A comprehensive global dataset of river nutrient concentrations, loads and streamflow Authors: Jordan Peters, Alasdair Noble, Crile Doscher, Richard McDowell Published: 2026 ABSTRACT AbstractAccess to high quality river water quality data is crucial for environmental research and informing policymakers about management decisions. Currently, publicly available river data is spread across several large national and international databases and sparse in both spatial and temporal coverage in much of the world. In response to this, we produced three related data outputs which we have combined into a single comprehensive dataset: (1) a global dataset of total nitrogen, phosphorus and streamflow measurements compiled from seventeen river databases, containing over around 20,000 unique monitoring sites for total nitrogen, 40,000 sites for total phosphorus and 32,000 streamflow sites, (2) a set of 1,292 total nitrogen sites and 2,302 phosphorus sites with daily water quality values from a concentration estimation model and (3) global estimates of nutrient loads and concentrations produced by extrapolating the second output to areas with low data coverage using random forests. The dataset is intended to aid in future hydrological research and can also be used to advise on local river management, identify hotspots which may be prone to high nutrient loads, and compute global nutrient fluxes form land to sea. Background & Summary Access to unpolluted river water is vital for ecosystems across the planet. Humans rely on rivers as a source of water for drinking, cultivation of food and energy production1. However, global river water quality has been declining since the late 20th century under climate change and from the intensification of agricultural practices2,3, with increasing eutrophication in rivers and lakes of particular concern. Eutrophication refers to the process in which high nutrient (nitrogen and phosphorus) concentrations in freshwater lead to large algal blooms, which are known to cause negative effects including inducing hypoxic and anoxic conditions by limiting biologically available oxygen4, disrupting drinking water production5, harming recreational appeal, and causing fish kills6. In 2008, the combined annual costs associated with eutrophication were estimated to be around US $2.2 billion7. Modelling and forecasting of nutrient concentrations in rivers are important for monitoring and controlling eutrophication at both local and regional levels. However, scientists frequently highlight a lack of data as a key issue in carrying out river water quality related research2,3,8,9. Without data, the scope and extent of global river studies to quantify risk and management of nutrient impairing water quality are limited, producing less verifiable and reproducible outputs. Machine learning approaches are often used to simulate global river streamflow, but applications for water quality modelling are limited by a lack of training data10. Furthermore, simulated data are usually estimates of means and not extreme events like droughts or floods which may be of interest to freshwater researchers3,9,11–15. Implementation of informed river management policy is reliant on the availability of accurate nutrient load data, particularly for extremes in flows which are known to dominate total nutrient loads16. Existing global river databases like the Global Runoff Data Centre17, GEMStat18 and the Global River Water Quality Archive19 have two main issues: (1) most countries in Africa, Asia and South America carry out little river monitoring so coverage in those regions is sparse, and (2) the data which have been collected are split between multiple databases, making informed calculation of loads difficult. The first problem limits the usefulness of global river water quality models as many climate zones and terrestrial biomes will not appear in the input data. Poor representativeness is difficult to remedy, although the combination of many publicly available databases will highlight the areas which are currently lacking in data. This paper is motivated by the second issue and aims to develop robust load estimations by collating sixteen publicly available river water quality and quantity databases into one data product able to be used for a variety of freshwater research. The data product focuses on total nitrogen (TN) and phosphorus (TP) loads, with the goal of maximising the number of sites with nutrient concentration and streamflow measurements over the same period. This dataset may be of interest to those involved with local river management as it can potentially be used to identify hotspots and periods of time where nutrient loads or concentrations take extreme values or exceed regulated limits. Furthermore, the dataset can be analysed to calculate estimates for global nutrient transfers, such as from rivers to the ocean. These values could be compared against thresholds like planetary boundaries, which are “the environmental limits within which humanity can safely operate20,21”. Tracking of finite quantities such as phosphorus could also be investigated. To calculate loads, we also produce estimates of daily nutrient concentrations and loads for the days between observations. Most existing nutrient concentration data are measured at a monthly frequency which is not suitable for many purposes such as research into water quality during short-term extreme weather events. For this reason, we apply the Weighted Regression in Time, Discharge and Seasonality (WRTDS)22 model. Other global nutrient load models have been developed in the past, although they generally use less comprehensive models than WRTDS or are fitted using smaller input datasets23–25. An example of a potential use case of these estimates is explored by modelling mean TN and TP concentrations and loads based on the land and climate characteristics around each site, which will then be used to predict loads and concentrations across the world using the HydroBASINS river catchment boundaries26. Figure 1 shows the methodology used to construct the data product. A graphical summary of the approach taken to produce the data. The numbering corresponds to the stages in the methods section. Methods Source data The water quality data used are a combination of thirteen national and international databases (Table 1, Fig. 2, Supplementary Table 1 for data access instructions) chosen to provide the greatest spatial coverage. For each dataset, all sites with measurements of TN, TP or streamflow were used. The Global Freshwater Quality Database (GEMStat18) was the starting point for inclusion, supplemented by regional datasets from Europe (Waterbase27, WQA28, Naïades29 and HMS30), North America (ECCC31 and USGS32), Australia (Victoria Water Measurement Information System33 Department of Water and Environmental Regulation, Western Australia34 and the River Murray Water Quality Monitoring Program35) and New Zealand (LAWA36, NRWQN37). Data from the Global River Water Quality Archive (GRQA19) were also used in areas where coverage was still sparse. The full extent of the GRQA was not used because it already includes several of the other contributing datasets. The site IDs for the included GRQA sites can be found in the metadata files contained in the dataset associated with this manuscript. Table 1Summary of water quality data sources.DatabaseRegionNumber of sitesTimespan of observationsMedian length of series (years)Median number of observationsTotal number of observationsEnvironment and Climate Change Canada (ECCC)31Canada203 (229)2000–2024 (1991–2024)17.3 (23.0)90 (107)27,899 (37,618)Global Freshwater Quality Database (GEMStat)18Global3,368 (5,142)1973–2023 (1954–2023)7.7 (7.8)40 (41)232,640 (418,331)UK Harmonised Monitoring Scheme (HMS)30United Kingdom0 (37)NA (1974–2013)NA (10.9)NA (133)NA (5,255)Land, Air, Water Aotearoa (LAWA)36New Zealand1,026 (1,020)2004–2019 (2004–2019)13.4 (13.9)125 (132)126,622 (130,291)River Murray Water Quality Monitoring Program35Murray-Darling Basin, Australia10 (10)1980–2017 (1980–2017)37.0 (37.0)1270 (1271)12,540 (12,571)Eau France (Naïades)29France1 (12,755)2007–2017 (1978–2024)10.9 (13.1)255 (49)255 (1,250,266)National River Water Quality Network (NRWQN)37New Zealand76 (76)1989–2020 (1989–2020)31.4 (31.4)363 (375)26,727 (27,697)United States Geological Survey (USGS)32United States6,512 (8,967)1900–2024 (1900–2025)7.8 (8.2)33 (36)492,313 (768,409)Water Measurement Information System (VIC)33Victoria, Australia167 (416)1993–2024 (1978–2024)4.6 (19.7)30 (179)14,157 (105,040)Department of Water and Environmental Regulation, Western Australia34Western Australia113 (115)1979–2024 (1971–2024)12.4 (13.4)67 (78)19,962 (21,245)Waterbase27Europe7,396 (13,456)1971–2022 (1970–2022)6.7 (8.1)26 (29)276,860 (564,259)Water Quality Data Archive (WQA)28United Kingdom810 (1,871)2000–2024 (2000–2024)3.4 (2.9)65 (36)104,329 (173,723)Global River Water Quality Archive (GRQA)19Global645* (553*)1969–2020 (1965–2020)13.1 (17.0)69 (98)69,241 (77,824)Total20,327 (44, 647)1900–2024 (1900–2025)7.5 (9.0)37 (39)1,403,541 (3,590,533)The first numbers in each field correspond to the total nitrogen sites, with the total phosphorus sites in parentheses.*Only a subset of the GRQA was used. Summary of water quality data sources. The first numbers in each field correspond to the total nitrogen sites, with the total phosphorus sites in parentheses. *Only a subset of the GRQA was used. Locations of total nitrogen (TN), total phosphorus (TP) and streamflow monitoring sites for all data sources combined. The streamflow data are a combination of nine databases (Table 2, Fig. 1, Supplementary Table 1 for data access instructions), chosen to provide the greatest coverage possible in areas where water quality data is present. The streamflow databases include five of the same sources as for the water quality data (GEMStat, LAWA, Murray-Darling, NRWQN, USGS) along with three additional databases: CAMELS-BR38 for Brazil, HYDAT39 for Canada, and NRFA40 for the United Kingdom. Other publicly available water quality datasets exist, but most are measured too infrequently41–43 to be included. Table 2Summary of streamflow data sources.DatabaseRegionNumber of sitesTimespan of observationsMedian length of series (years)Median number of observationsTotal number of observationsCatchment Attributes and MEteorology for Large-sample Studies – Brazil (CAMELS-BR)38Brazil3,5321911–201925.5800435,724,490Global Runoff Data Centre (GRDC)17Global8,6531806–202448.216,312145,914,897National Water Data Archive (HYDAT)39Canada5,8951860–202422.65,69351,600,827United States Geological Survey (USGS)32United States12,3771857–202534.210,228175,319,698Global Freshwater Quality Database (GEMStat)18Global1321974–202232.4389207,777Land, Air, Water Aotearoa (LAWA)36New Zealand3901960–202130.910,9833,526,790River Murray Water Quality Monitoring Program35Murray-Darling Basin, Australia101934–201743.315,688168,093National River Flow Archive (NRFA)40United Kingdom1,5711879–202449.117,16725,093,712National River Water Quality Network (NRWQN)37New Zealand761989–202031.411,443866,260Total32,6361806–202536.411,134438,422,544 Summary of streamflow data sources. Data preprocessing The combined water quality and streamflow monitoring databases included around 4,000 duplicates from different databases, or rarely, two sites in close spatial proximity with varying periods of measurement or recorded values. Combining these sites removed temporal and spatial duplication of sites in the dataset, and in rare cases, combined time series to provide a longer period of measurement than either series alone. This was done using the following process: All data from the source databases were extracted and written to individual files for each site in the same format and units (mg L−1 for concentration and m3 s−1 for streamflow), retaining only those with measurements over a period of more than one year. Water quality parameters were kept consistent by only including measurements reported as total nitrogen and total phosphorus rather than combining individual chemical parts (e.g. NH3, NOX) or physical forms (e.g. dissolved or solid P). For this reason, all TN and TP data were of unfiltered samples. Based on previous uses of the source databases (e.g. McDowell, et al.23), we trust the results have been through quality assurance and control procedures. If values were marked as being below the detection limit from their source database, they were set to that detection limit. All analytical methods were included, unless they were specified in the original database as being unreliable (e.g. as in some USGS data). If a site had more than one measurement on a single day, the mean of the measurements was taken. All data from the source databases were extracted and written to individual files for each site in the same format and units (mg L−1 for concentration and m3 s−1 for streamflow), retaining only those with measurements over a period of more than one year. Water quality parameters were kept consistent by only including measurements reported as total nitrogen and total phosphorus rather than combining individual chemical parts (e.g. NH3, NOX) or physical forms (e.g. dissolved or solid P). For this reason, all TN and TP data were of unfiltered samples. Based on previous uses of the source databases (e.g. McDowell, et al.23), we trust the results have been through quality assurance and control procedures. If values were marked as being below the detection limit from their source database, they were set to that detection limit. All analytical methods were included, unless they were specified in the original database as being unreliable (e.g. as in some USGS data). If a site had more than one measurement on a single day, the mean of the measurements was taken. The “Spatial Join” tool from the GIS software ArcGIS44 was used to create a list of all pairs of sites for the same parameter (TN, TP or streamflow) within 200 metres of each other. The “Spatial Join” tool from the GIS software ArcGIS44 was used to create a list of all pairs of sites for the same parameter (TN, TP or streamflow) within 200 metres of each other. For each pair, the mean percentage difference of the days with measurements in both series was calculated. If there was no overlap in the period of measurement, the mean percentage difference was set to zero to guarantee that the sites would be combined. For each pair, the mean percentage difference of the days with measurements in both series was calculated. If there was no overlap in the period of measurement, the mean percentage difference was set to zero to guarantee that the sites would be combined. Iterating through the list of sites, all pairs with a mean percentage difference of less than 0.1% were combined to ensure that the sites with overlapping measurements contained the same values. Sites were combined by taking the mean of the series on each day with at least one measurement (meaning that if there was only one measurement on a day that value was used). Outliers were detected and removed when combining duplicate sites, although they are still present in the source data included in the dataset (folders TN/TP/Streamflow.1). Grouped sites then were removed from the list so they were not included in more than one group. If the difference was more than 0.1%, the sites were treated as separate groups. If there were more than two matching sites with a mean percentage difference less than 0.1%, the mean of all sites was taken. Iterating through the list of sites, all pairs with a mean percentage difference of less than 0.1% were combined to ensure that the sites with overlapping measurements contained the same values. Sites were combined by taking the mean of the series on each day with at least one measurement (meaning that if there was only one measurement on a day that value was used). Outliers were detected and removed when combining duplicate sites, although they are still present in the source data included in the dataset (folders TN/TP/Streamflow.1). Grouped sites then were removed from the list so they were not included in more than one group. If the difference was more than 0.1%, the sites were treated as separate groups. If there were more than two matching sites with a mean percentage difference less than 0.1%, the mean of all sites was taken. The coordinates for the new combined sites were the mean of the coordinates of all contributing sites, although changes were small due to the maximum distance of 200 metres between sites. This step reduced the number of global TN sites from 20,327 to 19,424, and TP sites from 44,647 to 41,639. The GRQA and GEMStat databases were responsible for most of the overlaps, containing many duplicates of sites found in the national databases. The number of streamflow sites reduced from 32,636 to 32,103. Pairing water quality and streamflow sites Loads and concentrations on days without water quality measurements can only be estimated at sites with water quality and streamflow data measured in the same location and over the same period. ArcGIS44 was used to identify pairs of streamflow and water quality sites matching in location by snapping sites to the nearest HydroRIVERS global river network reach45. The maximum snapping distance (i.e. the furthest distance a site could be moved to be placed on the river network) was 500 metres. There were less than 100 sites which were further than 500 metres away which were all measurements of small coastal drainage channels, so these were removed from the dataset. Streamflow and concentration sites were paired if they were within 50 km of each other, directly upstream or downstream of each other, and located on rivers of the same Strahler stream order (an indicator of the level of branching of a river segment, a variable in the HydroRIVERS dataset). These restrictions were put in place to ensure that sites are located on the same river and are close enough to reflect the same changes in flow each day. Fifty kilometres was chosen as an upper limit to maximise the chance of identifying sites with overlapping periods of measurement. The distances between paired sites can be found in the metadata presented alongside the river data. Water quality sites were matched with any streamflow site meeting these requirements, meaning there were usually multiple streamflow sites paired with each water quality site. These lists of site pairings are presented along with the dataset (TN/TP.2_downstream_matches.csv). An overlap in period of measurement between water quality and streamflow sites is dependent not only on the start and end dates of each series, but also the frequency of measurement throughout each series. This is because concentration and load estimations will be most reliable when the input data is measured regularly throughout the period of estimation. For these reasons, it was necessary to restrict modelling to series meeting some threshold of frequency of measurement. Many concentration estimation models, including the R implementation of WRTDS from the “EGRET” package46, require a complete daily series of streamflow measurements (i.e. no missing days). Hence, for streamflow data, years were only considered when they had at least 300 out of 365 daily measurements. The missing measurements can be approximated by linear interpolation provided there are not multiple consecutive missing days. The water quality data were restricted to years with at least ten months having at least one measurement. This guarantees that there are minimum of ten measurements each year and that they are spread throughout the year, capturing all series measured at least monthly which is the minimum frequency that is useful for nutrient load estimation16,47. Each water quality site was ultimately paired with the streamflow site within a 50 km radius on a river of the same Strahler stream order and connected via the HydroRIVERS network with the largest total number of years where both series met the above restrictions in frequency of measurement for a period of at least five years. This temporal pairing method was developed specifically to prepare data for modelling with WRTDS which is very restrictive, so the lists of all spatial pairings are made available alongside the dataset to allow users to pair sites for other uses (files TN/TP.2_downstream_matches.csv). This result is composed of 2,635 TP sites and 1,520 TN sites with at least five consecutive years with more than 300 streamflow measurements and concentration measurements in at least ten months per year (Table 3, Fig. 3). This output is almost entirely made up of data from Europe, North America and Oceania, with very few data in Africa, Asia and South America. A short investigation into the representativeness of the dataset relative to terrestrial biomes is presented in the Technical Validation section. Table 3Global coverage of total nitrogen (TN) and total phosphorus (TP) sites meeting the spatial and temporal coverage restrictions, with proportion of total shown in parentheses.ContinentTN sitesTP sitesAfrica25 (1.8%)23 (0.9%)Asia9 (0.6%)29 (1.2%)Europe335 (23.8%)1,122 (44.5%)North America409 (29.0%)526 (20.9%)Oceania631 (44.8%)807 (32.0%)South America0 (0%)13 (0.5%)Total1,4092,520 Global coverage of total nitrogen (TN) and total phosphorus (TP) sites meeting the spatial and temporal coverage restrictions, with proportion of total shown in parentheses. Location of total nitrogen (TN) and total phosphorus (TP) sites after filtering by overlap in space and time with streamflow data. Estimating missing water quality values The observational source data is missing many daily nutrient concentration measurements which may make it unsuitable for some analyses such as quantifying the flux of nutrients lost from land to freshwater or oceans48–50. For this reason, a second output was produced containing complete daily streamflow and water quality time series, estimated using the Weighted Regression in Time, Discharge and Season (WRTDS) model22. This model was chosen as it can more accurately capture temporal and seasonal effects in water quality parameters than other common models such as LOADEST51–53, and is frequently used for modelling nitrogen and phosphorus levels54–57. WRTDS is a concentration estimation method based on the L5 model where Qi and Cˆi are streamflow and estimated concentration at time ti, lnQ*® is the mean of the natural log of all streamflow values, t® is the mean time, i is an index for time, and the β symbols are the five regression parameters. Rather than fitting the L5 model once and using it to predict across the whole set of input streamflow data, WRTDS applies a moving window across the input data in the three dimensions of time, streamflow and seasonality, fitting a unique L5 model for each input data point. This moving window ensures that only the input data which are like the day to be estimated in all three dimensions influences the prediction. More information can be found in the original paper22. The “EGRET” package for the programming language R was developed to fit the WRTDS model22 to data from the National Water Information System (NWIS) from the U.S. Geological Survey58. Some care needed to be taken when applying the model to our dataset. To prepare the water quality and streamflow data for the WRTDS fitting process, the following steps were taken for each matched concentration and streamflow site from the source data: The water quality and streamflow pairs were truncated so they had the same start and end dates. The water quality and streamflow pairs were truncated so they had the same start and end dates. Any streamflow measurements for periods of up to two consecutive days of missing values were linearly interpolated. A maximum period of two days was set as a somewhat conservative limit to reduce the chances of poor flow estimation. Linear interpolation was chosen as it is one of the best performing flow estimation methods across two-day periods59–61. There were around 1,500 days with interpolated flows across the entire dataset (<0.01%), which are flagged in the files found in folders TN/TP.4. Any streamflow measurements for periods of up to two consecutive days of missing values were linearly interpolated. A maximum period of two days was set as a somewhat conservative limit to reduce the chances of poor flow estimation. Linear interpolation was chosen as it is one of the best performing flow estimation methods across two-day periods59–61. There were around 1,500 days with interpolated flows across the entire dataset (<0.01%), which are flagged in the files found in folders TN/TP.4. Both series were split at any point where there were more than four consecutive years without any water quality measurements. This was done to ensure that any sites with separated periods of data collection were only modelled together if they were not far apart in time. Four years was a conservative choice based on experimentation with fitting WRTDS to a variety of rivers with different gaps in data collection. This experimentation found that estimates became highly variable when there were gaps in water quality measurements of at least five years, provided there was at least five years of continuous measurement on either side of the gap. This is explored in more detail in the Technical Validation section. Both series were split at any point where there were more than four consecutive years without any water quality measurements. This was done to ensure that any sites with separated periods of data collection were only modelled together if they were not far apart in time. Four years was a conservative choice based on experimentation with fitting WRTDS to a variety of rivers with different gaps in data collection. This experimentation found that estimates became highly variable when there were gaps in water quality measurements of at least five years, provided there was at least five years of continuous measurement on either side of the gap. This is explored in more detail in the Technical Validation section. Following these steps, the WRTDS model needed to be tested on a variety of rivers to check its suitability for modelling on a global scale, to investigate the effects of varying the model parameters on predictions, and to formulate a consistent method which could be applied to all rivers in the dataset. The WRTDS model is intended to be used on time series with the following characteristics20: The number of water quality measurements is at least 200. The number of water quality measurements is at least 200. The period of water quality data collection is at least 20 years. The period of water quality data collection is at least 20 years. There are streamflow values for every day in the period being analysed. There are streamflow values for every day in the period being analysed. The water quality data should be representative of the entire cross-section of the river to ensure that loads calculated by multiplying concentration by volume are accurate. The water quality data should be representative of the entire cross-section of the river to ensure that loads calculated by multiplying concentration by volume are accurate. The river should not be so “flashy” at the sampling points such that data recorded at the time of measurement is not substantially different to the daily average. In this context, flashiness refers to the rate of fluctuation of flow in response to watershed input. The river should not be so “flashy” at the sampling points such that data recorded at the time of measurement is not substantially different to the daily average. In this context, flashiness refers to the rate of fluctuation of flow in response to watershed input. These conditions are restrictive when developing a method which is to be applied to thousands of rivers varying in size and behaviour around the world. For this reason, the dataset containing the output of the WRTDS predictions on the global dataset is presented with the disclaimer that predictions for individual sites should be treated cautiously if they do not meet all the above criteria. The original paper notes that for the fourth condition, “there is every reason to believe that this constraint can be eliminated”22. It was not feasible to check that the fifth condition holds for every river in the dataset as this would need to be evaluated on a site-by-site basis, so we assume that any sites with unrepresentative measurements will not significantly influence the results when considering concentrations and loads on a global scale. The preparation process above ensured that the third condition is met, meaning all that needed to be considered is the duration, the number of data points, and the flashiness of the river. Only 197 TN sites (12.6%) and 694 TP sites (23.8%) had continuous water quality measurements for a period of at least 20 years, meaning it was worth considering if the limits of WRTDS could be pushed to improve the number of viable sites and the spatial coverage of the data, acknowledging that estimation may be less reliable for shorter series. This motivated the design of the WRTDS testing process: Using a selected set of rivers with varying flow regimes and periods of data availability, compare the loads estimated by WRTDS for a range of model hyperparameter values. This (1) tested the performance of WRTDS for rivers of varying flow regimes; (2) provided insight into how altering the model hyperparameters affects the output; and (3) informed on what model hyperparameter values should be used when running WRTDS on the entire dataset, and whether these values should change based on the properties of each river. This analysis is described in detail in the Technical Validation section. The decision was made to restrict the application of WRTDS to series at least five years long, and the windowQ parameter was varied depending on the length of the series according to equation (3) where hQ is windowQ and T is the length of the series. This means that for series at least 10 years long, the default value of 2 is used. For sites with less than 10 years of measurements, users are advised that the hydrological profile of the concentration estimates may have been flattened due to the wide windowQ and concentrations during high flow events may be understated. There was no clear reason to alter windowY from the default value of seven years as it is not so small as to substantially alter predictions for small rivers, and altering windowY does not have much of an effect on larger rivers, even for extreme values. More information about the function of the model hyperparameters can be found in the original WRTDS documentation22,46. The model still failed to converge for around 3% of the sites which is mostly a result of unusual features in the input data such as sites with many days having zero streamflow. Out of the 1,409 TN sites suitable for WRTDS, 18 sites failed the WRTDS run and 38 of the 2,520 TP sites failed. Standard errors were calculated for each daily estimate which can be found alongside the estimates in the files found in folders TN/TP.4_-_WRTDS_Output and are investigated in the technical validation section. After fitting the model to all possible sites, Kalman filtering was applied to the estimates using the WRTDSKalman function from the EGRET package which is recommended for producing the best possible estimates rather than investigating trends53,62. After recombining the series that were split due to breaks in streamflow or concentration measurements as described above, 1,292 sites remain for TN and 2,302 for TP. Global extrapolation of load and nutrient models The WRTDS output was used to develop estimates of global river nutrient losses based on the Level 7 HydroBASIN catchment boundaries26 by fitting random forest models to the dataset and making predictions of TN and TP annual loads and mean concentrations across the world based on 46 geological, climate and land use characteristics (Table 4). Principal component analysis (PCA) was used to group similar highly correlated characteristics together. Four random forest models were fitted, estimating annual load and concentration for TN and TP separately. Loads and mean concentrations for input data were calculated as a mean of the five most recent years in each time series. The median time these five-year periods were centred around was the year 2017 for TN and 2016 for TP. Table 4Catchment characteristics used as predictors in the random forest model.ParameterResolutionDetailsTotal atmospheric N deposition (kg km−2 yr−1)732.5° × 2.5°Original source contains values for three years each decade between 1980s and 2010s, but WRTDS output ranges from 1962 to 2024, so a mean was taken and used as a fixed value. Wet and dry deposition were summed for each cell to calculate total and converted from 2° × 2.5° resolution to 2.5° × 2.5° via the “Point to Raster” tool in ArcGIS44.Wastewater production (106 m3 yr−1)85 arcminPCA of WorldClim Bioclimatic variables BIO1-BIO11 (Temperature)7430 arcsecEstimated value in 2010.PCA of WorldClim Bioclimatic variables BIO12-BIO19 (Precipitation)7430 arcsecEstimated value in 2010.PCA of cattle, chicken, sheep and goat livestock counts755 arcminEstimated count in 2010.Population density (count km−2)7630 arcsecEstimated population density in 2010.Potential Evapotranspiration (mm)7730 arcsecOlsen P (mg kg−1)781 kmSoil porosity (dimensionless)790.25° × 0.25°PCA of clay, silt and sand soil fractions790.25° × 0.25°Forested area (dimensionless)8030 arcsecProportion of catchment covered by land classes 1–10 according to Bartholomé and Belward80Rangeland (dimensionless)68,8030 arcsecProportion of catchment covered by land classes 11–15 according to Bartholomé and Belward80, excluding the tundra terrestrial biome68.Pasture cover (dimensionless)80,8130 arcsecProportion of catchment covered by land classes 16–18 according to Bartholomé and Belward80, split between pasture and crop based on the ratio between pasture and crop cover according to Ramankutty, et al.81Crop cover (dimensionless)80,8130 arcsecProportion of catchment covered by land classes 16–18 according to Bartholomé and Belward80, split between pasture and crop based on the ratio between pasture and crop cover according to Ramankutty, et al.81Bare land (dimensionless)8030 arcsecProportion of catchment covered by land class 19 according to Bartholomé and Belward80Proportion of catchment permanently covered by water (dimensionless)8030 arcsecProportion of catchment covered by land class 20 according to Bartholomé and Belward80Snow or ice cover (dimensionless)8030 arcsecProportion of catchment covered by land class 21 according to Bartholomé and Belward80Urban landcover (dimensionless)8030 arcsecProportion of catchment covered by land class 22 according to Bartholomé and Belward80Soil erosion (Pg yr−1)8212.5 arcminRunoff (104 km3 yr−1)830.5° × 0.5°Mean slope (degrees)8430 arcsecDerived from HydroSHEDs DEMMax slope (degrees)8430 arcsecDerived from HydroSHEDs DEMMean elevation (m)8430 arcsecDerived from HydroSHEDs DEMMean streamflow (m3 s−1)26For the input data, the mean streamflow is calculated from the time series for each site. For the prediction across Level 7 HydroBASINS, data is supplied directly for each HydroBASINS catchment from Linke, et al.26 Catchment characteristics used as predictors in the random forest model. Principal component analysis was applied separately to the bioclimatic variables BIO1–11 and BIO12–19 as well as to the livestock count and soil fraction predictors to account for high correlation between predictors. The first principal component (explaining 46.7% of the variance) was used in the final model for livestock, and the first two principal components were kept for the other groups of predictors to capture most of the variation in the predictors. The mean streamflow, population density and wastewater production predictors were also transformed by taking the natural log. The mean annual concentrations (calculated from the WRTDS output) were also modelled on the log scale, with bias correction preformed upon back transformation according to Miller63. The “Watershed” tool in ArcGIS44 was used to generate catchments for each of the points in the WRTDS output. Averages for each of the predictors in Table 5 across these catchments were then calculated using the “Zonal Statistics as Table” tool in ArcGIS44, which served as input data for the models. Table 5Lists of the five catchment characteristics with the highest importance in the random forest models for total nitrogen and total phosphorus loads and concentrations.TN mean concentrationTN annual loadVariableVariable importance (%)VariableVariable importance (%)Population density48.70Median flow (log)83.83Crop land-cover proportion26.53Crop land cover proportion5.78Wastewater produced5.49Population density2.29Mean slope of catchment3.72Water land cover proportion2.07Pasture land-cover proportion2.53Total atmospheric N deposition1.64TP mean concentrationTP annual loadVariableVariable importance (%)VariableVariable importance (%)Population density31.65Median flow (log)81.69Wastewater produced15.96Water land cover proportion3.41Crop land-cover proportion12.84Crop land cover proportion2.69Forest land-cover proportion10.80Forest land-cover proportion1.92Rangeland land-cover proportion7.97Rangeland land-cover proportion1.34 Lists of the five catchment characteristics with the highest importance in the random forest models for total nitrogen and total phosphorus loads and concentrations. The input data were split into training and testing sets at a 70:30 ratio for each region in the dataset to guarantee equal proportional regional coverage. Performance statistics for the models were then calculated based on the predictions made for the testing sets (Table 6). Then, all available input data were used to train the final model used for global predictions across Level 7 HydroBASINS, the results of which are shown in Figs. 4, 5. Predictions were not made for catchments with median annual streamflow values of 0 (marked as “Low flow” on the maps). The regressionforest function from the grf package64 for R was used to fit the random forest model, using 4,000 trees with a minimum node size of 5. Table 5 shows the five variables with the highest importance for each model. For both the TN and TP mean concentration models, population density and other measures of human influence like crop cover proportion and wastewater produced were the variables which were most influential on the model. For TN and TP annual loads, median flow was the most important by a large margin, followed by the land cover proportions of cropland and water (i.e. lakes and wide rivers). Table 6Performance statistics for the total nitrogen (TN) and total phosphorus (TP) load and concentration models.ModelRoot mean square errorMean absolute errorR2 (Nash-Sutcliffe efficiency)TN Concentration (log mg L−1)0.6450.4940.709TN Load (log kg yr−1)0.8290.6310.875TP Concentration (log mg L−1)1.0700.6860.478TP Load (log kg yr−1)1.2340.7870.601 Performance statistics for the total nitrogen (TN) and total phosphorus (TP) load and concentration models. Total nitrogen (TN) and total phosphorus (TP) global concentration estimates. “Low flow” refers to areas with a median streamflow of zero. Higher resolution tiff files can be accessed from https://doi.org/10.25400/lincolnuninz.30546371. Total nitrogen (TN) and total phosphorus (TP) global load estimates. “Low flow” refers to areas with a median streamflow of zero. Higher resolution images can be accessed from https://doi.org/10.25400/lincolnuninz.30546371. Data Records The dataset and all R code can be accessed from https://doi.org/10.25400/lincolnuninz.3054637165 including associated README and metadata files. Folders of data, csv files and R files are numbered sequentially according to their relevant stage of methodology: Source data prior to removal of duplicate sites Source data prior to removal of duplicate sites Source data after removal of duplicate sites Source data after removal of duplicate sites Paired streamflow and concentrations sites with overlapping periods of measurement in preparation for WRTDS modelling Paired streamflow and concentrations sites with overlapping periods of measurement in preparation for WRTDS modelling Output of WRTDS modelling and global extrapolation via random forest models Output of WRTDS modelling and global extrapolation via random forest models The folders with names beginning with TN/TP/Streamflow.X contain csv files of the observational and estimated concentration and streamflow data for each site (Note that there is no Streamflow.4 folder as streamflow values are contained along with the WRTDS concentration estimates in TN/TP.4). Metadata and summary statistics can be found in the corresponding csv files named TN/TP/Streamflow.X_meta.csv. All concentration and streamflow values are measured in mg L−1 and m3 s−1 respectively. The data from the UK National River Flow Archive40 and some of the sites from the GEMStat18 and GRDC17 databases are not able to be included due to their respective licensing agreements. These sites along with their associated site IDs are listed in the metadata files in the data folder and can be accessed from the original data repositories. Lists of publishable sites for GEMStat and GRDC were sourced from the UNEP GEMS/Water Global Freshwater Quality Archive66 and GRDC-Caravan67 respectively. Further information relevant to the methods used such as all possible concentration and streamflow site pairings can be found across the other csv files included in the dataset. All R scripts used for removal of duplicate sites, matching sites and modelling concentrations and loads are also made available. More detailed descriptions of the content of each file and column names can be found in the README file. Technical Validation Filtered data The source data is missing coverage in much of the world so some types of land cover may not be well represented. For this reason, the spread of data across the terrestrial biomes as described by Olson, et al.68 is presented in Table 7. Two chi-square tests were performed to test whether the distribution of TN and TP sites were significantly different from the expected distribution based on the total areas covered by each biome. Both produced p-values of <10−6, indicating that many of the biomes are under- or overrepresented in the site-matched source data. Most prominently, biome numbers 1, 6, 7, 11 and 13 are underrepresented, as well as the Rock and Ice terrestrial region which does not fall under any biome (Table 7). Conversely, biomes 4 and 8 (Temperate Broadleaf and Mixed Forests and Temperate Grasslands, Savannas, and Shrublands) are overrepresented in the data. There is not a strong link between the representation of a biome in the model input and the mean size of the residuals across the biome. For the biomes which we were able to compute residuals (i.e. biomes which are present in the input data), residuals for the TN model were similar, with most between 0.6 and 0.8 log units. The only outlier with larger residuals was the Tundra biome, although the mean of the residuals was only 0.92, and the biome contains less than 0.1% of the global population. Residuals varied more for the TP model, ranging from 0.26 to 1.12 log units. Of the biomes with mean residuals greater than 0.9, only Deserts and Xeric Shrublands contains a sizeable population (9.9%), although this is not unexpected as rivers in deserts are unusual when compared to the rest of the world. Overall, this shows that underrepresentation of certain land cover types in the input data for our nutrient models do not cause the model to perform substantially differently across the world. Table 7Coverage of site-matched source data relative to terrestrial biomes, with percentages of totals shown in parentheses.BiomePercentage of world covered by biomeTN sitesTP sitesPopulation in 2010 (millions)Mean of TN load absolute residuals (log kg yr−1)Mean of TP load absolute residuals (log kg yr−1)1. Tropical and Subtropical Moist Broadleaf Forests13.3%No Coverage9 (0.3%)2,217.4 (32.6%)No Coverage0.462. Tropical and Subtropical Dry Broadleaf Forests2.6%8 (0.5%)9 (0.3%)524.3 (7.7%)0.780.263. Tropical and Subtropical Coniferous Forests0.5%No CoverageNo Coverage58.0 (0.8%)No CoverageNo Coverage4. Temperate Broadleaf and Mixed Forests8.6%997 (65.6%)1972 (74.8%)1,711.1 (25.1%)0.520.645. Temperate Coniferous Forests2.6%73 (4.8%)104 (3.9%)85.4 (1.3%)0.610.706. Boreal Forests/Taiga10.5%27 (1.8%)41 (1.6%)40.3 (0.6%)0.781.107. Tropical and Subtropical Grasslands, Savannas, and Shrublands14.6%11 (0.7%)18 (0.7%)593.8 (8.7%)0.450.398. Temperate Grasslands, Savannas, and Shrublands7.2%331 (21.8%)400 (15.2%)289.0 (4.2%)0.650.899. Flooded Grasslands and Savannas0.8%No CoverageNo Coverage103.0 (1.5%)No CoverageNo coverage10. Montane Grasslands and Shrublands3.3%31 (2.0%)31 (1.2%)117.0 (1.7%)0.740.8411. Tundra5.9%4 (0.3%)2 (0.1%)2.2 (<0.1%)0.921.1412. Mediterranean Forests, Woodlands, and Shrublands2.3%19 (1.3%)30 (1.1%)306.3 (4.5%)0.560.9513. Deserts and Xeric Shrublands18.0%19 (1.3%)19 (0.7%)674.0 (9.9%)0.750.9614. Mangroves0.2%No CoverageNo Coverage85.4 (1.3%)No CoverageNo CoverageRock and Ice (Terrestrial Ecoregion)9.7%No CoverageNo Coverage1.2 (<0.1%)No CoverageNo Coverage Coverage of site-matched source data relative to terrestrial biomes, with percentages of totals shown in parentheses. WRTDS parameter testing Because WRTDS was applied to thousands of rivers around the world, all with different sizes, in different climates and with different flow regimes, it was important to determine that the model is suitable for a range of rivers. Also, it would be informative to see how the model performs when applied to data which do not comply with the recommendations made in the original paper22, specifically those relating to the length of the time series and the flashiness of the rivers. For these reasons, the following two objectives must be achieved: Determine the minimum length of time series for which the model converges. Determine the minimum length of time series for which the model converges. Determine how altering the hyperparameters of the model affect the output for a variety of river types. Determine how altering the hyperparameters of the model affect the output for a variety of river types. Three New Zealand rivers were chosen as testing sites: The Waimakariri River, the Wairua River, and the Rangitopuni Stream (Table 8). These sites were selected because they each contain over 20 years of consecutive overlapping daily streamflow data and monthly water quality data for TN and are all different in terms of size and flow regimes. The Waimakariri River is the largest New Zealand river in the dataset meeting all the filtering criteria and has a small proportional change between median and 90th percentile flows. The Rangitopuni Stream is one of the smallest New Zealand rivers in the dataset, having a median flow of 0.4850 m3s−1 with extreme spikes during periods of rainfall. The Wairua River lies in the middle and is a modal representation of the sites in the dataset. Median flow, 90th percentile flow and the ratio of the two are presented to summarise how the flow regimes differ between the three rivers. Table 8Summary of sites used for testing of WRTDS.RiverCoordinates (WGS84)Period of measurementMedian daily flow flowmed (m3s−1)90th percentile flow flow90 (m3s−1)flow90/flowmedWaimakaririLat: −43.35900Lon: 172.05121989-02-01 - 2020-03-04100.0224.462.245WairuaLat: −35.65192Lon: 174.15081989-01-31 - 2020-03-317.18042.2485.884RangitopuniLat: −36.74428Lon: 174.61791989-01-25 - 2020-03-110.48503.18626.569 Summary of sites used for testing of WRTDS. Lat: −43.35900 Lon: 172.0512 Lat: −35.65192 Lon: 174.1508 Lat: −36.74428 Lon: 174.6179 There are three hyperparameters which can be adjusted in the EGRET implementation of WRTDS: windowQ, windowY and windowS which are the half-widths of the weighting window for streamflow, time and seasonality respectively. In simple terms, they represent the width of the window in each of the three dimensions used to assign weights to the input data. The default values used by EGRET are 2 log units for windowQ, 7 years for windowY and 0.5 years for windowS. The original WRTDS paper notes that any choice of window widths is largely arbitrary and is a potential topic for further research22, although there has been no development in this area at the time of writing. Exploratory testing of the model parameters found that windowQ was the only parameter which had a large effect on model performance and convergence. We needed to determine what the minimum length of time series used for prediction should be. The data from each of the three rivers were divided into shorter segments from 3 to 20 years long over the period from 2000-01-01 to 2019-12-31, ignoring the last segment if they did not divide into 20 exactly. Then, WRTDS was fitted to each of these smaller segments to estimate TN concentrations, the outputs of which were then pieced together to form the full series. This was then compared to the output of running WRTDS on the full 20-year period. A visual explanation of the testing method and results can be seen in Fig. 6. Explanation of methodology and results for WRTDS testing on estimated total nitrogen concentrations over time for the Rangitopuni Creek. The width and height of the blue box indicate the length of the time segments and the windowQ value used, respectively. For the Waimakariri River, the predictions for the two models were generally close for segments of more than 10 years, but many of the segments shorter than 10 years contained 1-2 predictions which were hundreds or thousands of times larger than the estimates from the full series (Table 9). Similar trends are seen for the smaller Wairua River, although the occasional spikes were smaller than for the Waimakariri (Table 9). Also, for the three- and four-year segments, there were 20–30 days over the 20-year period for which the model failed to converge, meaning no estimates were made. For the Rangitopuni Stream, results were poor and contained extreme (>1010) spikes all the way up to the 20-year segment (Table 9). Every segment length had days which failed to converge, usually more than 50. Overall, the results were passable for the two larger rivers when the segments were more than 10 years long, but poor for the small stream. Table 9Differences in TN concentration estimates using default WRTDS hyperparameters between the full 20-year period and the spliced segmented periods for the Waimakariri River, the Wairua River and the Rangitopuni Stream.RiverLength of segments (years)Mean difference (mg L−1)Median difference (mg L−1)Maximum absolute difference (mg L−1)Non-convergence days (count)Waimakariri38.38 × 10−2−1.05 × 10−35.26 × 1020Waimakariri45.19 × 108−1.16 × 10−32.28 × 10120Waimakariri52.15 × 108−1.15 × 10−31.57 × 10120Waimakariri18−1.71 × 10−4−3.83 × 10−72.43 × 10−10Waimakariri194.37 × 10−4<10−151.90 × 1000Waimakariri20−3.80 × 10−8<10−152.77 × 10−40Wairua3 > 1015−2.60 × 10−3 > 101526Wairua42.95 × 10−3−7.72 × 10−41.87 × 10124Wairua57.73 × 10−4−1.13 × 10−31.17 × 1000Wairua18−1.75 × 10−3−4.36 × 10−62.62 × 10−10Wairua196.24 × 10−44.06 × 10−67.84 × 10−20Wairua20<10−15<10−15<10−150Rangitopuni36.05 × 100−6.97 × 10−43.64 × 104236Rangitopuni4−5.39 × 10−4−1.04 × 10−33.18 × 100213Rangitopuni55.90 × 101−7.41 × 10−42.28 × 105120Rangitopuni184.80 × 10−4<10−151.70 × 10074Rangitopuni191.90 × 1042<10−151.30 × 104677Rangitopuni201.24 × 10−1<10−151.08 × 10278Differences greater than 0.1 mg L−1 are bolded. Positive values indicate that the segmented estimate is larger than the 20-year estimate. Differences in TN concentration estimates using default WRTDS hyperparameters between the full 20-year period and the spliced segmented periods for the Waimakariri River, the Wairua River and the Rangitopuni Stream. Differences greater than 0.1 mg L−1 are bolded. Positive values indicate that the segmented estimate is larger than the 20-year estimate. Upon experimenting with the hyperparameters, it was discovered that increasing windowQ (the streamflow window width) reduced the frequency of the extreme spikes in the predictions and improved the success rate and similarity to the full series for the shorter segments. Altering the other two hyperparameters did not make any noticeable changes. Using a windowQ value of 7 log units (input streamflow values must be within e7≈1100 times larger or smaller than the streamflow at estimation to have a weight above 0) with the same segment method while keeping windowQ for the full series as 2, yielded quite different results for the three rivers. For the Waimakariri River, there were none of the spikes which had been seen before even in the shortest segments (Table 10). Also, mean and median errors were generally lower than with the smaller window. The same occurred for the Wairua River, which also saw that the days which previously failed to converge for the short segments were now successful (Table 10). There were also no more sudden spikes for the Rangitopuni Stream, although the same number of days with failed convergence was still around the same (Table 10). Upon further investigation, it was apparent that these days were the ones with sudden spiked in streamflow in the input data, with values sometimes spiking to over 100 times the median. It seems that there are not enough data points which are within the moving window for streamflow (windowQ) for these days to resolve, particularly when they are close to the start and end of the series. Table 10Differences in TN concentration estimates using WRTDS with a windowQ value of 7 between the full 20-year period and the spliced segmented periods for the Waimakariri River, the Wairua River and the Rangitopuni Stream.RiverLength of segments (years)Mean difference (mg L−1)Median difference (mg L−1)Maximum absolute difference (mg L−1)Non-convergence days (count)Waimakariri3−1.16 × 10−4−1.31 × 10−37.52 × 10−10Waimakariri4−7.64 × 10−4−1.29 × 10−37.62 × 10−10Waimakariri5−8.71 × 10−4−1.46 × 10−37.34 × 10−10Waimakariri18−8.27 × 10−4−6.52 × 10−57.58 × 10−10Waimakariri19−5.53 × 10−45.12 × 10−57.55 × 10−10Waimakariri20−6.81 × 10−4−1.17 × 10−47.55 × 10−10Wairua3−3.36 × 10−3−3.07 × 10−31.05 × 1000Wairua4−1.41 × 10−3−1.44 × 10−48.89 × 10−10Wairua5−2.79 × 10−3−7.87 × 10−46.85 × 10−10Wairua18−4.18 × 10−3−6.82 × 10−46.72 × 10−10Wairua19−2.11 × 10−37.13 × 10−46.72 × 10−10Wairua20−2.66 × 10−3−4.78 × 10−56.72 × 10−10Rangitopuni3−4.77 × 10−3−6.13 × 10−45.89 × 10−173Rangitopuni4−4.48 × 10−3−7.39 × 10−45.64 × 10−178Rangitopuni5−4.60 × 10−3−5.11 × 10−46.15 × 10−178Rangitopuni18−4.41 × 10−36.97 × 10−45.84 × 10−173Rangitopuni19−4.33 × 10−37.46 × 10−45.84 × 10−177Rangitopuni20−4.36 × 10−39.12 × 10−45.84 × 10−178Differences greater than 0.1 mg L−1 are bolded. Positive values indicate that the segmented estimate is larger than the 20-year estimate. Differences in TN concentration estimates using WRTDS with a windowQ value of 7 between the full 20-year period and the spliced segmented periods for the Waimakariri River, the Wairua River and the Rangitopuni Stream. Differences greater than 0.1 mg L−1 are bolded. Positive values indicate that the segmented estimate is larger than the 20-year estimate. Another note for concern is that for all three rivers, increasing windowQ to 7 seemed to cause the model to underfit when compared to the smaller window as the variation in predicted concentrations decreased substantially, and high flow days often became indistinguishable from the norm (Fig. 6). However, increasing windowQ clearly has its benefits as it allows for shorter series to be modelled more accurately for large and medium rivers. Also, the model fitting process is prone to fail if the time series is shorter than three years, and that increasing the windowQ parameter (the half-window width for streamflow) makes the model more likely to succeed for shorter series. Hence, the decision was made to scale windowQ according to Eq. (2) so series of less than 10 years of length use a wider window to improve the rate of convergence of the model. WRTDS Performance Standard errors for each daily concentration estimate were produced as a part of the WRTDS fitting process via the EGRET R package. These can be found alongside the estimates in the TN/TP.4_-_WRTDS_Output folders and means of the standard errors for each site can be found in the metadata files TN/TP.4_meta.csv. Means were also calculated for the standard errors on a continental basis (Table 11). Table 11Means of the standard errors for the WRTDS concentration estimations across the continents.Mean of TN standard errors (log mg L−1)Mean of TP standard errors (log mg L−1)Africa0.381 (16)0.509 (14)Asia0.309 (9)0.500 (21)Europe0.227 (330)0.410 (1,062)North America0.322 (357)0.528 (462)Oceania0.325 (580)0.485 (730)South AmericaNA0.575 (13)Global0.300 (1,292)0.460 (2,302)Numbers of sites are shown in parentheses. Means of the standard errors for the WRTDS concentration estimations across the continents. Numbers of sites are shown in parentheses. The standard errors for TN were consistently lower than for TP across all continents regardless of the representativeness of each continent. This may be because TP concentrations are tied to the mobilisation of sediment during periods of high precipitation which in general are not as well represented by weekly or monthly measurements compared the diffuse processes behind TN flux. Concentration estimates were generally made with higher standard errors in Africa, Asia, South America and North America, although the first three continents listed are underrepresented in the dataset. While other metrics like root mean squared error would usually also be investigated, the Kalman filtering applied to the estimates results in all the observed measurements being expressed exactly in the estimates, so the error is zero. Comparison of results with existing literature To assess the validity of the results of the nutrient load and concentration modelling, the estimated loads and concentrations were compared to other published literature. Results were compared across the entire global dataset, as well as regionally by considering only the sites contained within the Mississippi River System in the United States. This location was chosen as there have been several studies which simulate nutrients fluxes across the Mississippi Basin69–71. The coefficients of determination for our models were similar to or better than those reported in other studies with varying methodology (Table 12). Table 12Comparison of coefficients of determination between our model and those in similar studies.ParameterRandom forestsMcDowell, et al.23Beusen, et al.85Mayorga, et al.86 (Dissolved inorganic N and P)Stackpoole, et al.70Alvarez-Cobelas, et al.87TN Load/Yield0.710.590.540.64TP Load/Yield0.600.360.510.840.73TN Concentration0.870.59TP Concentration0.480.58 Comparison of coefficients of determination between our model and those in similar studies. The distribution of yield estimates from our models were compared against those of McDowell, et al.23 and Beusen, et al.24 to assess the validity of our results. McDowell, et al.23 used a seven-parameter version72 of the equation used by WRTDS (Eq. 1) to estimate concentrations, but without the moving window approach of WRTDS. Concentrations were then used to calculate yields, which were subsequently modelled based on a variety of land characteristics and predicted for the whole world (nTN = 787, nTP = 1,236). Beusen, et al.24 took a different approach, using IMAGE-GNM, a nutrient model simulating the inputs, retention, and outputs of riverine nutrients to estimate TN and TP loads at a 0.5 by 0.5-degree resolution. The values estimated for 2015 were used for comparison. We converted these gridded estimates to catchment estimates by taking the mean of the IMAGE-GNM load raster across the Level 7 HydroBASINs catchment boundaries and dividing by the upstream catchment area (nTN = 57,450, nTP = 57,450). Our models, which have the benefit of a greater number of sites and spatial representativeness of source data, performed similarly or better to the other global models. Nutrient predictions for the Mississippi River System were also assessed to determine how results compare to estimates produced by models fit locally on a smaller scale. Again, two studies were compared with our estimates. Stackpoole, et al.70 also used IMAGE-GNM to predict loads, this time only across the Mississippi Basin. Yields were calculated by dividing the load estimates by the cumulative catchment sizes provided by the study (nTN = 1,379,904, nTP = 1,379,904). Krasovich, et al.69, unlike the other studies, is a dataset containing only measured data rather than simulated estimates. However, the concentration data is not a full daily series of measurements, so annual loads and yields cannot be easily calculated. For this reason, the mean concentrations for each of the sites provided (nTN = 24,688, nTP = 61,743) were compared with the concentration estimates from our model rather than loads. Our models had comparable results with regional models despite being fit at a larger scale which improves our confidence in the validity of our results when used at a regional level. We then compared the total continental and global nutrient loads between models. Continental nutrient loads refer to the total nutrient flux from rivers to the ocean. The simplest way to calculate these loads was to take the sum of the loads measured at every river outlet to the ocean in the continent. We could approximate this well using Level 7 HydroBASINs as each inland basin generally contained only one river network. Hence, by summing our load estimates in only the coastal catchments in each continent, we could estimate continental loads well (Table 11). Furthermore, we produced a measure of the uncertainty around these estimates using the confidence intervals calculated as a part of the modelling process. The 10th and 90th percentiles for the load estimates for each of the coastal basins were added together to calculate a lower and upper bound representing the uncertainty for the continental and global load estimates (Note that these are not equivalent to 80% confidence intervals for the continental estimates but still serves as a measure of the uncertainty around the estimates). These can be seen along with the continental loads in Table 13. Table 13Estimates of continental nutrient loads.ContinentTN random forest estimates (Tg yr−1)Beusen, et al.24 TN (Tg yr−1)McDowell, et al.23 TN (Tg yr−1)TP random forest estimates (Tg yr−1)Beusen, et al.24 TP (Tg yr−1)McDowell, et al.23 TP (Tg yr−1)Africa2.52 (1.87–3.44)2.472.280.30 (0.21–0.43)0.290.56Asia5.64 (4.16–7.72)18.574.110.49 (0.34–0.73)1.701.05Europe4.77 (3.61–6.36)7.901.260.38 (0.26–0.56)0.980.35North America4.29 (3.19–5.81)3.562.510.42 (0.30–0.60)0.340.63Oceania4.56 (3.31–6.32)2.490.540.44 (0.33–0.61)0.380.13South America3.14 (2.31–4.33)14.171.830.34 (0.24–0.48)1.440.45Global24.92 (18.44–33.99)43.1112.542.36 (1.68–3.42)5.123.16The numbers in parentheses next to out load estimates are an uncertainty interval calculated form the confidence intervals for each catchment. Estimates of continental nutrient loads. The numbers in parentheses next to out load estimates are an uncertainty interval calculated form the confidence intervals for each catchment. Beusen, et al.24 provides an estimate of global nutrient loads, however these are supplied as 0.5 by 0.5-degree raster grids. We used these to calculate continental loads by taking the mean of the gridded estimates across each Level 7 basin, the summing the coastal catchments in the same way as our estimates. This method lines up closely with the global values given in the original paper of 43 Tg yr−1 for TN and 5 Tg yr−1 for TN, although continental values are not given. These load estimates are compared with estimates published in McDowell, et al.23 in Table 13. The main differences in the estimates from our models compared to the other two studies were that our TP load for Asia was lower than the other studies by a large margin and our TN estimate for Oceania was larger than the others. All other estimates were in similar ranges to the those produced by the other two models, many of which were inside our uncertainty bounds. This increased our confidence in the reliability of our load estimate data we present. Our models have similar or better coefficients of variation than those from other studies, and other published results mostly fall within our global uncertainty bounds. Usage Notes The three main data outputs composing the dataset are intended to be useful for a variety of types of hydrological analysis and management. The first output, the harmonised site data from the seventeen source databases, is designed to be a useful starting point for global or regional hydrological studies. All data contained is observational rather than being simulated, and users are free to either utilise the entire dataset as a whole or filter by options such as parameter (TN, TP and flow), location or period of measurement using the corresponding metadata. By combining and harmonising several large databases, we hope to streamline the data sourcing process for future studies. A dataset of this size would be ideal for identifying long-term trends in river nutrient levels, which may be of interest for researchers as a topic for future study. The second data outcome if the output from running WRTDS on the sites with consistent water quality and quantity measurements over an extended period (at least five years). To the authors’ knowledge there are currently no other datasets of this size composed daily concentration estimates corresponding to in-stream measuring stations rather than as global gridded estimates. This provides a unique opportunity for researchers to study phenomena which require high-frequency data like changes in water quality during extreme weather events either on a site-by-site basis, or regionally. Due to the experimental nature of the application of WRTDS, users may wish to filter the data to only include data with an extended period of data collection (e.g. <10 years) or by mean standard error to only include the most reliable estimates. Some shorter series may be missing some individual daily TN or TP estimates due to failed convergence of WRTDS for those days, although this is very uncommon across the dataset. Additionally, the spatial coverage of this stage was limited by data availability so only North America, Europe, Australia and New Zealand are well covered. The global concentration and load estimates produced by the random forests may be useful for identifying areas with poor water quality which may be at risk of eutrophication. Either the concentration estimates could be used to highlight areas with high in-stream concentrations of TN or TP, or the load estimates could indicate where nutrients may build up in sinks like lakes or be transported into the ocean in high volumes. This may be of interest to local river management policymakers or other freshwater scientists. All the data along with associated metadata and R scripts used for analysis are available for access from the figshare site (https://doi.org/10.25400/lincolnuninz.30546371)65, with the exception of the data from the UK National River Flow Archive40 and some of the sites from the GEMStat18 and GRDC17 databases which are not able to be included due to their respective licensing agreements. These sites along with their associated site IDs are listed in the metadata files in the data folder and can be accessed from the original data repositories. Lists of publishable sites for GEMStat and GRDC were sourced from the UNEP GEMS/Water Global Freshwater Quality Archive66 and GRDC-Caravan67 respectively. The sites which are included in the data repository {Peters, 2026 #482} can be seen in Fig. 7. Maps of the sites which are included in the data repository. Supplementary information Supplementary Information Supplementary Information Author contributions J.P. collated and analysed the data and wrote the manuscript; R.W.M. initiated the study, secured funding, provided supervision and co-wrote the manuscript; A.N. and C.D. provided co-supervision and co-wrote the manuscript. Funding The work was not funded by any external sources. Code availability All R code used for data extraction, filtering, WRTDS application and modelling can be found along with the data from https://doi.org/10.25400/lincolnuninz.3054637165. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Supplementary information The online version contains supplementary material available at https://doi.org/10.1038/s41597-026-07625-1. References Associated Data Supplementary Materials Supplementary Information Supplementary Information Data Availability Statement All R code used for data extraction, filtering, WRTDS application and modelling can be found along with the data from https://doi.org/10.25400/lincolnuninz.3054637165.