Sustained deoxygenation in global flowing waters under climate warming

Abstract

Dissolved oxygen (DO), as a vital material sustaining aquatic ecosystems, has declined markedly in oceans, lakes, and coastal waters, yet unbiased understandings of changing DO concentrations in each individual river segment globally remain a challenge. Here, we estimate DO concentrations in 21,439 rivers globally between 1985 and 2023, based on Landsat observations and climatic data, and examine their patterns and trends. We find sustained deoxygenation in global rivers, at a rate of −0.045 mg liter⁻¹ decade⁻¹, with 78.8% experiencing fluvial deoxygenation, driven mainly by oxygen solubility and temperature. Moreover, short-term heatwaves and dam impoundment exert non-neglecting influence on these changes. Future projections demonstrate that global fluvial DO concentrations decline by 1.1% ± 1.6% under SSP1–2.6 and 4.7% ± 2.7% under SSP5–8.5 throughout the 21st century. Our study provides an unbiased baseline for escalating deoxygenation in global fluvial ecosystems that underscores targeted measures to mitigate deoxygenation threats and protect ecosystem health.

Sustained deoxygenation in global rivers primarily results from climate warming.

INTRODUCTION

The concentration of dissolved oxygen (DO) is fundamental to fluvial ecosystem health and services, as it can sustain ecosystem functioning, regulate biogeochemical cycle, protect aquatic biodiversity, and measure drinking water quality (1–4). Fluvial DO concentrations generally depend on covarying physical, biological, and chemical processes that encompass atmospheric oxygen dissolution, ecosystem metabolism, and microbial decomposition (5). Oxygen dissolution, defined by oxygen solubility (OS), depends on water temperature, air pressure, and salinity (6). As temperature increases, OS will decline and thus accelerate deoxygenation (7, 8). Ecosystem metabolism (photosynthesis and respiration that sustain fluvial ecosystems) varies with multi-drivers, such as temperature, light and flow regimes, land use, and nutrients (3, 9, 10). In photosynthesis-dominated rivers, aquatic vegetation and phytoplankton can produce oxygen to enhance DO levels and even supersaturate DO in surface waters (11). Conversely, when respiration exceeds photosynthesis, fluvial ecosystems will consume oxygen and be vulnerable to hypoxia (<2 mg liter⁻¹), which severely threatens well-oxygenated habitats and survival conditions for aquatic species and organisms (12). Moreover, microbial decomposition, as an oxygen-consuming process, can rapidly deplete oxygen and thereby form anoxic conditions, named as dead zones (<0.5 mg liter⁻¹), when massive phytoplankton biomass decays (13), leading to mass fish kills, water quality degradation, and ecosystem service loss (14, 15). However, little is known about how these processes regulate fluvial DO concentrations on a global basis.

Over the past half-century, widespread and frequent deoxygenation has been reported in oceans (16, 17), lakes (8, 18, 19), and coastal waters globally (20), in response to both climate change and human activities. Currently, understandings of changing fluvial DO concentrations primarily rely on field monitoring and numerical modeling. Long-term measurements indicate that fluvial DO concentrations in China’s 10 largest rivers markedly increase at a rate of +0.93 mg liter⁻¹ decade⁻¹ between 2006 and 2020, due to the strict nutrient management (21, 22). Decades of field monitoring confirm that flowing waters in Puerto Rico experience DO enhancement since the implementation of the Clean Water Act in 1974 (23). However, numerical modeling reveals widespread deoxygenation in temperate rivers of the US and Central Europe under climate warming (7). Heavily polluted Gange River, India is reported to suffer from unprecedented deoxygenation, with associated rates varying from −0.03 to −0.1 mg liter⁻¹ decade⁻¹ between 2001 and 2015 (24). Unfortunately, these previous studies delineate changes in fluvial DO concentrations in several regions and temperate rivers, challenging to upscale to Arctic and tropical rivers because of complex regulation from amplified climate warming and concomitant nutrient (i.e., nitrogen, phosphorus, and organic carbon) release from thawing permafrost in Arctic (25, 26), and from flow regimes and land use changes in the tropical regions (27, 28). A recent study used a hybrid process–based and machine-learning model to investigate low-oxygen and hypoxic stress for global rivers, and to identify river warming as the top contributor (29), while its spatial resolution of ~10 km seems to be inadequate to well capture changing DO concentrations in individual river segments. Furthermore, the hybrid model does not consider oxygen supply from photosynthesis reactions, possibly leading to systematic biases in some productive rivers and reservoirs. In this regard, global patterns, trends, and variations in fluvial DO concentrations are worthwhile and necessary to be systematically assessed.

Satellite observations are recognized as promising ways to accurately monitor water quality in surface waters, and have been successfully applied to delineate sediment transport, water clarity, and algal bloom extents at global scales (30–32). Here, we used Landsat satellite imagery between 1985 and 2023 to systematically examine changes in fluvial DO concentrations, drawing an unbiased and comprehensive picture of global river deoxygenation risks and addressing three fundamental questions: (i) Where and how have fluvial DO concentrations changed globally? (ii) What have been the primary drivers behind these changes? Whether differ from regions? (iii) How does fluvial DO respond to future climate warming? Systematically understanding these changes is crucial for enhancing the resilience of fluvial ecosystems to sustained deoxygenation risks through targeted measures and strategies, and helps to achieve sustainable management in global rivers.

RESULTS

Establishment of global fluvial DO database

We established a global fluvial DO database (GDOD) to characterize spatial patterns and temporal trends in DO concentrations in 21,439 river reaches (fig. S1A), based on 3.4 million 30-m-resolution Landsat images from 1985 to 2023 and a stacking machine-learning retrieval algorithm. Using an independent dataset to validate the stacking retrieval algorithm yielded high accuracy levels, with an absolute and relative mean error of <0.9 mg liter⁻¹ and 13%, despite a wide range of 0.19 to 19.1 mg liter⁻¹ globally across different sensors. Performance on the training dataset was greatly more satisfactory, with its absolute and relative mean errors below 0.4 mg liter⁻¹ and 6% (fig. S2). Such accuracy levels can capture variations in fluvial DO concentrations spatially and temporally.

Our GDOD demonstrated that the mean DO concentration in a total of 21,439 global river reaches (76.2% and 23.8% in the northern and southern hemisphere, respectively) was 8.31 ± 1.74 mg liter⁻¹ between 1985 and 2023. The prominent latitudinal gradients exhibited fluvial DO concentrations increasing from equatorial to polar regions. Of six continents, Europe and North America showed apparent higher DO concentrations, reaching 10.07 ± 0.42 mg liter⁻¹ and 9.78 ± 1.09 mg liter⁻¹, respectively, while the lowest level characterized South America and Africa (<7 mg liter⁻¹). The DO concentrations in high-altitude rivers tend to be higher than those in the same latitude plain rivers, such as the Tibetan Plateau (Fig. 1A and fig. S3).

Fig. 1. Spatial and temporal patterns in global fluvial DO concentrations.

(A) Spatial distributions of DO concentrations in global flowing waters, where the inset boxplot represents variations among the different continents. (B) Global patterns in temporal trends of fluvial DO concentrations between 1985 and 2023, with an inset panel showing comparison of river segments with increasing and decreasing trends in each continent. (C) Long-term changes in the fluvial DO concentrations at the global scale. The dashed line and shade represent the linear regression and 95% confidence intervals. (D) Latitudinal profiles of the fractions of river segments experiencing different trends: decreasing and increasing trends, as well as corresponding significant trends.

Four decades of fluvial DO changes globally

Global fluvial DO concentration experienced an overall downward trend, at a rate of −0.045 ± 0.001 mg liter⁻¹ decade⁻¹ [confidence interval (CI) = −0.052 to −0.037, P < 0.001] between 1985 and 2023 (Fig. 1C), nearly equivalent to previously documented −0.042 mg liter⁻¹ decade⁻¹ by a hybrid process–based and machine-learning model (29), and greater than that (−0.038 mg liter⁻¹ decade⁻¹) reported in temperate rivers of the US and Central Europe (7). Of 21,439 studied river reaches, 78.8% of them (7.4 × 10⁵ km) exhibited a decreasing trend, and 33.0% (3.2 × 10⁵ km) were statistically significant (P < 0.05), while only 2.2% (1.7 × 10⁴ km) showed significantly increasing trends (Fig. 1B). Largely spatial variability indicated that the largest decrease in DO concentrations was observed in South America; the fractions of decreasing and significant decreasing trends were 86.4% and 51.5% of total river length, respectively, primarily attributed to pervasive deoxygenation in the Amazon River basin and Brazilian highlands. Moreover, the most pronounced deoxygenation (>0.2 mg liter⁻¹ decade⁻¹) dominated the India rivers. By contrast, DO enhancement collectively occurred in the northern India, Tibetan, and Mongolian Plateau, as well as some rivers in the eastern Siberia and western North America. We also observed considerable difference in fluvial DO trends over latitude where the most notable decrease characterized subtropical and tropical flowing waters (30°S to 30°N), with a fraction of rivers having significant decreasing trends exceeding 60%, and the peak fraction reaching >90% between 20°S and 20°N (Fig. 1D).

Fluvial DO patterns and trends varied with flow conditions globally, with the largest concentration (8.16 mg liter⁻¹) and deoxygenation rate (−0.043 mg liter⁻¹ decade⁻¹) in normal-flow seasons (Fig. 2, A and B). Global deoxygenation rates reduced to −0.040 and −0.035 mg liter⁻¹ decade⁻¹ under high- and low-flow conditions, respectively. Compared with normal flow, low-flow DO concentrations were generally higher in temperate and tropical rivers, especially in China, India, and Pakistan, while the US and western Europe rivers displayed lower concentrations. Conversely, the eastern US and western Europe rivers had higher DO concentrations under high-flow conditions (Fig. 2 and fig. S4C). Globally, most rivers experienced slower deoxygenation in low- and high-flow seasons, and associated hotspots primarily focused on eastern China and southeast Asia. The fraction of river reaches holding significantly downward trends decreased by 6.7% from normal- to high-flow conditions. Unexpectedly, there were some rivers in the northern India showing faster deoxygenation within high-flow seasons (fig. S4D).

Fig. 2. Variations in global fluvial DO concentrations across varying flow conditions.

(A) Probability density of climatological mean DO concentrations in global flowing waters from low- to high-flow seasons. The dashed lines indicate the median values of fluvial DO concentrations under three flow conditions. (B) Variations in long-term changes of global fluvial DO concentrations among different flow regimes. (C and D) Global patterns of climatological mean DO concentrations and their temporal trends during the low-flow season between 1985 and 2023. (E) Fractions of increasing and decreasing trends in fluvial DO concentrations from south to north. (F to K) As in (C) to (E) but for normal- and high-flow seasons.

Short-term heat extremes caused both fluvial DO decline and deoxygenation acceleration (fig. S6). Upon excluding heatwaves, global fluvial DO concentrations increased by 0.04 mg liter⁻¹, with the largest increase reaching up to 0.15 mg liter⁻¹ in Arctic rivers (fig. S6A). However, heatwaves exerted contrasting effects to some tropical rivers and promote DO levels. Over the past four decades, heatwaves contributed 22.7% to global river deoxygenation; long-term trend decreased by 0.01 mg liter⁻¹ decade⁻¹ when removing heatwaves (fig. S6B). Meanwhile, 22.5% of fluvial ecosystems witnessed mitigation of deoxygenation risks by heatwaves; most of them are located in mid- and high-latitude regions, possibly due to enhanced photosynthesis under sudden rising temperature (33).

Dam impoundment also regulates trends in fluvial DO concentrations, indicated by 457 dams constructed between 1995 and 2015 (text S2 and fig. S7). Both before and after dam impoundment, river deoxygenation was dominant, while the proportion of rivers displaying significantly deoxygenation increased by 18.3% (fig. S8). Changes in fluvial DO trends were evident since dam impoundment (fig. S9A) and strongly dependent on reservoir depths (R² = 0.15, P < 0.001; fig. S9B). More interestingly, such dependence was amplified in the warmer extended summer (June to September) (fig. S9, C and D), possibly associated with hydro-dynamics alteration and land-water conversion since reservoir impoundment. Specifically, impounded reservoirs become lentic and vulnerable to climate warming (34), while substantial expansion of impoundment areas by land-water conversion promotes local evaporation and thus generates the cooling effects to the water and atmosphere (34, 35). For large and deep reservoirs, such as Three Gorges Reservoir (TGD), the cooling effects are more likely to offset and even exceed reservoir warming related to altered thermal regimes in lentic waters (36).

Primary drivers of global fluvial DO trends

Pervasive deoxygenation in global flowing waters was undoubtedly associated with changing climate (2, 7). As global warming, declining OS was dominant (fig. S10B). Partial correlation analysis revealed that OS was the most primary driver of river deoxygenation and contributed mostly (62.7%) when excluding the covariate effects from other drivers, that is, temperature, shortwave radiation, discharge, and wind speed (Fig. 3A). These contributions were markedly greater in temperate and frigid rivers, reaching 84.8% and 68.8%, respectively. Furthermore, the correlation between fluvial DO concentrations and OS nearly exceeded 0.9 in temperate regions, especially in the US, Europe, and eastern China (Fig. 3B and fig. S11A). DO changes in 8.6% of global rivers were attributed to surface air temperature (SAT). Over half of them were tropical rivers (fig. S11B). SAT overall exhibited negative correlation with fluvial DO concentrations, due to asymmetric metabolic response to temperature increases and net heterotrophy shifts in fluvial ecosystems (37). Light and flow conditions are also key regulators of river metabolism (9). However, only 365 (1.7%) and 812 (3.7%) rivers were primarily driven by light and flow regimes, of which 81.2% were tropical and 16.9% were temperate rivers, predominantly in the Tibetan Plateau. Wind speed, as a key driver of air-water O₂ exchange (38), exhibited the most significant correlation in 388 (1.8%) rivers and streams that are mainly located in the Amazon River basin, central Africa, and southeast Asia. Positive correlation demonstrated the decline in DO concentrations with slowing wind speed in tropical flowing waters (fig. S11E).

Fig. 3. Primary drivers of long-term trends in fluvial DO concentrations.

(A) Partial correlation between fluvial DO concentrations and various potential drivers, including oxygen solubility (OS), surface air temperature (SAT), shortwave radiation (SR), discharge (D), and wind speed (WS). The inset pie chart illustrates the proportion of each driver with the highest correlation. Flowing waters labeled as “others (OT)” indicate no significant partial correlation with any of these drivers. Histogram shows the number of flowing waters having significant correlation in tropical, temperate, and frigid regions, respectively. (B) Highest significant partial correlation coefficients between fluvial DO concentrations and primary drivers across the globe.

Fluvial DO saturation percentage (DO SP), defined as the ratio between DO concentrations and OS (text S3), proved that metabolism played crucial roles in regulating DO levels in tropical rivers. DO SP measures the contributions of metabolic, chemical, and biogeochemical processes without confounding relationship from temperature (19). Overall, global DO SP significantly decreased at a rate of −1.4% decade⁻¹ (P < 0.001); associated pronounced decline collectively centered on tropical regions (20°S to 20°N, fig. S12D). In tropical regions, 63.1% of flowing waters experienced DO SP decreases, potentially indicating river respiration exceeding photosynthesis. Conversely, DO SP increases characterized central Amazon River basin and some river reaches in the southeast Asia (fig. S12B), suggesting that photosynthetic oxygen supply outweighs climate warming–induced deoxygenation. Such mechanism could certainly support increasing trends in fluvial DO concentrations (Fig. 1B).

Future projections of sustained fluvial deoxygenation

Sustained fluvial deoxygenation is projected to continue in the future warmer world. On a global basis, the deoxygenation rate and significance increase with the severity of climate warming, from −0.013 mg liter⁻¹ decade⁻¹ (R² = 0.78, P < 0.001) under SSP1–2.6 to −0.066 mg liter⁻¹ decade⁻¹ (R² = 0.99, P < 0.001) under SSP5–8.5 (Fig. 4A). Spatial patterns in deoxygenation are consistent across the different climate scenarios. We observed the fastest deoxygenation in high-latitude and highland rivers, such as the pan-Arctic and Tibetan Plateau witnessing amplified climate warming. However, 2.4% of global rivers are expected to experience persistent increases in DO levels between 2024 and 2100, and primarily focus on the Amazon River basin and southeast Asia under four climate scenarios (Fig. 4, B to E). As a result, worldwide deoxygenation reduces fluvial DO concentrations by 1.1% ± 1.6% under SSP1–2.6 and 4.7% ± 2.7% under SSP5–8.5 throughout the 21st century (fig. S13). Several regions, including India, southeast US, and Parana Plateau, experience the largest magnitude decline in fluvial DO concentrations, exceeding 12% under SSP5–8.5 (fig. S13D), suggesting that local fluvial ecosystems more easily suffer from hypoxia and water quality degradation. By comparison, there are some rivers located in the Amazon River and southeast Asia showing slight increases in fluvial DO concentrations, with a magnitude of <3% under SSP5–8.5, markedly smaller than deoxygenation magnitude globally.

Fig. 4. Future projections of changes in global fluvial DO concentrations.

(A) Long-term changes in fluvial DO concentrations from historical (1985 to 2023) to future periods (2024 to 2100). The solid lines and associated shade areas represent the mean values and standard deviations among different general circulation models (GCMs) in the latest CMIP6 project, respectively. Significant trends in fluvial DO concentrations are annotated by “*.” (B to E) Spatial distributions in the projected trends under four representative climate scenarios, including SSP1–2.6 (B), SSP2–4.5 (C), SSP3–7.0 (D), and SSP5–8.5 (E).

DISCUSSION

We established a GDOD that provides an unbiased picture and overwhelming evidence of sustained river deoxygenation on a global basis in response to climate warming, albeit with large variations spatially. Our analysis corroborated previously documented worldwide river deoxygenation and escalating low-oxygen and hypoxic stress, as well as the dominant roles of climate warming (29), whereas we also revealed that high-altitude rivers, such as the Tibetan Plateau, have higher DO levels than those plain rivers at the same latitude areas. Such disagreement in spatial patterns might be associated with photosynthetic regulation to fluvial DO concentrations. Our results indicate a net DO decline in China’s rivers between 1985 and 2023, which seems to conflict with previously reported increasing trends in large-proportion rivers from 2006 to 2020 (21, 39). However, these patterns overall agree with increasing DO SP, at a rate of 5.7% decade⁻¹ in China’s fluvial ecosystems from 2006 to 2020, especially in the northern regions, which suggested that net autotrophy flowing waters are dominant. Under this condition, our different finding is likely because of substantial expansion of temporal coverage or net autotrophy-induced DO enhancement offset by deoxygenation under climate warming.

Mechanism behind changing fluvial DO concentrations is complex due to many covarying variables. For example, both OS and metabolic regimes can regulate fluvial DO concentration globally, while temperature dependence makes them covary with each other. Moreover, fluvial metabolism could also be driven by various covarying factors, such as light and flow regimes, as well as intrinsic nutrient and organic matter (2, 3, 9). We apply partial correlation to isolate the contribution of each driver to fluvial DO concentrations, and ultimately conclude that fluvial metabolism (photosynthesis versus respiration), reflected by SAT, shortwave radiation, and discharge, plays non-neglecting roles in changing DO concentrations, despite declining OS as the top primary driver. However, there are 21.6% of global river reaches categorized into “other,” mainly focusing on tropical flowing waters (77.7%), indicating that changes in fluvial DO concentrations are attributed to other drivers without consideration in our analysis (Fig. 3A). Of these flowing waters, 74.8% (3459) experience consistent trends between fluvial DO and DO SP, including 54.9% as decreasing trends and the remaining 19.9% as increasing. Such patterns suggest that fluvial DO concentrations in “other” rivers might be explained by ecosystem metabolism, especially in tropical zones. In this regard, previous studies demonstrated significant increases in dissolved organic matter in some tropical rivers, such as the Amazon, Orinoco, Parana, and Nile, throughout the late half of 20th century (40). High-level organic matter can greatly enhance benthic and plankton respiration (41), potentially serving as a potential driver of deoxygenation in “other” tropical fluvial ecosystems. In addition, substantial nutrient (nitrogen and phosphorus) discharge from urban wastewater (42, 43), agricultural fertilization (44), atmospheric deposition (45), and biological fixation (46) can stimulate phytoplankton growth and blooms, and potentially regulate fluvial DO concentrations. Thus, the relationship between intrinsic-related metabolism and long-term changes in DO concentrations are adequately assessed in “other” rivers.

Our work presents an unbiased and comprehensive picture of pervasive deoxygenation in global fluvial ecosystems, underscoring that much more attention and efforts are necessary for flowing waters to mitigate rapid deoxygenation, especially those experiencing unprecedented hypoxic stress, such as India and eastern US (Fig. 1B and fig. S13), due to the importance of sustaining aquatic biodiversity and ensuring domestic water supply (1, 47). The India’s rivers suffer from the largest decline in fluvial DO concentrations, ultimately leading to DO levels that dropped below 5 mg liter⁻¹ by the end of 21st century. Such oxygen levels severely damage well-oxygenated habitats and survival conditions for aquatic species and organisms, and potentially trigger oxygen-sensitive species loss, commercial fish kills, and water quality degradation. Furthermore, we find that dam construction alters trends in fluvial DO concentrations in its impoundment area: negative in shallow reservoirs but positive in deep reservoirs, with its turning point as ~30 m for water depth (fig. S9, B and D). Currently, over 50,000 dams have been constructed worldwide and impede approximately half of the planet’s major fluvial ecosystems; 70% of them, however, have reservoir depths below 30 m (48, 49) and certainly contribute to global river deoxygenation. There are over 3700 new dams either planned or under construction worldwide that are expected to further exacerbate river deoxygenation globally (50). Under these conditions, it is imperative to formulate and implement targeted measures and strategies to mitigate river deoxygenation risks and prevent concomitant ecological disasters worldwide, particularly for India and eastern US.

MATERIALS AND METHODS

Satellite data and preprocessing

We collected the entire archive of Landsat 5–9 surface reflectance images between 1985 and 2023 from the Google Earth Engine (GEE) cloud platform. All these surface reflectance images were atmospherically corrected based on the Landsat Ecosystem Disturbance Adaptive Processing System (LEDAPS) and Landsat Surface Reflectance Code (LaSRC) (51), and widely used to monitor long-term changes in water quality across the entire globe (30, 52, 53). The “CFmask” provided by the GEE cloud platform was used here to remove the outlier pixels that are regarded as cloud coverage, thick aerosols, and snow cover (54). The central pixels of global rivers and streams were extracted using the Global River Widths from Landsat (GRWL) dataset to determine flowing water extents (55), and then masked the pixels as occasional contamination (i.e., bridges, boats, and floating algae) using a predefined threshold (>0.05) of the modified normalized difference water index, as used in (31).

Considering the difference in temporal coverage among sensors, the surface reflectance of Landsat images was normalized to conduct cross-sensor correction using the following equation

$${\text{SR}}_n\left({\mathrm{\lambda }}_i\right)=\frac{\text{SR}\left({\mathrm{\lambda }}_i\right)}{{\displaystyle {\sum }_{i=1}^n}\text{SR}\left({\mathrm{\lambda }}_i\right)}$$

where SR_n (λ_i) represents the normalized surface reflectance at λ_i nm; SR (λ_i) is the surface reflectance; λ and n are the wavelength and number of the Landsat bands, respectively. Note that the normalization of surface reflectance was conducted on blue, green, red, and near-infrared (NIR) bands.

In situ fluvial DO and auxiliary data

We extracted the fluvial DO measurements from Global River Water Quality Archive (GRQA) that combines five regional and global datasets—Global Freshwater Quality Database (GEMStat), global river chemistry dataset (GLORICH), Canadian Environmental Sustainability (CESI), European waterbase, and Water Quality Portal (WQP)—and contains over 1.7 million water quality measurements in global rivers and streams from 1898 to 2020 (56). Using all fluvial DO concentrations with river width greater than 90 m, the concurrent satellite observations were matched as the source data for developing and validating the retrieval algorithm, based on the following criteria: (i) The date difference between satellite overpassing and in situ data was within 1 day to minimize the effects of dynamics in water conditions, and (ii) the 3 by 3 pixel windows centered on the sampling sites were homogeneous in the surface reflectance (i.e., coefficient of variation < 0.15). As a result, the selection criteria reduced millions of originally fluvial DO monitoring records to a total of 10,944 high-quality matching pairs, including 4710 pairs for Landsat 5 TM, 4294 for Landsat 7 ETM+, 1835 for Landsat 8 OLI, and 114 for Landsat 9 OLI. These matching pairs primarily covered the flowing waters in North America and Europe, with a range of 0.19 to 21.8 mg liter⁻¹ and an average concentration of 8.92 ± 2.46 mg liter⁻¹ (fig. S2A).

In addition to satellite surface reflectance, we also extracted the elevation from global digital elevation model (DEM) collected by the shuttle radar topography mission (SRTM), and SAT from ERA5-Land daily dataset. We regarded the elevations and latitudes as static properties for estimation of fluvial DO concentrations.

Development of the stacking retrieval algorithm

We developed the retrieval algorithm of fluvial DO concentrations through ensemble learning that combines and integrates various algorithms to exploit the advantages of conventional single learner, with an ultimate result of improving the model performance. Here, we used the random forest (RF) and the extreme gradient boosting (XGBoost) as the base models due to their robust performance and high popularity. We separately trained and tested the generic RF and XGBoost models using high-quality matching pairs of input parameters and observed DO concentrations distributed globally, and then calculated the average estimates from these two well-trained models as the stacking one.

Concentrations of DO in flowing waters vary with various factors, such as water temperature, light availability, and nutrient supplies. Specifically, water temperature can not only control gas solubility (57) but also affect ecosystem metabolism (i.e., primary production and respiration) that represents crucial oxygen sources and consumption (37). High light availability associated with water clarity and turbidity could carve favorable niches for phytoplankton photosynthesis (9). Nutrient supplies potentially manage the chlorophyll concentrations and phytoplankton biomass, and thus regulate the DO concentrations in aquatic ecosystems (3). Considering these complex mechanisms above, the input parameters of the DO retrieval algorithms included SAT as an alternative of water temperature due to their strong dependence, surface reflectance at blue, red, and NIR bands, as well as blue-green, red-green, and NIR-red reflectance ratios due to their sensitivity to turbidity and chlorophyll a concentrations in inland waters (58, 59). Moreover, we also used the absolute latitudes and elevations as input parameters owing to the latitudinal and altitudinal gradients of water temperature in flowing waters (60).

We randomly separated all high-quality matching pairs of input parameters and fluvial DO measurements into the training (70%, 7667) and testing subsets (30%, 3277), respectively. The fluvial DO measurements were uniformly normalized using Z-score method, while the normalization schemes for input parameters of two base models were manually optimized (refer to table S1). We then applied the grid iterative approach to optimize the hyperparameters, that is, the number of trees and maximum depth of each tree in the RF model, the learning rates, and maximum depth of each tree in the XGBoost model. The other hyperparameters, such as the booster and the number of estimators, remained default; loss function was set as the mean square error. Ultimately, the combined hyperparameters with the best performance were obtained as the optimal ones, with a tree number of 70, a maximum tree depth of 18 in the RF model, a learning rate of 0.14, and a maximum tree depth of 10 in the XGBoost model.

Creation of a GDOD

The well-trained stacking retrieval algorithm was applied to global river center pixels of Landsat 5–9 satellite images to establish a global fluvial DO concentration database. We aggregated fluvial DO concentrations into river reaches in the GRWL dataset and regarded them as the smallest units, and used several criteria as follows to select global river reaches with sufficient and high-quality fluvial DO estimates for inclusion in the GDOD.

First, we only focused on the retrievals of fluvial DO concentrations in ice-free periods between 1985 and 2023. The river ice cover in each segment was calculated based on a logistic regression model between ice extent and corresponding 30-day prior mean SAT, developed in (61). When ice extent exceeds 50%, the ice period will begin and last until ice extent below 50%. These two dates were the freeze- and break-up days, and ice duration for the time between freeze- and break-up dates. We determined the ice-free periods as the time spanning from the latest break-up day and the earliest freeze-up day between 1985 and 2023. Consequently, the ice-free periods in low- and mid-latitude flowing waters lasted throughout four seasons of the year, while the entire year was almost covered by ice in the Arctic rivers (fig. S1B).

Second, we excluded all river channel segments with the width smaller than 90 m and ice-free satellite observations fewer than 12 times per year to minimize the potential effects of pseudo-signals associated with adjacent land pixels and sparse observations. Moreover, river reaches with sufficient satellite observations in less than 10 years were also excluded to ensure the validity of long-term trend examination.

We ultimately obtained a total of 21,439 river reaches globally, with the length ranging from 0.06 to 877.6 km and total length accounting for 78.1% of global rivers and streams wider than 90 m in the GRWL dataset; 89.6% of them are less than 100 km. Of these river reaches, the number of satellite observations distributed uneven among regions, where India, southern US, and Europe, as well as the eastern South America had greatly more observations (>2000 times from 1985 to 2023) than other regions. Overall, the mean number of satellite observations in global flowing waters reached 36.3 per year during the study period (fig. S1A). We estimated fluvial DO concentrations in each river reach based on all available Landsat images and then calculated monthly mean values to construct the fluvial DO database in global fluvial ecosystems.

Analysis of long-term trends in fluvial DO concentrations

Our analysis of long-term trends in fluvial DO concentrations was conducted on their annual anomalies between 1985 and 2023. We first calculated monthly anomalies as the differences from monthly mean values during the study period and then estimated annual mean anomalies within each river reach. The linear regression was used to determine the long-term trends (i.e., the linear slopes) and associated significance in fluvial DO concentration over the past four decades.

Moreover, our analysis focused on temporal trends in fluvial DO concentrations during three different flow seasons, that is, low-, normal-, and high-flow seasons, to demonstrate the regulation of flow regimes to fluvial DO trends globally. We extracted daily river discharge from Global River Discharge Reanalysis (GRDR) dataset (62) and calculated flow exceedance probability to develop flow-duration curves between 1985 and 2018. Flow exceedance probability was defined here as the proportion of the number of river discharge exceeding that in one given day within the period of 1985 to 2018 (63). We then applied a 15-day running mean to smooth daily flow exceedance probability and to further facilitate detection of low- and high-flow seasons (fig. S14). When smoothed flow exceedance probability exceeded 75th percentile threshold over the entire study period and lasted at least 7 consecutive days, this date was the onset day of a low-flow season. Similarly, flow exceedance probability in next 7 consecutive days below 75th percentile threshold indicated the end of the low-flow season. A low-flow season persisted between these two dates. By contrast, high-flow seasons were defined to start when smoothed flow exceedance probability in the next 7 consecutive days dropped below 25th percentile threshold, and to end until exceeding 25th percentile threshold. The remaining days within 1 year were thus normal-flow seasons. Overall, low-flow seasons tended to occur in winter and early spring, albeit with considerable variations among regions. For most of mid- and high-latitude flowing waters, low-flow seasons primarily focused on March and April, with an occurrence probability nearly reaching 100%. However, as latitude decreases, low-flow seasons shifted toward January and December; associated occurrence probability dropped down to 50%. There were several exceptional examples, such as eastern US and western Europe showing the largest low-flow occurrence probability in summer and autumn (fig. S15, A and B). Conversely, high-flow seasons were likely to occur in summer (June to August), with the occurrence probability exceeding 90% in South Asia, western Canada, and Africa, while eastern US and western Europe had high-flow seasons between February and May (fig. S15, C and D).

We partitioned all fluvial DO estimates in each river reach into these three flow seasons and examined global patterns and trends over the past four decades. Note that three different flow-seasons were assumed to remain unaltered from 2018 to 2023, due to the limited availability of river discharge.

Examination of primary drivers for changing fluvial DO

Fluvial DO concentrations are generally regulated by OS, air-water exchange, and metabolism. Individually, OS depends on water temperature and air pressure. We determined fluvial OS between 1985 and 2023 based on a lookup table, using SAT and surface pressure from ERA5-Land daily dataset, as in (19). Wind speed and river discharge can affect air-water exchange, while fluvial metabolism encompasses photosynthesis and respiration, and varies with multi-variables, such as temperature, light and flow regimes, as well as nutrient and organic matter (3). Owing to unavailable nutrient and organic matter time series at large scales, we selected OS, SAT, shortwave radiation, discharge, and wind speed as potential primary drivers of changing fluvial DO concentrations globally. Shortwave radiation and wind speed were available through ERA5-Land daily dataset.

As these potential drivers may covary with one another, we needed to carefully deal with such covariance when detecting the linkage between fluvial DO trends and one independent variable. Partial correlation, as an effective statistical tool to isolate two-variable dependence from the covariate effects of many correlated variables, has been widely used in regional and global studies (64–66). We thus applied it here to examine the linkage between fluvial DO and each potential driver without the confounding effects of other variables, where the correlation coefficient represents its contribution to fluvial DO changes. The highest significant partial correlation (P < 0.05) in each river reach was ultimately derived as the primary driver of long-term trends in fluvial DO concentrations between 1985 and 2023. If all these variables showed nonsignificant partial correlation, river reach was categorized into “other,” meaning that there were other factors responsible for changing fluvial DO concentrations.

Future projections of fluvial DO response to climate warming

We projected future changes in fluvial DO concentrations across the globe through climate outputs from 10 general circulation models (GCMs) under four Shared Socio-economic Pathways, that is, SSP1–2.6, SSP2–4.5, SSP3–7.0, and SSP5–8.5 in the latest CMIP6 project. Of these GCMs, bias corrections were conducted, where climatological mean difference between daily ERA5-Land and historical SAT from 2000 to 2014 was regarded as the bias, and thereby added to future climate outputs under each SSP. We then extracted bias-corrected SAT to drive the well-trained retrieval algorithm, while the other input parameters, including surface reflectance and its band combinations, were assumed as unaltered mean values during the last 5 years (2019 to 2023); the latitude and elevation remained static in the projection of future fluvial DO concentrations. We also assumed that the minimum ice-free periods were unaltered under both the historical and future periods to extract fluvial DO concentrations in ice-free periods between 2024 and 2100.

Supplementary Materials

This PDF file includes:

Texts S1 to S4

Figs. S1 to S16

Table S1

References

Data, code, and materials availability:

All data and code needed to evaluate and reproduce the results in the paper are present in the paper and/or the Supplementary Materials. Landsat surface reflectance and ERA5-land dataset used in this study is accessible through Google Earth Engine (GEE) at https://earthengine.google.com. The GRQA and GRWL datasets are available at https://zenodo.org/records/6347038 and https://zenodo.org/records/1297434, respectively. The satellite-derived global fluvial DO database (GDOD) and associated code are publicly available at https://zenodo.org/records/19148556.