Satellites reveal widespread deoxygenation of large global reservoirs from 1984 to 2023

Dissolved oxygen (DO) is a fundamental indicator of water quality and ecosystem health in aquatic systems^1,2,3. It regulates the survival, growth, and distribution of fish, invertebrates, and microorganisms, while also controlling key biogeochemical processes such as nutrient cycling, carbon turnover, and greenhouse gas emissions⁴. Deoxygenation—manifesting as hypoxia (DO < 2 mg/L) or anoxia (DO < 0.5 mg/L)—has severe ecological and societal consequences, including fish kills, harmful algal blooms, loss of biodiversity, and deterioration of drinking water quality^5,6,7. These processes are especially pronounced in reservoirs, where water-column stratification and anthropogenic regulation often exacerbate oxygen depletion compared with many natural lakes or rivers, although some deep natural lakes also develop strong stratification and some shallow lakes experience substantial anthropogenic impacts^3,8,9,10.

Reservoirs are highly sensitive to deoxygenation due to their dual roles as engineered systems and aquatic ecosystems^11,12. Thermal and chemical stratification restrict vertical mixing, confining oxygenated waters to the surface and promoting hypoxia or anoxia in bottom layers¹³. This process not only reduces habitat availability for aquatic life but also enhances the release of phosphorus, nitrogen, and reduced metals from sediments, fueling eutrophication and impairing water supply⁵. Moreover, anthropogenic drivers—including damming, altered hydrological regimes, land-use change, and nutrient loading—intensify oxygen depletion and amplify climate-induced warming effects^14,15. Collectively, these processes make reservoirs hotspots of rapid and widespread deoxygenation, with cascading consequences for aquatic biodiversity, nutrient cycling, and human health^2,3,7.

Despite its importance, long-term monitoring of DO in reservoirs remains challenging. Traditional in-situ measurements provide accurate but highly localized data^16,17, often failing to capture the strong spatial and temporal heterogeneity of oxygen dynamics in large and complex reservoirs^6,9,18,19. The labor-intensive nature of field sampling limits the temporal coverage needed to assess seasonal and interannual trends, while sparse monitoring networks are insufficient to characterize basin-wide deoxygenation patterns¹⁸. These constraints underscore the need for innovative monitoring strategies capable of reconstructing historical DO variability across broad spatial and temporal scales^20,21,22.

The integration of remote sensing and machine learning provides significant advantages for analyzing deoxygenation in large spatio-temporal scales. Satellite data offers broad-scale, high-frequency coverage, overcoming the limitations of traditional point sampling methods²³. Specifically, Landsat archives provide spectral data with a 16-day revisit period and 30-m spatial resolution, containing information on water constituents like chlorophyll-a, suspended sediment, and dissolved organic matter, all of which are closely related to DO production and depletion²⁴. Furthermore, machine learning algorithms are crucial for interpreting complex, multi-dimensional remote sensing data, enabling the identification of intricate, non-linear relationships between spectral signatures and DO levels. Recent studies have successfully used machine learning to retrieve DO concentrations from satellite data and even predict deoxygenation events^6,7,9,18,19. This allows us to reconstruct historical DO data over the past four decades, facilitating in-depth analysis of long-term trends and patterns. Thus, the integration of remote sensing and machine learning offers a powerful tool for understanding deoxygenation of reservoirs globally.

In this study, we present a generalized global reservoir DO retrieval method utilizing four decades of Landsat data and multiple machine learning algorithms to reconstruct the temporal dynamics of DO in 357 major reservoirs (area > 100 km²) globally. We first compile a comprehensive DO dataset covering the globe, ensuring sufficient data for the development of robust machine learning models. We then use our DO retrieval model to: (1) characterize multi-decadal spatiotemporal patterns of reservoir deoxygenation, (2) quantify the four-decade trends in reservoir deoxygenation, and (3) disentangle the relative contributions of climate warming and anthropogenic regulation to observed DO changes. This study introduces a practical methodology for large-scale spatiotemporal reconstruction of DO using remote sensing and machine learning, enhancing our understanding of human impacts and climate change on DO dynamics in global reservoirs.

In-situ DO data and matchup

Field DO observations are downloaded from several public sources including the Global River Water Quality Archive (GRQA)¹⁷ and the Global Freshwater Quality dataset (accessible at https://gemstat.bafg.de/). We first filter the in-situ observations to the period 1984–2023, corresponding to the temporal coverage of the satellite record. We then implement additional manual quality control (e.g., applying a strict spatio-temporal matching window and removing low-quality pixels) to identify suitable satellite-field data matchups for algorithm development (Fig. 1). To accurately identify water surfaces, we employ the widely used Dynamic Surface Water Extent (DSWE) method²⁵. We adopt a widely used cross-sensor harmonization approach to ensure consistency in sensor characteristics and minimize discrepancies from multi-sensor integration, including Landsat ETM, ETM+, and OLI. Specifically, we apply cross-sensor calibration following established procedures²³, which have been demonstrated to reduce systematic biases among Landsat sensors. We implement rigorous temporal matching criteria, restricting the time difference between field measurements and satellite overpasses to a strict ± 6 h window. Spatial data processing involved extracting a 3 × 3 pixel window surrounding each sampling location to mitigate adjacency effects. This is followed by comprehensive quality screening using QA_PIXEL band bitmask technology to systematically exclude pixels contaminated by clouds, snow, or cloud shadows. Only windows retaining over 50% valid pixels, defined as pixels free from cloud, cloud shadow, and snow contamination according to the QA_PIXEL bitmask, with a spatial coefficient of variation below 0.15 (calculated within the 3 × 3 pixel window) underwent further analysis. The final reflectance values for all six Landsat spectral bands (blue, green, red, near-infrared, SWIR1, and SWIR2) calculated as the mean of these quality-controlled pixels. Following these procedures, we successfully match a total of 32,065 DO measurements globally, including 1,732 matchups for the studied reservoirs.

Satellite data and pre-processing

The USGS Landsat archive offers valuable long-term satellite data (1984–2023), enabling scientific research on historical water quality parameter retrieval. We first select major reservoirs (area > 100 km²) based on boundaries provided by the Global Surface Water (GSW)²⁶. Given that reliable global reservoir volume data are still limited and often inconsistent due to variations in water depth and bathymetric uncertainty²⁶, we used surface area as a practical and consistent criterion for reservoir selection. Reservoir surface area and volume are strongly correlated globally (R² = 0.76; Supplementary Fig. 1), supporting the use of area (> 100 km²) as a robust proxy for identifying major reservoirs in this study. We define permanent water as pixels with ≥ 75% water occurrence frequency, based on global imagery from the GSW dataset (30 × 30 m resolution) and considering seasonal water body variations¹⁸. We acquire Landsat-5, Landsat-7, and Landsat-8 images from the United States Geological Survey (USGS) archive and process them using ACOLITE²⁷, a widely used atmospheric correction method. We then use a widely used method to maintain sensor consistency across the four-decade study period²⁸. Cloud-contaminated scenes are filtered using Landsat’s internal metadata, retaining only images with less than 30% cloud cover. Additionally, rigorous pixel-level quality control was applied using the QA_PIXEL band bitmask to eliminate unreliable observations, including those affected by clouds, snow, or cloud shadows. Water pixels are identified using the USGS-developed DSWE algorithm²⁵. Finally, we obtain 159,969 processed images for DO retrieval (Fig. 1).

Machine learning retrieval model development

To predict DO, we select three representative machine learning algorithms: Random Forest (RF), eXtreme Gradient Boosting (XGBoost), and Support Vector Regression (SVR). Model development is conducted using the scikit-learn package (version 1.3.1) in Python 3.8. Random samples from the integrated dataset are partitioned into training (70%) and testing (30%) sets. Hyperparameters (e.g., number of trees, maximum depth, learning rate, and kernel parameters) are optimized using a grid search strategy coupled with 10-fold cross-validation. We incorporate remote sensing reflectance (e.g., visible bands and near-infrared band), hue angle, longitude, latitude, and DEM information into the model following a recent study¹⁸. SWIR bands are excluded because strong water absorption in the SWIR (especially > ~ 1,300 nm) yields negligible water-leaving reflectance and low signal/noise ratio for most inland waters²⁹. Hue angle serves as a robust proxy for water color, and is widely used to assess water quality in various water bodies¹⁸. More details of the hue angle can be found in Hu et al.¹⁸. Feature importance is quantified via SHapley Additive exPlanations (SHAP) values. Details of the optimized hyperparameters are provided in Table 1. Each of these three algorithms offers distinct advantages: RF is an ensemble learning method that aggregates predictions from multiple independent decision trees. It is well-regarded for its robustness and is widely applied in water quality research; XGBoost is a powerful boosting ensemble algorithm built on gradient-boosted decision trees. It incorporates regularization techniques to control model complexity, prevent overfitting, and enhance computational efficiency. Previous studies have also highlighted its strong predictive capabilities and superior performance in water quality parameter estimation. Support Vector Regression (SVR), grounded in structural risk minimization theory, excels at handling small sample sizes and nonlinear problems, commonly applied in fields like water quality retrieval by satellites.

Spatiotemporal analysis

To minimize the influence of ice cover on DO estimates, we focus on satellite imagery acquired during the ice-free periods: July-September in the Northern Hemisphere and January-March in the Southern Hemisphere^7,30. The average DO concentration during these months is used to represent each reservoir and year. To analyze long-term trends in annual DO, we use the non-parametric Mann-Kendall test, setting the significance level (α) to 0.05. We also investigate the relationships between DO and other environmental variables using Pearson correlation. The statistical significance of these correlations, as well as differences between group means, is assessed using a t-test with a significance threshold of p < 0.05. While Pearson’s correlation coefficient quantifies the strength and direction of the linear relationship between two continuous variables, the t-test determines whether the observed relationship or difference is statistically significant, taking into account the sample size and variability.

Quantification of contributions of DO change drivers

The pervasive decline in reservoirs DO concentrations stems from a complex interplay of environmental factors, anthropogenic perturbations, and biogeochemical processes^18,20,21. Disentangling the roles of these various drivers is paramount for understanding reservoir ecosystem development and safeguarding aquatic biomass health²¹. To precisely quantify their individual contributions, we employ a machine learning attribution method. This comprehensive analysis encompassed environmental factors (e.g., temperature, precipitation, wind, and evaporation), human activities (e.g., populations, urban area, and cropland area), and other key biogeochemical processes (e.g., turbidity and primary production). To improve the physical representation of temperature effects on DO solubility, we reprocess Landsat Level-2 Surface Temperature (LST) data across all studied reservoirs during ice-free months and use these observations to calibrate ERA5 air temperature via regression correction (Supplementary Fig. 2). The corrected temperatures are subsequently used for model development and driver attribution. Our data sources include the ECMWF Reanalysis v5 (ERA5) Land product for environmental factors, the MODIS Land Cover Type Product (https://lpdaac.usgs.gov/products/mcd12q1v006/) for urban and cropland areas, and LandScan (https://landscan.ornl.gov) for population data. Reservoir turbidity is precisely calculated using the Normalized Difference Turbidity Index (NDTI), which is a widely used proxy for aquatic turbidity levels.

Model performance evaluation

Model performance is comprehensively evaluated using two key metrics: the coefficient of determination (R²) and the root mean squared error (RMSE). R² quantifies the proportion of variance in the dependent variable that is predictable from the independent variables, with values ranging from 0 to 1. Higher R² values indicate a better model fit to the observed data. Conversely, RMSE measures the average magnitude of the errors between predicted and observed values. Lower RMSE values denote higher model prediction accuracy. The formulas for calculating these evaluation metrics are as follows:

where (:{X}_{i}) represents the in-situ measured data, and (:{Y}_{i}) represents the retrieved data.

Characteristics of observed DO data

The global assessment of reservoir dissolved oxygen (DO) concentrations reveals pronounced intercontinental disparities (Fig. 2). Europe exhibits a notably high mean DO level of 9.8 ± 1.3 mg/L, supported by a robust sample size (N = 11,525). Africa records the highest mean concentration (10.8 ± 2.1 mg/L), which may reflect relatively undisturbed, pristine aquatic environments; however, this interpretation warrants caution given the smaller sample size (N = 903). In contrast, the moderate DO levels observed in Asia (8.2 ± 1.3 mg/L; N = 13,300) and North America (7.9 ± 1.2 mg/L; N = 2,158) likely result from the complex interplay between extensive anthropogenic pressures and ongoing mitigation measures. More concerning patterns emerge in South America and Oceania, where the comparatively low mean DO concentrations (6.7 ± 0.8 mg/L and 6.3 ± 1.1 mg/L, respectively) suggest heightened anthropogenic stress or distinct regional environmental constraints, underscoring the need for further investigation and targeted management strategies.

Model performance

Our comparison of three machine learning methods reveals marked differences in their ability to accurately estimate reservoirs DO concentration (Fig. 3). The RF model demonstrates the highest performance, achieving an R² of 0.77 and a RMSE of 1.12 mg/L in the training set, and an R² of 0.73 with a RMSE of 1.23 mg/L in the testing set. The RF also shows a relatively tight clustering of predicted versus observed DO concentrations closely aligned with the 1:1 line, indicating a strong capability in capturing the complex nonlinear relationship between satellite data and DO. In contrast, the XGBoost model shows a moderate performance with an R² of 0.58 and an RMSE of 1.59 mg/L in testing set, with some uncertainties at higher DO concentrations (e.g., higher than 15 mg/L), and some banding at lower values (e.g., lower than 5 mg/L). The SVR model exhibits lower accuracy, with an R² of 0.45 and an RMSE of 1.93 mg/L in testing set, showing a considerable spread of data points and a tendency to under-predict higher DO values. With 10-fold cross-validation, the RF achieves robust DO estimation (R² = 0.72, RMSE = 1.27 mg/L), representing a significant improvement over other methods, which have R² values ranging from 0.42 to 0.56 and RMSE values from 1.63 to 1.91 mg/L. Based on these comparative metrics, the RF clearly stands out as the most effective algorithm among the tested methods for reliably estimating reservoir DO concentrations in this study (Fig. 3). To evaluate the model performance at the study sites, we compared satellite-derived and measured DO concentrations across 326 reservoirs with available in situ measurements (total 1,732 matchups). The RF model achieve an R² of 0.69 and an RMSE of 1.33 mg/L (Fig. 4).

The superior performance of the RF model may be attributed to its ensemble structure, which effectively captures complex nonlinear relationships and interactions among multiple predictors while being less sensitive to parameter tuning³¹. By contrast, SVR can struggle with highly nonlinear dynamics in large, heterogeneous datasets, often leading to underestimation of high DO values³². XGBoost, although powerful, may require more extensive parameter optimization and larger training datasets to achieve comparable performance, and can be prone to overfitting when the number of features is relatively small³³. These characteristics likely explain the observed differences in model accuracy in our study context.

The spatial pattern of global major reservoirs DO concentrations

Our four-decade global assessment of DO dynamics across global 357 major reservoirs reveal pronounced spatial heterogeneity (Fig. 5). From 1984 to 2023, the global mean annual surface DO concentration is 8.4 ± 1.1 mg/L (mean ± SD). A distinct hemispheric contrast emerges: reservoirs in the Northern Hemisphere maintain higher DO levels (9.8 ± 0.9 mg/L) compared to those in the Southern Hemisphere (7.7 ± 1.4 mg/L) (Fig. 6). These concentrations generally remain sufficient to support diverse fish assemblages and avoid the stress associated with extreme oxygen depletion, positioning many reservoirs as critical ecological benchmarks for sustaining aquatic biodiversity.

Latitudinal gradients further highlight this heterogeneity (Fig. 5). Polar reservoirs (latitude > 66.56°) exhibit the highest DO concentrations (10.8 ± 0.5 mg/L), whereas equatorial reservoirs (5°S to 5°N) display the lowest (6.9 ± 0.8 mg/L). The extremes of this spectrum range from a minimum of 5.2 mg/L near the equator to a maximum of 11.5 mg/L in Arctic systems. The elevated DO levels in cold-region reservoirs exemplify pristine oxygenation, largely sustained by oligotrophic conditions, cold montane hydrology, and minimal anthropogenic disturbance. When classified by climatic zones, the North Frigid (10.8 ± 0.5 mg/L) and North Temperate zones (9.6 ± 0.7 mg/L) consistently maintain higher DO than the Tropical (6.9 ± 0.6 mg/L) and South Temperate zones (8.7 ± 1.1 mg/L, median 9.1 mg/L). Variability, expressed as the standard deviation of multi-decadal DO, is also higher in the North Frigid (0.6 ± 0.1 mg/L) and Tropical zones (0.6 ± 0.3 mg/L) compared with the more stable North Temperate (0.5 ± 0.1 mg/L) and South Temperate zones (0.4 ± 0.1 mg/L).

At the continental scale, North American (9.2 ± 0.8 mg/L) and European (7.5 ± 0.8 mg/L) reservoirs show significantly higher DO levels than those in Africa, Asia, and South America, where concentrations typically range between 7 and 8 mg/L (Fig. 6). However, localized low-oxygen anomalies occur in parts of Africa, Asia, and South America, with DO values falling well below the continental mean (8.4 ± 1.3 mg/L). This north–south disparity reflects interacting climatic and ecological drivers: northern reservoirs benefit from lower thermal regimes that enhance oxygen solubility, extensive riparian forest cover that mitigates nutrient loading, and wind-driven mixing that promotes vertical oxygenation, whereas equatorial and subtropical systems remain more vulnerable to thermal stratification and eutrophication^2,34. These findings not only highlight the global heterogeneity of reservoir oxygenation but also signal potential climate vulnerabilities. Especially, we find that reservoirs in North and South America, already exhibiting oxygen stress with some outliers significantly below the median, are particularly at risk of further deoxygenation (Fig. 6).

The temporal pattern of DO concentrations

We find that reservoirs worldwide show a decreasing trend in surface DO, with a mean rate of − 0.13 mg/L per decade (Fig. 7). This trend is not uniform across all reservoirs; globally, 74% of the studied reservoirs experienced a decrease in surface DO, averaging − 0.07 mg/L per decade, while 26% show an increase at a rate of 0.07 mg/L per decade. Reservoirs with increasing DO are predominantly located in tropical regions. This decline in reservoir surface DO is widespread on a continental level, with all six continents studied showing decreases over the past four decades: Africa (− 0.07 mg/L), Asia (− 0.06 mg/L), Europe (− 0.12 mg/L), North America (− 0.10 mg/L), Oceania (− 0.05 mg/L), and South America (− 0.05 mg/L, Fig. 8).

Seasonal surface DO of the studied reservoirs show greater variability compared to their interannual variations (Fig. 9). Seasonal surface DO fluctuations are greater in the Northern Hemisphere (1.8 ± 0.3 mg/L) and smaller in the Southern Hemisphere (1.1 ± 0.2 mg/L). Seasonal variations in DO concentrations display divergent patterns between the Southern Hemisphere and the Northern Hemisphere (Fig. 9). In the Southern Hemisphere, mean DO levels peak around July (8.1 mg/L) and reach their minimum in March (7.0 mg/L). Conversely, in the Northern Hemisphere, surface DO levels peak around January (11.3 mg/L) and are lowest around July (7.4 mg/L). Furthermore, the monthly averaged DO in the Southern and Northern Hemispheres shows contrasting trends: a convex pattern predominates in the Southern Hemisphere, while a combination of convex and concave patterns is observed in the Northern Hemisphere (Fig. 9). This pattern aligns with typical temperate reservoir dynamics where colder months favor higher oxygen solubility and reduced biological oxygen demand, while warmer periods experience both decreased solubility and increased respiratory consumption.

The controls of widespread decreases of DO concentrations

Our results reveal that the rates of change in reservoir surface DO is faster than those observed in the lakes⁷ (− 0.08 mg/L per decade), oceans²⁰ (− 0.02 mg/L per decade) and in rivers³ (− 0.04 mg/L per decade) over a similar time period. This highlights that reservoirs are particularly vulnerable to deoxygenation, likely due to the combined effects of anthropogenic regulation and climate forcing. SHAP analysis identifies temperature (19.4%) as the predominant driver of DO decline, a finding consistent with global warming impacts⁷ (Fig. 10). Although oxygen solubility is temperature-dependent³⁴, our analysis specifically focuses on summer (ice-free) months when intra-hemispheric temperature variability is relatively small. Moreover, by explicitly quantifying temperature’s contribution through SHAP analysis, we effectively account for thermal effects on DO dynamics without converting DO concentrations to percent saturation. This approach ensures that the observed spatial and temporal patterns reflect genuine biogeochemical and anthropogenic influences rather than artifacts of solubility normalization. Rising water temperatures directly reduce oxygen solubility while enhancing microbial respiration, creating a synergistic mechanism for oxygen depletion¹⁸. Cropland (13.9%) emerges as the secondary driver, largely through agricultural runoff (fertilizers, pesticides), which triggers eutrophication and promotes algal blooms, whose decomposition by heterotrophic bacteria further depletes DO via biochemical oxygen demand^2,34 (Fig. 10). However, we acknowledge that reservoirs artificially regulated by dams may be less affected by agricultural fertilization activities compared to natural lakes, yet previous studies have demonstrated that nutrient enrichment can exert long-term controls on DO dynamics through eutrophication processes³⁵. Primary production (12.7%) affects DO by increasing oxygen through photosynthesis, but when nutrient enrichment promotes algal biomass accumulation, the later breakdown of this biomass intensifies microbial oxygen consumption and reduces^5,21,36. Together, these results reveal a dual-threat framework, which climate-driven thermal stress and nutrient pollution act in concert to amplify hypoxia risk in aquatic ecosystems (Fig. 10).

Beyond primary drivers, secondary mechanisms exert nuanced but critical influences on deoxygenation dynamics. Wind speed (12.3%) demonstrates dual modulation, e.g., wind-driven mixing enhances surface aeration, whereas prolonged calm periods promote stratification, isolating hypoxic waters in benthic zones⁷ (Fig. 10). Turbidity (10.7%), largely driven by sediment influx, attenuates photic zone depth, concurrently inhibiting photosynthetic O₂ production and stimulating organic matter mineralization^5,21. Anthropogenic factors are equally pivotal, e.g., urban land cover (8.9%) and population density (8.2%) drive hypoxia through wastewater effluents and nutrient leaching from impervious surfaces²¹ (Fig. 10). Precipitation (5.4%) exhibits paradoxical effects, which dilutional oxygenation versus pulse-loaded nutrient surges³⁴. Ecological shifts captured by NDVI change rate (5.0%) modulate nutrient retention, while evaporation (3.4%) concentrates dissolved constituents, exacerbating oxygen deficits through volumetric compression⁷. Collectively, these interdependencies generate nonlinear feedbacks that govern the basin’s spatiotemporal deoxygenation gradients.

Beyond above climatic and watershed drivers, several reservoir-specific factors also play critical roles in shaping DO dynamics. Unlike natural lakes, reservoirs are subject to direct human regulation, including dam release operations, seasonal water storage, and hydropower generation, all of which can alter vertical mixing and stratification regimes³⁵. For instance, selective withdrawal during dam releases often exports hypolimnetic water with low oxygen concentrations, further exacerbating downstream hypoxia³. Moreover, reservoir morphology—such as mean depth, bathymetry, and shoreline complexity—can interact with meteorological processes to influence oxygen dynamics³⁷. Shallow or morphologically complex reservoirs tend to experience stronger wind-driven mixing and shorter stratification periods, whereas deeper reservoirs with smoother bathymetry are more prone to persistent hypolimnetic oxygen depletion. Although these morphological and operational factors are challenging to quantify globally, their potential impacts on reservoir deoxygenation warrant further investigation in future research.

Model advantage and limitation

In this study, we develop a generalized method to robustly reconstruct the spatio-temporal patterns of deoxygenation across major reservoirs globally. Our machine learning framework further enhances the ability to downscale and harmonize satellite-derived water quality metrics with historical DO measurements, enabling a multi-decadal reconstruction of deoxygenation trends with R² is 0.73 and RMSE is 1.23 mg/L in testing set (Figs. 3 and 4). This approach not only captures interannual variability driven by climate oscillations but also identifies hotspots of hypoxia linked to anthropogenic pressures such as agricultural runoff and urbanization.

However, while satellite remote sensing provides invaluable insights, several limitations must be acknowledged. First, cloud cover and atmospheric interference can lead to data gaps, particularly in temperate regions with frequent cloudiness^6,38. Second, current satellite sensors are constrained by their revisit cycles (e.g., the temporal resolution of Landsat at 16 days), making it challenging to achieve daily-scale monitoring of rapidly changing oxygen dynamics, such as those driven by storm events or sudden algal blooms³. Despite these constraints, our methodology optimizes the use of available satellite data by leveraging statistical gap-filling techniques and multi-sensor fusion to improve temporal resolution. Future advancements in hyperspectral satellites (e.g., NASA’s PACE mission) and geostationary water quality monitoring could further enhance our ability to track reservoir deoxygenation at finer spatio-temporal scales³⁹. By combining satellite observations with emerging autonomous in situ sensors (e.g., gliders, buoys), we can move toward a more comprehensive understanding of deoxygenation mechanisms in inland waters²⁸.

In addition, although our DO retrieval model achieves reasonable predictive results (R² = 0.73, RMSE = 1.23 mg/L), we acknowledge that this level of uncertainty should be carefully considered. Specifically, an error of ~ 1.23 mg/L represents approximately 15% of the global mean DO concentration (8.4 ± 1.1 mg/L), which may affect the precise quantification of DO at individual reservoirs, especially in aquatic systems with relatively low DO concentrations where small biases could alter ecological thresholds such as hypoxia (DO < 2 mg/L). Overall, our study demonstrates that remote sensing, despite its limitations, is indispensable for large-scale, long-term reservoir deoxygenation assessments, providing critical data to inform water resource management and climate adaptation strategies.

While our study focuses on reservoirs larger than 100 km², the methodology could, in principle, be extended to smaller reservoirs or other types of inland waters. However, smaller water bodies often exhibit more rapid hydrodynamic responses, higher variability in thermal stratification, and distinct catchment influences¹¹, which may require additional calibration of the machine learning model. Future work could leverage higher-resolution satellite imagery, autonomous in situ sensors, and targeted field campaigns to adapt and validate the model for these smaller systems. Such efforts would expand the applicability of our framework, enabling comprehensive assessments of oxygen dynamics across a wider range of inland waters.

Implication for reservoirs management and the sustainable development goals

Our reconstruction of four decades of surface DO dynamics in global reservoirs provides critical insights into both ecological processes and reservoirs management strategies¹¹. Reservoirs serve as vital water sources for drinking, irrigation, and hydropower production, while simultaneously sustaining diverse aquatic ecosystems³⁷. Progressive deoxygenation directly threatens these functions by reducing habitat suitability for fish, altering food-web dynamics, and accelerating biodiversity loss^40,41. Low oxygen conditions also facilitate the release of nutrients and metals from sediments, thereby promoting harmful algal blooms and compromising drinking water quality^1,42. These effects are of particular concern in densely populated regions where reservoirs are the primary sources of freshwater supply⁴³. Despite their relatively high present-day DO concentrations, reservoirs in Europe and North America exhibit steeper DO declines compared with other continents. This accelerated decline likely reflects the combined effects of reservoir aging and legacy nutrient loading, which enhance eutrophication and oxygen depletion over time^7,14. Additionally, intensified thermal stratification under regional warming reduces vertical mixing and oxygen replenishment, further exacerbating hypoxia risks in temperate reservoirs with strong anthropogenic influence. These findings highlight that even technologically advanced regions are not immune to the escalating challenges of deoxygenation, emphasizing the need for adaptive management strategies tailored to regional hydroclimatic and socio-economic contexts.

Our findings of widespread deoxygenation of global reservoirs underscore the need to incorporate oxygen dynamics into reservoir operation and design. Thus, integrating DO monitoring into water resource management frameworks is crucial for safeguarding ecosystem services and ensuring reliable water supplies under conditions of climate warming and increasing anthropogenic pressures^13,34,44,45. Oxygen depletion in the hypolimnion not only suppresses microbial CH₄ oxidation, thereby reducing its consumption, but also creates favorable anaerobic conditions that enhance CH₄ production, together leading to elevated methane emissions^46,47,48. Given that reservoirs are already recognized as hotspots of methane release, the amplification of emissions under deoxygenated conditions poses significant feedback to the global climate system⁴⁹. Thus, addressing reservoir deoxygenation is not only essential for local water security but also for mitigating global greenhouse gas budgets.

Although current mean DO concentrations in most reservoirs remain above the generally accepted ecological threshold of 5 mg/L⁵, the observed rates of decline suggest that some regions—particularly in Asia and North America—may approach critical oxygen limits within the coming decades if current trends persist. Given the considerable spatial variability in reservoir morphology, climate sensitivity, and management regimes, we refrain from providing a global first-order estimate of when these thresholds might be crossed, as such extrapolation could introduce substantial uncertainties. Nevertheless, our findings underscore the urgency of developing regional-scale models to predict oxygen depletion timelines and inform early adaptive management strategies for maintaining ecological integrity and water quality under ongoing climate and anthropogenic pressures.

In this context, our findings align strongly with multiple Sustainable Development Goals (SDGs). By enabling accurate long-term assessments of reservoir oxygen status, our results contribute to improved water quality management (SDG 6: Clean Water and Sanitation) and protection of aquatic ecosystems (SDG 14: Life Below Water). The observed climate sensitivity of DO emphasizes the urgent need for climate adaptation strategies (SDG 13: Climate Action), while the direct link to greenhouse gas emissions integrates reservoir management into global carbon mitigation efforts. Reliable monitoring also supports sustainable agriculture (SDG 2: Zero Hunger) by improving the understanding and management of irrigation water quality, fosters technological innovation (SDG 9: Industry, Innovation, and Infrastructure) in environmental monitoring, and strengthens international cooperation (SDG 17: Partnerships for the Goals) by providing a global benchmark for reservoir management under climate change. Overall, the long-term perspective presented here highlights reservoir deoxygenation as an emerging global water quality challenge with far-reaching implications. Incorporating oxygen dynamics into reservoir design, operation, and governance frameworks is essential for sustaining aquatic biodiversity, securing safe water supplies, and contributing to global sustainability agendas.

This study develops a robust method for monitoring global reservoir deoxygenation by integrating multi-decadal Landsat observations with machine learning. Our novel dissolved oxygen (DO) retrieval method achieve an R² is 0.73 and RMSE is 1.23 mg/L in testing set. Using this method, we reveal that 74% of the global major reservoirs exhibited widespread deoxygenation over the past four decades. The widespread decreasing of reservoirs DO drive primarily by climate warming (~ 46%) and human perturbations (~ 31%), including agricultural runoff and land use changes. Our analysis also reveals pronounced spatial heterogeneity in global reservoir oxygenation, with consistently higher DO concentrations in cold northern and polar systems and lower, more variable conditions in tropical and subtropical regions. Temporally, reservoirs undergo a significant global decline in surface DO at a mean rate of − 0.13 mg/L per decade, a pace exceeding that observed in natural lakes, rivers, and oceans. Seasonal fluctuations further amplify this variability, with contrasting hemispheric patterns driven by temperature-dependent solubility and biological oxygen demand. Our satellite-based approach developed here provides a scalable solution for long-term reservoir DO monitoring, underscoring both the climatic sensitivity and ecological vulnerability of reservoir oxygen dynamics. Future advancements in remote sensing, coupled with expanded in-situ validation, will further enhance our ability to track and predict these changes, supporting evidence-based conservation efforts in a warming world.