Extent, patterns, and drivers of hypoxia in the world's streams and rivers

Hypoxia in coastal waters and lakes is widely recognized as a detrimental environmental issue, yet we lack a comparable understanding of hypoxia in rivers. We investigated controls on hypoxia using 118 million paired observations of dissolved oxygen (DO) concentration and water temperature in over 125,000 locations in rivers from 93 countries. We found hypoxia (DO < 2 mg L−1) in 12.6% of all river sites across 53 countries, but no consistent trend in prevalence since 1950. High‐frequency data reveal a 3‐h median duration of hypoxic events which are most likely to initiate at night. River attributes were better predictors of riverine hypoxia occurrence than watershed land cover, topography, and climate characteristics. Hypoxia was more likely to occur in warmer, smaller, and lower‐gradient rivers, particularly those draining urban or wetland land cover. Our findings suggest that riverine hypoxia and the resulting impacts on ecosystems may be more pervasive than previously assumed.

than watershed land cover, topography, and climate characteristics. Hypoxia was more likely to occur in warmer, smaller, and lower-gradient rivers, particularly those draining urban or wetland land cover. Our findings suggest that riverine hypoxia and the resulting impacts on ecosystems may be more pervasive than previously assumed.
Oxygen is fundamental to sustaining life on Earth by supporting aerobic respiration and the conversion of stored chemical energy into biomass and work. Since oxygen from the atmosphere diffuses slowly through water, aquatic ecosystems are particularly vulnerable to low dissolved oxygen (DO) levels. Hypoxia-the depletion of DO below levels known to be harmful to fish and other aquatic aerobic life (Saari et al. 2018;Croijmans et al. 2021)-has been extensively reported in lake (Jenny et al. 2016;Jane et al. 2021), coastal (Diaz and Rosenberg 2008), and marine (Rabalais et al. 2010;Breitburg et al. 2018) ecosystems globally over the last half century. Although hypoxia has been previously documented in rivers, we lack a comparable understanding of its prevalence and dynamics in inland rivers and streams at a global scale relative to lakes and coastal ecosystems. Recent publicized fish kills associated with hypoxic conditions in rivers ( Fig. S1; Table S1) suggest riverine hypoxia, and its attendant impacts on ecosystems and people, may be more common than previously assumed.
Rates of photosynthesis and organic matter mineralization (i.e., respiration) are important biological controls on aquatic oxygen dynamics (Garvey et al. 2007;Piatka et al. 2021). Diel DO variation in the photic-zone of water columns is welldocumented and arises from daytime photosynthetic activity by autotrophs (e.g., algae, cyanobacteria, submerged plants). This time-varying production of DO occurs simultaneously with sustained DO consumption by heterotrophic and autotrophic respiration and continuous oxygen exchange between the water and atmosphere to control patterns in DO concentrations (Odum 1956). When respiratory consumption of DO outpaces physical reoxygenation from the atmosphere, usually at night, DO declines until daytime production starts again at sunrise (Hornberger and Kelly 1975). Oxygen production is largely controlled by light and flow disturbances (Kirk et al. 2021;Bernhardt et al. 2022), while consumption is determined by temperature and organic matter supply to heterotrophic organisms (Streeter and Phelps 1925;Arroita et al. 2019). Rivers draining both natural and humandominated watersheds can receive large inputs of organic matter, making rivers well known landscape control points for exceptionally rapid oxygen consumption (Webster and Meyer 1997;Hotchkiss et al. 2015).
Although the same biochemical processes of photosynthesis and respiration govern oxygen supply and demand in rivers as in lakes and coastal ecosystems, gas exchange rates are often much higher in flowing waters (10-1000 m d À1 ; Hall and Ulseth 2020) than in the quiescent waters of lakes ($ 1.5 m d À1 ; Winslow et al. 2016) and estuaries (1-3 m d À1 ; Rosentreter et al. 2017). The exchange of oxygen between water and air is determined by departure from atmospheric equilibrium, mediated by temperature-dependent diffusivity and DO solubility as well as the gas transfer velocity, which depends on turbulent dissipation at the air-water interface that changes with channel slope (Hall and Ulseth 2020). While reaeration can be rapid in turbulent streams (Ulseth et al. 2019) and thus limit the potential for hypoxia, low gradient rivers with less rapid gas exchange (1-10 m d À1 ) are still vulnerable. Higher water temperatures limit DO solubility, elevating the risk of hypoxia in warmer streams (Rajesh and Rehana 2022), while the magnitude and directionality of hydrologic exchange with the hyporheic zone can amplify or dampen the influence of benthic oxygen consumption on water column DO (Grimm and Fisher 1984). Temporary stratification of the water column in relatively deep and slower moving streams and rivers is not unusual (Siders et al. 2017), yet persistent stratification in flowing waters is less common unless tidally influenced (Mackay and Fleming 1969;Nelson et al. 2017). Therefore, it is commonly assumed that turbulent mixing in streams and rivers limits the occurrence of hypoxia in comparison with other aquatic ecosystems.
Hypoxia in rivers can manifest as various contrasting regimes, ranging in duration, timing, and severity. Hypoxic regimes are relevant to management but understudied, despite the emergent utility of sensors to document hypoxia (Carter et al. 2021). Episodic hypoxia during flow stagnation is characterized by low DO at extremely low flows, during which reoxygenation via gas exchange, advection, or primary productivity is limited. This regime is exacerbated by anthropogenic water abstraction (Pardo and García 2016) and stream channel modification (Blaszczak et al. 2019), but can occur during droughts even in undeveloped landscapes (G omez-Gener et al. 2020). Episodic hypoxia also occurs following pulses of organic matter from anthropogenic waste or natural storage (e.g., hippopotamus waste [Dutton et al. 2018] or organic-rich blackwaters [Hladyz et al. 2011]). Nighttime hypoxia is characterized by large diel swings in DO caused by high daytime primary production and attendant elevated respiration rates that consume oxygen at night faster than it can be resupplied causing repeated oxic-hypoxic cycles (e.g., downstream of sewage inputs [Streeter and Phelps 1925] or nonpoint-source agricultural inputs [Heiskary et al. 2013]). Finally, persistent hypoxia can occur in sites dominated by high groundwater contributions of low-DO water (e.g., seasonal drainage from wetlands; Hamilton et al. 1997). No systematic analysis of the frequency and drivers of different hypoxic regimes exists at the global scale.
To evaluate the spatial extent and controls of riverine hypoxia and hypoxic regimes, we assembled an extensive global dataset of riverine DO concentration and associated physical and chemical characteristics-over 118 million paired observations of DO concentration and water temperature from over 125,000 river locations across the globe. We used this database to (1) quantify global patterns in riverine DO over space and time, (2) document spatial variation and drivers of different hypoxia regimes, and (3) evaluate local river and watershed characteristics as predictors of hypoxia.

Global riverine DO database
We compiled data from publicly available sources including government-supported data portals (e.g., Water Quality Portal; Read et al. 2017) and published data to create a riverine DO dataset that spans all continents except Antarctica but is dominated by data from North America . For each location, we required a geographic coordinate, at least one measurement of DO in units of percent saturation or mg L À1 , a corresponding water temperature, and a date-time stamp. We removed any data (1) marked as questionable if data quality notes were available, (2) identified as before 1 January 1900 or after 1 January 2020, and (3) with DO measurements less than À1 mg L À1 or above 35 mg L À1 . DO measurements between À1 and 0 mg L À1 were set to 0 mg L À1 to include low-DO measurements that fell within a reasonable range of calibration error. We paired all locations in the database with the HydroATLAS database (Linke et al. 2019) for estimated stream channel slope, contributing catchment land cover, and other potential covariates (Table S3) and paired locations with MODIS Land Cover (Friedl and Sulla-Menashe 2019) from Google Earth Engine to describe land cover surrounding each site. For sites within the continental United States, we also joined each unique DO sampling location to its nearest National Hydrography Dataset flowline segment (i.e., COMID) to link locations to the StreamCat database (Hill et al. 2016). Refer to the Supporting Information for details.
We assessed the distribution, frequency, and global extent of riverine hypoxia using 118,369,376 paired DO and water temperature measurements from 125,158 unique locations in 93 countries and territories across six continents (Figs. 1, S2; Blaszczak et al. 2021). Measurement locations in the United States accounted for 89.7% of the database (see Table S2 for breakdown by country). The longest periods of record are from the United States, including Puerto Rico, followed by Canada, Germany, and the Netherlands, with newer measurements increasing the number of countries from 10 in 1975 to 93 in 2018 (Fig. S3). We use the widely adopted threshold of < 2 mg DO L À1 to define hypoxia (CENR 2010), while noting that negative effects on aquatic organisms are observed as high as 5 mg DO L À1 (Saari et al. 2018). For analyses necessitating high-frequency data, we use a subset of 587 sites with highfrequency DO data (≤ 60 min time step between records) from 2007 to 2017 previously published in Appling et al. (2018), which we refer to as the "USGS High-Frequency Subset."

Temporal trends in riverine hypoxia
We examined temporal trends in riverine hypoxia by calculating the proportion of hypoxic observations over time across all observations and the length of continuous hypoxic events observed in the USGS High-Frequency Subset. We quantified trends in the proportion of hypoxic observations through time by fitting a Sen's slope using the "sen.slope" function in the "trend" package (Pohlert 2020) in R with a 95% confidence level.

Probability of riverine hypoxia
We evaluated relationships between the proportion of hypoxic measurements for each location and hypothesized drivers, including within-river and surrounding land-use attributes. To evaluate the relative importance of climatic, geomorphic, land-use, and in-stream variables for predicting the occurrence of hypoxia (Table S3), we fit boosted regression tree (BRT) models to the full global dataset (Elith et al. 2008). All analyses were performed in R (R Core Team 2021).
To further evaluate in-stream controls, we bootstrapped (n = 5000) two logistic regression models to predict the occurrence of hypoxia at a river location based on characteristics of the river and surrounding land use. In the first model, we predicted the occurrence of hypoxia at a location based on the static properties of slope and river size (Strahler order), and the site mean water temperature (n = 66,470). We predicted the probability of hypoxia across a range of water temperatures (2-30 C) using each bootstrapped combination of coefficients in groups determined by slope and river size. We used an inverse logit function to convert the parameter estimates into probabilities. In the second model, we predicted the occurrence of hypoxia at a location based on the surrounding land cover type using the same set of data as the first model (n = 66,470). Finally, we used logistic regression to evaluate the relationship between the occurrence of hypoxia in a river and watershed population density (StreamCat 2010 watershed population density [people km À2 ]) and estimated the probability of hypoxia across the most probable range (2.5-97.5%) of watershed population densities (n = 94,176; 0.23-1060 [people km À2 ]).
We further explored the joint roles of biological consumption (i.e., net ecosystem production [NEP]) and physical replenishment (i.e., gas transfer velocities) of DO, by adapting a widely used oxygen mass-balance model by Odum (1956) and estimating gas exchange using empirical equations from Raymond et al. (2012). We used a t-test to compare whether the mean NEP values of the distributions of sites with and without detected riverine hypoxia were different. We calculated thresholds past which hypoxic measurements are expected (low gas exchange, high ER) as a function of oxygen saturation (Garcia and Gordon 1992) across a range of temperatures, assuming constant barometric pressure of 760 mmHg. We then evaluated how well persistent or frequent hypoxia was captured by mean site behaviors.

Results and discussion
We detected hypoxia (<2 mg DO L À1 ) in rivers across 53 countries (Fig. 1A), with 12.6% of locations exhibiting at least one hypoxic measurement (n = 15,818) and 35% exhibiting at least one measurement below 5 mg DO L À1 (n = 43,823). A similar fraction of sites within the United States exhibited hypoxia (14,334 locations, 12.8% of U.S. sites), while Canada (379 locations, 24.6%) and Brazil (307 locations, 35.0%) had the greatest number of river locations with detected hypoxia by country outside of the United States. Locations with high-frequency sensor measurements (≤ 60 min time step between records) accounted for < 1% of locations but 95% of the total database observations, thus the median number of measurements per location was nine. Of all locations globally with at least 10 DO measurements (49% of the database), 5.6% (n = 7018) were hypoxic for 5% or more of the time series.
It is likely our global assessment under predicts the frequency of hypoxia in rivers and that prediction uncertainty is greatest outside North America given the geographic bias in the data. Until recent advances in field-deployable DO sensor Within-site proportion of DO observations that are < 2 mg DO L À1 (left y-axis) and within-site standard deviation of DO concentration (right y-axis; gray points) for sites with at least three measurements (n = 84,010; 67% of all sites). Sixty-two sites with mean DO concentration > 15 mg L À1 were excluded from this plot. Loess (locally estimated scatterplot smoothing) lines (black for the proportion of observations that are hypoxic; gray for the standard deviation of DO concentration) have a span of 1% of the dataset. The marginal distribution of within-site mean and standard deviation (SD) of DO concentration are shown at the top in black and the right side of the plot in gray, respectively. technology, DO was typically measured manually at frequencies of once per day or less (Fig. 2A). Manual DO measurements are typically made during the day, yet hypoxia is more likely to occur at night when photosynthetic oxygen production ceases. For example, 90% of water samples in a global river chemistry database (GLORICH; Hartmann et al. 2019) were collected between 08:10 h and 15:55 h with a median sampling hour of 11:25 h (Fig. S4). This daytime sampling bias underestimates the number of sites exhibiting hypoxia by approximately 25% in the USGS High-Frequency Subset (Fig. 2B). Pervasive underdetection of hypoxia underscores the value of continuous high-frequency data in aquatic ecosystem monitoring programs (Venkiteswaran et al. 2015;Arroita et al. 2019) and the importance of nighttime sampling to capture the full range of biogeochemical conditions (G omez-Gener et al. 2021).
Despite an increase in DO measurements from less than 1000 per year in 1950 to over 10 million per year in 2018 (Fig. 2C), the proportion of observations exhibiting hypoxia has remained static near 3% (Fig. 2D). This proportion reached as high as 5.6% in 2011 if excluding sites with highfrequency measurements. Similarly, the annual proportions of hypoxic measurements within U.S. Environmental Protection Agency regions have not systematically changed between 1970 and 2018 ( Fig. S5; Table S4; statistically significant Sen's slopes < 0.03%). Indeed, only 20 out of 484 sites (4% of sites) in the USGS High-Frequency Subset with data across at least 3 yr had a positive or negative trend in DO (Fig. S6). Our analysis strongly suggests that regions differ in their natural susceptibility to hypoxia necessitating implementation of regional DO standards (Jankowski et al. 2021). Individual hypoxic events in the USGS High-Frequency Subset had a median duration of 3 h (Fig. 3A). Riverine hypoxic events persisting longer than 1 week were rarely detected (n = 700; 2% of all hypoxic events). Six of the 10 locations with the longest duration of hypoxic events were in Florida (Fig. 3B), a state with low gradients, high temperatures, and significant aquifer discharge. Differences in the timing of the onset and end of a hypoxic event were most apparent for events lasting 6-12 h (Fig. 3C). The onset was most frequently between 5:00 h and 8:00 h local time, matching our expectations that DO is often lowest just before sunrise following an extended period without photosynthetic oxygen production. The end of a hypoxic event was most frequently between 13:00 h and 15:00 h when cumulative DO production rates peak under maximal light availability. For longer-duration events, the timing of the beginning and end was less clear (Fig. S7), suggesting other controls on the duration of events (e.g., storms) and illustrating the dynamic nature of DO in rivers beyond diel light and temperature controls (Hensley et al. 2018).
The occurrence of hypoxia in rivers is determined by both instream attributes (e.g., channel slope) as well as watershed controls (e.g., land cover) which interact to determine the biochemical and physical conditions within a channel (Piatka et al. 2021). Our BRT analysis (n = 124,519) revealed that instream attributes including maximum observed water temperature and stream slope were the two strongest predictors of hypoxia occurrence (Fig. 4A). The BRT model had few (n = 16) false positives (i.e., the model rarely predicted hypoxia at a site if there was no evidence of hypoxia), but many (n = 3,790) false negatives (i.e., the model did not predict hypoxia at a site despite evidence of hypoxia; Fig. S8). Persistent or frequent hypoxia was well captured by mean site behaviors of low gas exchange (< 10 m d À1 ) and negative NEP (< À10 g O 2 m À2 d À1 ), but these two controls alone were insufficient predictors of episodic hypoxic events in rivers (Fig. 5). These results highlight that although global models of river hypoxia occurrence can effectively predict instances of hypoxia that match our understanding of physical and thermal constraints on hypoxia occurrence, the models cannot capture occurrences of hypoxia that might be driven by uncertainty in gas exchange estimates or excess organic matter and nutrient loading, for which additional corresponding data were unavailable. Therefore, the use of regional models is likely to offer a compromise between sitespecific approaches which lack generality, and global approaches that can lack specificity given data sparsity.
To further explore in-stream controls on spatial variation in hypoxia, we examined relationships between DO concentrations and local river characteristics. Across sites paired with Fig. 4. Within-river and landscape controls on the probability of occurrence of river hypoxia. (A) The relative influence of each predictor variable in a boosted regression tree model on the percent of variance in riverine hypoxia occurrence explained (x-axis). Predictor variables were grouped into categories of stream attributes (blue), watershed climate (yellow), watershed land cover (green), and watershed topography (gray). Refer to methods for predictor variable sources and explanations. (B) The probability of riverine hypoxia depends on river size and slope, with both small and flat streams being most vulnerable to hypoxia at higher mean water temperature. Coefficients for each prediction are from a bootstrapped (n = 5000) logistic regression for sites with NHD slope estimates (n = 66,470). Lines are the median predictions with 95% confidence intervals shown for each line. Slope categories were divided into flat (0.05 m km À1 ), medium (0.67 m km À1 ), and steep (9.67 m km À1 ) corresponding to the mean AE one standard deviation of the log 10 transformed and standardized NHD slope. River size groups were based on Strahler order classification and were split into small (orders 1-3), medium (orders 4-6), and large (orders 7-9). (C) The posterior distributions of bootstrapped estimates (n = 5000) of the probability of riverine hypoxia, derived from a logistic regression based on MODIS-based IGBP land cover types immediately surrounding each site. Overlapping distributions (e.g., wetland and urban land use) indicate that the probability of riverine hypoxia is the same. the U.S. National Hydrography Dataset (McKay et al. 2012;n = 66,470), the probability of riverine hypoxia increased with mean water temperature, and this effect was dependent on river slope and size (Fig. 4B). The probability of any stream experiencing hypoxia was greatest in small (Strahler order 1-3), low-relief (slope = 0.05 m km À1 ) streams and increased by 19% AE 2% over a mean water temperature range of 30 C. The probability of hypoxia in large rivers (Strahler order 7-9) was 71% lower than in smaller streams with the same mean slope (0.67 m km À1 ) and temperature. As depth increases, the water column volume increases relative to stream benthic surface area necessitating greater DO consumptive processes to draw down the mass of DO in the water column. Together, our results suggest that rivers with steep slopes (≥ 9.67 m km À1 ) universally create conditions of high gas exchange that limit hypoxia, whereas small, lowland streams are most susceptible to hypoxia owing to limited turbulence to replenish DO that is depleted during downstream transport, as well as high ratios of benthic surface area-to-water column volume (Mallin et al. 2006;Garvey et al. 2007;Shields and Knight 2012).
We expected watershed land cover to affect the probability of riverine hypoxia and found that hypoxia occurs most frequently in streams draining wetland or urban land cover (Fig. 4C). Wetlands and urban areas can both be sources of high organic matter loading that support biochemical oxygen demand sufficient to outpace the physical reoxygenation of water under natural conditions (e.g., blackwater [Hladyz et al. 2011;Whitworth et al. 2012]), or due to human activities (e.g., treated or untreated sewage inputs [Chambers et al. 1997] including nonpoint-source pollution [de Carvalho Aguiar et al. 2011;Blaszczak et al. 2019]). Indeed, the probability of riverine hypoxia increased 6.4% across watershed population densities (0.23-1060 people km À2 ; n = 94,176; p < 0.0001). While in-stream attributes were more important for predicting spatial variation in riverine hypoxia (Fig. 4A), our results highlight the importance of land cover (Fig. 4C) and the strong covariance between watershed and in-stream attributes (e.g., slope and wetland cover).
We found widespread, short-term occurrence of hypoxia in rivers across the globe, with rivers draining both disturbed and undisturbed landscapes vulnerable to hypoxia. Despite being generally ephemeral, hypoxic conditions in rivers are of outsized relevance because of detrimental effects on aquatic organisms (CENR 2010;Sampaio et al. 2021) and implications for greenhouse gas emissions (Venkiteswaran et al. 2014). We did not find clear trends in the prevalence of hypoxia after the enactment of the U.S. Clean Water Act (33 U.S.C. §1251 et seq. 1972), consistent with other studies investigating Fig. 5. The occurrence of river hypoxia across mean ecosystem metabolism and gas exchange rates. Net ecosystem productivity (NEP) estimates between 0 and À50 g O 2 m À2 d À1 across estimated gas transfer velocities (k), assuming steady-state conditions and zero GPP (n = 87,992). Estimates are based on mean water temperature and mean annual discharge by location. Locations with no hypoxic observations are colored in black (n = 76,047) whereas locations with at least one hypoxic observation are colored according to the proportion of observations with [DO] < 2 mg L À1 (n = 11,945). Lines are labeled by water temperature threshold and delineate the hypothetical threshold above which we would expect hypoxic conditions within a site. The marginal distributions of NEP and k across individual sites are shown as histograms on the right side and top of the plot, respectively, and colored according to the detected occurrence of hypoxia. NEP was more negative in rivers with hypoxia (M = À10.3, SD = 6.5) than those with no events (M = À6.5, SD = 7.1; t(16,600) = À59.0; p < 0.001).
trends in river solute concentrations within the United States (Stets et al. 2020) but inconsistent with declining biochemical oxygen demand trends in European rivers (EEA 2018). Yet, historical DO datasets certainly underrepresent the prevalence of hypoxia in rivers, given the geographic bias in measurements and methodological improvements offered by newer sensor technology (Table S6). Ensuring the continuation of publicly available data and further developing our predictive capacity to identify when and where rivers are vulnerable to the development of hypoxic conditions can guide river management in the face of continuing climate and land-use change.