristics. When incorporating basin attributes and human activities data, the complex interplay of anthropogenic factors necessitates the use of mean basin values (such as DEM and slope) or aggregate totals (e.g., population size and fertilizer application amounts) as input variables. These spatial datasets are preprocessed using GEE and Arc Geographic Information System (ArcGIS) and subsequently standardized via the GeoPandas package in Python. Global runoff data are sourced from the globally resampled dataset produced by the Global Reach-level Flood Reanalysis (GRFR) model, with numerical evaluations conducted by aligning the river reaches to the corresponding monitoring sites. The GRFR provides a comprehensive 3-hourly record of river discharge for 2.94 million river reaches worldwide over a 40-year span (1980–2019). This dataset is underpinned by a modeling chain that includes the Variable Infiltration Capacity (VIC) land surface model (0.05°, 3-hourly), which is rigorously calibrated and bias corrected, along with the Routing Application for Parallel computation of Discharge (RAPID) routing model (covering 2.94 million river/catchment vectors). The input data for precipitation is derived from MSWEP (Multi-Source Weighted-Ensemble Precipitation), while other meteorological fields are downscaled from ERA5. Validation against 3-hourly flow records from over 6000 gauges in the continental US, and daily records from more than 14,000 gauges globally, demonstrates robust modeling performance across all flow ranges, with strong capabilities in reconstructing global discharge patterns. The global discharge data are averaged monthly and then used to compute global TP flux. TP flux is calculated using the relationship between TP concentration and discharge (78). For training the machine learning models, we use basins from the HydroBASINS dataset to adequately capture spatial heterogeneity, ensuring rigorous model training and validation. Uncertainty analysis of global phosphorus simulation-prediction method While the phosphorus simulation-prediction method provides accurate estimates of phosphorus concentration and flux in global rivers, with good model validation performance using in situ TP data (R2 = 0.7109, NSE = 0.7109, and bias = 0.65%) (fig. S2D), there are a few potential limitations of uncertainty. First, our monitoring data represent either single measurements at a specific point in time or averages of multiple measurements. These are thus influenced by sampling location, collection methods, and chemical analysis techniques and may not always reflect the actual phosphorus levels in rivers precisely. Meanwhile, most data in our simulation-prediction method are based on averaged hydrometeorological data, basin attributes, and human activities within the watershed or sampling location (67). This results in a more comprehensive evaluation of the phosphorus levels in the current watershed-river system. The disparity between these two approaches may introduce uncertainties in the data sources (79). Second, simulation differences among various machine learning methods lead to uncertainty. Singular models tend to perform poorly on the complex nonlinear problem in the present study. Only random forest regression achieves relatively acceptable results (R2 = 0.7069, NSE = 0.7062, and bias = 0.80%) (fig. S2A), while other singular models show R2 < 0.6 and are thus are excluded (69). By contrast, ensemble models demonstrate stronger predictive power for global phosphorus simulation (fig. S2). Gradient boosting trees regression (R2 = 0.7014, NSE = 0.6979, and bias = 0.56%), extra tree regression (R2 = 0.7109, NSE = 0.7109, and bias = 0.65%), and XGBoost (R2 = 0.6996, NSE = 0.6995, and bias = 0.56%) all performed well in simulating global phosphorus levels. However, XGBoost is more prone to overfitting, even after optimal parameter tuning (fig. S2B) (80). Gradient boosting trees have weak performance in training (fig. S2C). Therefore, we ultimately select extra tree regression as the primary method for this study, despite the uncertainties associated with this method being the optimal approach for global phosphorus simulation. Third, the simulation method in this study is based on machine learning algorithms, which carry uncertainty in parameter selection (81). Uncertainty mainly stems from the complexity of the model structure and the degree of nonlinearity. Taking the extra tree regression model as an example, parameters such as the number of trees, maximum depth, and maximum features determine the model’s structure and performance. Simplified structures can impair the model’s ability to capture nonlinearity in complex problems, while overly complex structures can lead to overfitting (82). We carefully select key model parameters on the basis of parameter uncertainty analysis (fig. S4). The number of trees is set at 200, with a maximum depth of 50 for each individual tree. The minimum number of samples required to split an internal node and the minimum number of samples required in a leaf node are both set to 3, while each tree considers a maximum of 200 features for each split. Additionally, each time the model is retrained, and it starts from scratch without reusing previous trees. Bootstrap samples are used to build the trees to further enhance model robustness (83). The multimodal machine learning framework integrates a global compilation of 280,000 in situ TP observations with diverse environmental and socioeconomic data as predictors. This approach provides a holistic and globally consistent estimation of TP flux, offering an unprecedented resolution compared to fragmented regional studies. The framework is particularly adept at capturing the complex, nonlinear relationships and subtle interactions between multiple environmental variables and TP flux, a capability that surpasses traditional statistical models and simpler machine learning algorithms. By integrating multimodal data, the framework achieves significantly enhanced accuracy over traditional methods. Change rate calculation In the present study, the change rate of TP flux is estimated by comparing the average TP flux over two distinct periods within a long-term time series. The segmentation of these periods is achieved using the Pettitt test to identify breakpoints in the TP flux time series. The Pettitt test is a nonparametric statistical method renowned for its robustness in detecting abrupt changes in environmental variables (84). Applying the Pettitt test, we identify the year 1999 as a significant change point in global TP flux. Subsequently, we calculate the average TP flux for the periods 1980–1999 and 2000–2019. The change rate is calculated using the following formulaChange rate=∑i=20002019TPfluxim−∑j=19801999TPfluxjn∑i=19801999TPfluxim×100(1)where i and j represent the years in the post- and prechange periods, respectively. TP fluxi and TP fluxj denote the annual TP flux values for the corresponding years. m and n are the number of years in the post- and prechange periods, respectively.