Mapping pan-Arctic riverine particulate organic carbon from space (1985 to 2022)

Abstract

Carbon release from high-latitude regions is intensifying, with profound consequences for the Arctic carbon cycle. Here, we provide a comprehensive analysis of changes in fluvial particulate organic carbon (POC) concentrations (C_POC) and fluxes (F_POC) during ice-free seasons of pan-Arctic rivers from 1985 to 2022 on the basis of satellite observations. Across 578,000 kilometers of river length, 18% of the total length experienced a significant increase in C_POC, which exceeds the 11% that exhibited declines, resulting in a net rise. Most increases occurred after 2005, contributing to a 12.6% (0.49 teragrams per year) increase in total F_POC to the Arctic Ocean between 1985 to 2005 and 2006 to 2022. Regional contrasts highlight distinct possible drivers: increased precipitation in the North American Arctic and atmospheric warming in the Eurasian Arctic. Deepening of the permafrost active layer is also significantly correlated with C_POC increases. These findings highlight climate-driven fluvial POC export as a key contributor to the Arctic carbon budget and provide a high-resolution, satellite-based dataset that can inform carbon cycle models and data assimilation efforts.

Satellite observations reveal increased delivery of particulate organic carbon by pan-Arctic rivers over the past four decades.

INTRODUCTION

The Arctic is undergoing rapid climate change, warming four times faster than the global average—a phenomenon known as Arctic amplification (1). Meanwhile, annual precipitation in the Arctic has increased by ~10% compared to the mid-20th century (2). This warmer and wetter climate has driven widespread permafrost thawing across the region, substantially affecting the global carbon cycle, as about half of the world’s soil carbon is sequestered in high-latitude permafrost (3, 4). Numerous studies have documented that extensive permafrost degradation in the Arctic (5–8) has led to substantial alterations in regional hydrological processes and the associated carbon transport from rivers to the oceans over the past few decades (9–14). These changes disrupt regional biogeochemical cycles and further contribute to global warming (15–19). Therefore, understanding the carbon transport processes of Arctic rivers is crucial for comprehending regional biogeochemical dynamics and assessing their potential impacts on Arctic ecosystems.

Many studies have focused on quantifying fluvial organic carbon fluxes in the Arctic. Starting in 2002, sampling efforts by the Pan-Arctic River Transport of Nutrients, Organic Matter, and Suspended Sediments (PARTNERS) project and the Arctic Great Rivers Observatory (Arctic-GRO) provided important insights into the fluxes of particulate organic carbon (POC) and dissolved organic carbon (DOC) and nutrient cycling within large Arctic river basins (20, 21). Samples were mainly collected at the mouths of the Yukon, Mackenzie, Yenisey, Ob’, Lena, and Kolyma rivers. Measurements limited to a few stations at river mouths are, however, unable to document the fate of organic carbon as it is transported through the river systems. Furthermore, previous estimates of the total fluvial carbon flux for the entire pan-Arctic were often upscaled using the data from these sampled major rivers (22), introducing large uncertainties resulting from the distinctive hydrological and geomorphological response of smaller rivers (23). The relatively low sampling frequency and the short time frame of available in situ data have prohibited the characterization of longer-term trends and any underlying responses to climate change. Such limitations could potentially be addressed through satellite remote sensing, given the frequency of observations worldwide. Recent studies have successfully mapped the POC and DOC concentrations and the carbon fluxes in various Arctic or other deltaic systems (24–27). Notably, riverine POC accounts for ~15% of the total organic carbon flux into the Arctic Ocean (28). However, a comprehensive study to quantify and characterize fluvial carbon transport over the entire pan-Arctic region is currently unavailable.

We used satellite observations from 1985 to 2022 to generate the first long-term, pan-Arctic dataset of POC, enabling a comprehensive examination of changes in fluvial POC concentrations and fluxes across Arctic rivers. Two important questions regarding the pan-Arctic carbon cycle are addressed: (i) where and how have fluvial POC concentrations within the pan-Arctic region changed over the past four decades; (ii) have POC fluxes into the Arctic Ocean increased or decreased during this period.

RESULTS

Mapping POC concentrations and fluxes of pan-Arctic rivers

We used 1.1 million Landsat satellite images captured between 1985 and 2022 to characterize the spatiotemporal patterns of fluvial POC concentrations (C_POC, in μg/liter) along river channels across the pan-Arctic region, as well as the associated riverine fluxes (F_POC, in Tg/year) delivered to the Arctic Ocean (Figs. 1 and 2). By compiling an extensive dataset of in situ POC measurements, we developed a robust algorithm specifically tailored for estimating C_POC in pan-Arctic rivers using satellite observations (figs. S1 and S2; see Materials and Methods). Our analysis focused on rivers wider than 120 m, given the 30-m spatial resolution of Landsat imagery and the increased uncertainty in narrower channels. The algorithm was applied to 10,470 river reaches—defined as river segments between two adjacent confluences (29), with reach lengths ranging from 1.4 to 305.9 km (fig. S3). In total, the analyzed river network spans 5.78 × 10⁵ km in length and drains ~90% of the entire pan-Arctic region. The pan-Arctic region defined in our study is similar to that used by Feng et al. (30), and we expanded it to include adjacent river basins with extensive permafrost coverage as a result of their geographic and climatic similarities to the pan-Arctic region (31).

Fig. 1. Patterns and trends of fluvial POC concentrations (C_POC, in μg/liter) in the pan-Arctic region from 1985 to 2022.

(A) Mean annual C_POC over the examined period. (B) Long-term trends of C_POC for different river reaches. The bar chart shows the proportions of different trends of no change, increase, and decrease, derived using the MK test (P > 0.05 represents no change). (C) Scatterplots showing the C_POC variation across various river widths. (D) Violin plots of the relative changes in C_POC (in percent/year) within the two periods: 1985 to 2005 and 2006 to 2022. Each violin’s width represents probability density, with white dots marking median values. The black rectangle spans the first to third quartiles. Note that the data during 1991 to 1995 were excluded because of interference caused by the Mt. Pinatubo eruption.

Fig. 2. Long-term trends of annual POC fluxes (F_POC, Tg/year) in pan-Arctic Rivers from 1985 to 2022.

(A) Trends of F_POC for 200 rivers and ice sheet outlets, derived using the MK test. Circles with and without outer boundaries indicate significant (P < 0.05) and nonsignificant (P > 0.05) changes, respectively, with the circle size representing the mean annual water discharge. (B) Change of the total F_POC drained into the Arctic Ocean between the two periods: 1985 to 2005 and 2006 to 2022, with error bars labeled (see Materials and Methods). Note that the data during 1991 to 1995 were excluded because of interference caused by the Mt. Pinatubo eruption. NA, North America; EA, Eurasia.

We estimated the C_POC for studied river reaches within the observation period and examined any corresponding long-term trend. We calculated the F_POC and associated changes for 200 rivers and also ice sheet outlets situated in Greenland (32, 33) that discharge into the Arctic Ocean. We did not perform the F_POC calculations for rivers that are located within our defined pan-Arctic region but do not discharge into the Arctic Ocean. The water discharge of these rivers and ice sheet outlets accounts for 85% of the total for pan-Arctic rivers. The F_POC was determined by integrating information on water discharge and C_POC for the river mouth over a year (see Materials and Methods). Our analysis only relates to the ice-free seasons, because the substantial difference in optical properties between water and ice from satellite observations could cause large uncertainties in POC estimates (34). Nevertheless, the contribution of the ice season to riverine POC delivery should be minimal, as river discharge and C_POC are both very low during this period.

Patterns and long-term trends

The length-weighted mean C_POC for pan-Arctic rivers is 1221 ± 33 μg/liter (uncertainty from remote-sensing-derived POC estimates, see Materials and Methods), with pronounced site-specific patterns (Fig. 1A). Extremely high C_POC values (>5000 μg/liter) were found along river reaches in northwestern Canada and western and eastern Russia, while low POC concentrations (<500 μg/liter) were identified in rivers in northeastern Canada. From 1985 to 2022, we detected significant increases in C_POC in 18% (or 1.04 × 10⁵ km) of the examined river length (Fig. 1B), longer than the rivers exhibiting a decreasing trend (11% or 0.64 × 10⁵ km). As a result, the length-weighted mean C_POC for the pan-Arctic increased by 11.5 ± 0.8% (140 ± 6 μg/liter) during the past four decades. Spatially, increasing trends occurred across the entire pan-Arctic, with a higher proportion in North America (25.5%) than in Eurasia (14.4%). Meanwhile, C_POC remained relatively unchanged along 71% of the total river length.

Prominent increases in C_POC were primarily observed after 2005, with abrupt changes detected in 65% of river reaches on the basis of long-term change point detection (see Materials and Methods and fig. S4). During 1985 to 2005, although the length-weighted mean C_POC showed an overall increasing trend (0.15 ± 0.002%/year) for all the river reaches (Fig. 1D), the length of river reaches with downward trends (58%) was even greater than those with upward trends (42%). However, during 2006 to 2022, the increasing rate of C_POC was 3.3 times greater (0.50 ± 0.09%/year) because of increases predominating in 58% of the river reaches.

Considerable disparities were also noted among distinct river reaches within the same river, with variations evident in both C_POC levels and their respective trends. In general, smaller rivers exhibited higher C_POC than larger rivers (Fig. 1C). These patterns corroborate previous findings suggesting that upstream sediments are largely deposited and stored in channels and floodplains rather than transported downstream (35–37). The substantial spatial differences are further highlighted by the observation that trends in C_POC at river mouths may not accurately reflect changes in C_POC for the entire river. Our data showed that the C_POC trend at the river mouth could deviate from or even contradict the trend for the length-weighted C_POC of the entire river basin (fig. S5A). For instance, we found that C_POC trends along different segments of the Yenisey River were notably influenced by dam trapping (fig. S5C). In addition, in the Lena River Delta, the Arctic’s largest delta, significant increases in C_POC were observed [consistent with previous in situ studies (5, 11)], contrasting with varied changes in its main streams (fig. S5D).

Length-weighted C_POC trends across six major Arctic basins also reveal pronounced spatial heterogeneity (fig. S5). The Ob’ River shows the most rapid increase (11.77 ± 0.32 μg/liter per year), followed by the Yukon River (11.04 ± 0.68 μg/liter per year) and Mackenzie River (2.86 ± 0.56 μg/liter per year). In contrast, the Yenisey River exhibits only a marginal rise (0.79 ± 0.17 μg/liter per year), while the Lena River shows a slight decline (−1.41 ± 0.06 μg/liter per year). The Kolyma River displays a modest upward trend (2.28 ± 0.04 μg/liter per year), although not statistically significant (P > 0.05). These basin-level assessments, enabled by our headwater-to-estuary POC mapping framework, provide the first spatially explicit quantification of long-term fluvial POC dynamics across the pan-Arctic domain.

The mean F_POC into the Arctic Ocean during the study period was 4.13 ± 0.01 Tg/year, contributing 2.1 to 3.0% of the total global F_POC (140 to 197 Tg/year from global models) (38, 39). Of the 200 rivers and ice sheet outlets examined, 56 (28%) showed significant (P < 0.05) increasing trends in annual POC flux between 1985 and 2022, while 15 (8%) displayed significant decreasing trends, and 129 (64%) remained stable. The increased F_POC was mainly discharged into the Barents Sea, Hudson Bay, and the Bering Sea (Fig. 2A). Moreover, the F_POC from several ice sheet outlets in southwestern and northeastern Greenland also increased substantially (fig. S6), albeit with smaller annual flows, leading to a rate of increase of 0.58 ± 0.29 × 10⁻³ Tg/year² (or 0.6 ± 0.3%/year, P < 0.05) in Greenland F_POC. Among the six major rivers in the Arctic, the Yukon River exhibited a significant increase in F_POC, while the other five remained relatively stable throughout the study period. In total, a net increase of 0.49 ± 0.09 Tg/year (12.6 ± 2.4%) in F_POC was observed across the pan-Arctic region from 1985 to 2005 to 2006 to 2022 (Fig. 2B).

DISCUSSION

Impacts of climate change

To examine the potential impacts of climate change on the recent trends in Arctic fluvial POC, we first analyzed the sensitivity of C_POC to changes in surface air temperature by using climate elasticity models (see Materials and Methods). We found significant (P < 0.05) positive correlations between changes in mean annual length-weighted C_POC and changes in air temperature in 97 river basins (Fig. 3A), demonstrating the positive effects of atmospheric warming on the recent C_POC increase over the pan-Arctic region. Such results agreed well with previous local studies, where warmer temperatures have led to permafrost instability and excessive carbon release (5, 40–43). Moreover, another potential consequence of warming is that elevated water temperatures can foster greater autochthonous production within rivers (44, 45), thereby contributing to elevated C_POC levels. In addition, we observed significantly increased precipitation over substantial areas of the pan-Arctic (fig. S7) and, therefore, further analyzed the sensitivity of C_POC to precipitation. Positive correlations were identified between precipitation and fluvial C_POC in 65 river basins (Fig. 3B), indicating that fluvial POC delivery may also be modulated by the wetter climate in the Arctic, which is linked to increased soil erosion and heat transfer (46–51).

Fig. 3. Drivers of C_POC trends.

(A and B) Climate elasticity models (see Materials and Methods) show the river basins where C_POC is significantly sensitive (P < 0.05) to air temperature (A) or precipitation (B). (C and D) Time series of annual average C_POC, precipitation, and air temperature for the Kolyma and Nottaway rivers, with their locations shown in (A) and (B). Pearson correlation coefficients (r) between C_POC and climate variables (precipitation and air temperature) are annotated. The colors of the curves and r correspond to the colors of their respective y axes. (E) Trends in the active layer thickness (ALT) across the pan-Arctic from 1997 to 2021. (F) Correlation between C_POC change rates and ALT deepening rates across all river basins. The gray-shaded area represents the 95% confidence interval of the best-fit line. The inset shows the extent and types of permafrost in the pan-Arctic.

Air temperature and precipitation exhibit distinct regional variability in their impacts on C_POC. In North America, increases in C_POC are primarily attributed to a wetter climate rather than warming, reflecting more sensitive precipitation-C_POC feedbacks in regional river basins (Fig. 3B). In contrast, many Eurasian river basins exhibit a higher correlation and greater sensitivity between air temperature and C_POC (Fig. 3A). The warming rate for the Eurasian Arctic (+0.06°C/year) was double that for the North American Arctic (+0.03°C/year) between 1985 and 2022, whereas increased precipitation is found in more North American Arctic river basins (fig. S7). However, our results also show that the C_POC trends in a river (Kolyma River) can significantly correlate with both air temperature and precipitation (Fig. 3C), and the response of C_POC to these factors may vary temporally within the same river. For example, in the Nottaway River, C_POC variations before and after 2008 were more closely aligned with air temperature and precipitation, respectively (Fig. 3D). This further illustrates that the processes by which air temperature and precipitation influence C_POC are highly complex, which are modulated by various other factors that may change simultaneously, including vegetation coverage, soil mobilization, and permafrost types, among others (40, 52, 53). Therefore, as changes in air temperature and precipitation are also likely intricately linked (54, 55), further isolating their individual effects on C_POC is challenging.

Given that permafrost degradation reflects the combined effects of increasing temperature and precipitation, we further analyzed the correlations between changes in the active layer thickness of pan-Arctic permafrost and C_POC (Fig. 3, E and F). Widespread increases in active layer thickness were observed, and the active layer thickness deepening rate was significantly correlated with C_POC change rates (r² = 0.50, P < 0.01). This relationship is primarily attributed to the positive feedback mechanisms in isolated permafrost regions, further highlighting the critical role of these most vulnerable permafrost changes in modulating river C_POC and potentially altering broader biogeochemical cycles (10, 56).

Our study supports the conclusions of previous investigations, where amplified warming of the polar regions has been shown to be responsible for the recent increase in regional carbon release into the rivers (17, 57, 58). In situ observations have shown that increased thermal erosion is contributing to higher organic carbon flux in some Arctic rivers (57). Furthermore, our findings emphasize the dominant role of precipitation in regulating fluvial POC dynamics in the North American Arctic. While the greening trend in polar regions may offer some benefits by reducing soil erosion and fluvial POC transport across the pan-Arctic (59, 60), these effects are likely outweighed by the combined impacts of a warmer and wetter climate.

Further implications

Using 30-m resolution satellite observations, we mapped the spatial and temporal dynamics of C_POC and F_POC during ice-free seasons across all pan-Arctic rivers. Our estimated total F_POC of 4.13 Tg/year is lower than the calculations based on upscaled data from the six largest rivers reported by the PARTNERS and Arctic-GRO projects (that was 5.8 Tg/year) (22). This difference can be attributed to disparities in observed periods and the exclusion of small rivers (width <120 m) from our study. Our F_POC estimates for individual major rivers also agree well with previous studies (fig. S8). Furthermore, we observed that the C_POC trend at the river mouth (i.e., F_POC) could diverge from, or even be opposite to, that for the length-weighted C_POC for the entire river basin (fig. S5A), underscoring the importance of satellite observations in providing a comprehensive characterization of fluvial POC delivery. Our study also sheds light on the linkage between changes in fluvial POC and active layer thickness in isolated permafrost regions. Given that future projections suggest that the observed degradation patterns in Arctic permafrost are likely to persist throughout this century (61, 62), our findings highlight the need for continued focus on the impacts of permafrost degradation on high-latitude river systems.

We acknowledge that our correlation analysis may be insufficient to characterize fully the impacts of climate changes on fluvial sediment and POC transport, which are often nonlinear, episodic, and spatially independent (10, 63–66). The C_POC could be elevated markedly during a short period after permafrost collapses or landslides-thaws when precipitation or air temperature reaches a required threshold (7, 50). Rising temperatures can trigger various complex processes, including decreased heating degree days, earlier ice-out events, and increased energy transmission to riverbanks resulting from warmer water (10). These changes may affect the feedback relationship between air temperature and C_POC. Similarly, increased flooding, a consequence of a wetter climate, might also influence the correlation between precipitation and C_POC (49, 50, 54). We also analyzed POC fluxes from glacier outlets in Greenland; other pan-Arctic glaciers, mainly along the western coast of North America, do not drain into the Arctic Ocean and were excluded in the flux analysis. Glacial rivers on Arctic islands are narrow and cannot be reliably detected by satellites and thus were not considered in the C_POC analysis here, although their dynamics merit further detailed investigation. In addition, a few river segments experienced dam construction during our study period, yet the impacts on POC transport were inconsistent, with some segments even showing increased transport postdamming. Together, these observations highlight the need for physically based models that incorporate the full range of geomorphological and biogeochemical processes to comprehensively quantify the role of climate dynamics on POC transport in pan-Arctic rivers.

Our results revealed widespread increases in C_POC in pan-Arctic rivers, which have the potential to alter water color, subsequently affecting light penetration and, in turn, the biological productivity within these rivers (67). Notably, the documented rise in riverine POC flux holds substantial implications for the Arctic Ocean, as terrigenous input of carbon and nutrients from rivers and coastal erosion contributes to 28 to 51% of the annual Arctic Ocean net primary production (68). Furthermore, the intensified transport of riverine POC carries far-reaching consequences for carbon dynamics. On the one hand, the elevated POC levels may undergo oxidation, leading to increased emissions of CO₂ (18). On the other hand, this amplified transport of POC has the potential to become sequestered in offshore marine sediments, thereby contributing to geological carbon sinks (18). These multifaceted processes have the potential to influence the trajectories of climate change. The dataset provided here could be integrated into models examining these carbon cycle processes in pan-Arctic regions to facilitate a more comprehensive understanding of how climate change has reshaped polar ecosystems at large.

MATERIALS AND METHODS

Data sources

Landsat images

Landsat surface reflectance products, with a spatial resolution of 30 m, were used to retrieve the POC concentrations of Arctic rivers. A total of 1.1 million Landsat-4/5/7 images captured between 1985 and 2022 were used in this study, and the image products are available through the Google Earth Engine platform. The products have been corrected to remove atmospheric effects by the Landsat Ecosystem Disturbance Adaptive Processing System (LEDAPS) (69). We used the quality assurance (QA) flag associated with each image pixel to remove low-quality observations caused by factors including cloud, cloud shadow, ice, snow, radiometric saturation, and atmospheric opacity (see table S1) (70). Although the nominal temporal resolution of Landsat observation is 16 days, the satellite swaths perpendicular to the orbits have large overlaps in the pan-Arctic regions, providing higher observation frequency superior to 16 days (71). The global surface water occurrence (GSWO) product (72) generated using historical Landsat observations was used to determine the boundaries of the river channels. The GSWO product provides the probability (ranges between 0 and 100%) of water presence for each 30-m resolution pixel grid.

In situ data

In situ–collected fluvial POC concentration data were used to calibrate a quantitative remote sensing retrieval algorithm. We compiled four datasets that recorded pan-Arctic fluvial POC, including the global river chemistry database (GLORICH, https://doi.pangaea.de/10.1594/PANGAEA.902360) (73), the Arctic great rivers observatory (ArcticGRO, https://arcticgreatrivers.org/), the modern river archives of particulate organic carbon (MOREPOC, https://zenodo.org/record/7055970#.Y9nUm3ZByUl) (74), the Arctic Data Center (https://arcticdata.io/catalog/data), and other publications (75–80). We used high-quality records determined by the associated quality flag, totaling 18,787 records.

Climate parameters

ERA5-Land provides a global reanalysis dataset of various climate variables from the 1950s to the present (www.ecmwf.int/en/era5-land), with a spatial resolution of 9 km. We used the daily precipitation and 2-m air temperature to analyze their impacts on fluvial POC.

Coastal river discharge dataset

Monthly river discharge data from pan-Arctic rivers were obtained from the global streamflow index time series dataset (until 2022) (81). Annual discharge datasets of 57 large ice sheet outlets in Greenland were obtained from Overeem et al. (33). We matched these data to the satellite-derived POC concentrations to estimate pan-Arctic POC flux into the ocean.

River networks

Pan-Arctic river networks were obtained from the Global River Widths from Landsat (GRWL) dataset (29), where the river courses were separated into 14,672 reaches. A river reach, the smallest statistical unit in this study, is defined as the segment between two neighboring confluences. The GRWL dataset was further used to identify the centerlines and widths of global rivers. The river basin boundaries provided by the HydroBASINS (www.hydrosheds.org/products/hydrobasins) datasets were used to identify different levels of pan-Arctic rivers and basins.

Permafrost and active thickness layer maps

The circum-Arctic map of permafrost was downloaded from the NASA National Snow and Ice Data Center (https://nsidc.org/data/ggd318/versions/2). It characterizes the extent and classifications of permafrost. The pan-Arctic active layer thickness dataset was accessed from the European Space Agency’s Climate Change Initiative Permafrost project. It provides annual ALT from 1997 to 2021 with a resolution of 0.01°.

Developing a POC retrieval algorithm

Two methods are often used to estimate POC concentrations (C_POC, in μg/liter) from remote sensing observations. The first is based on the empirical relationship between C_POC and satellite reflectivity (26, 82, 83), and the second is based on the intrinsic correlation between C_POC and suspended sediment concentration (SSC) in turbid waters and uses satellite-derived SSC as a basis for estimating C_POC (84). Here, we adopted the first method and directly correlated the surface reflectance of Landsat images and concurrently collected in situ C_POC. This decision was made because the alternative approach—inferring C_POC from satellite-derived SSC—may introduce uncertainties related to spatiotemporal variations in the POC-SSC relationship (i.e., changes in the organic carbon content of suspended sediments) (fig. S1E). Such variations can alter the mass-specific absorption coefficients of sediments (85–87), ultimately affecting the retrieval accuracy; such effects could have been avoided by directly correlating satellite reflectance (i.e., water color) and C_POC (88–90). Our algorithm can be expressed as follows

$$\begin{array}{c}\text{log}\left(C_{\text{POC}}\right)=c_1\times {\text{SR}}_{\text{green}}+c_2\times {\text{SR}}_{\text{red}}+c_3\times {\text{SR}}_{\text{nir}}\hfill \\ +c_4\times \frac{{\text{SR}}_{\text{green}}}{{\text{SR}}_{\text{blue}}}+c_5\times \frac{{\text{SR}}_{\text{red}}}{{\text{SR}}_{\text{blue}}}+c_6\times \frac{{\text{SR}}_{\text{nir}}}{{\text{SR}}_{\text{blue}}}\hfill \\ +{\mathrm{c}}_7\times \frac{{\text{SR}}_{\text{red}}}{{\text{SR}}_{\text{green}}}+c_8\times \frac{{\text{SR}}_{\text{nir}}}{{\text{SR}}_{\text{green}}}+\text{Intercept}\hfill \end{array}$$ (1)

where ${\text{SR}}_{\text{band}}$ is the atmospherically corrected surface reflectance (dimensionless) for a specific Landsat band, and c₁ … c₈ and Intercept are the algorithm coefficients. The rationale underlying the algorithm is as follows: Visible–to–near-infrared (NIR) reflectance is directly related to the absorption and backscattering effects caused by particulate organic matter, such as sediments and phytoplankton. Band ratios are used to minimize the potential influence of variations in particle size and mass concentration across different regions, which can affect the backscattering signal detected by satellites (91). We note, however, that quantitatively isolating the direct physical contribution of each term to POC remains challenging at present. To further mitigate potential residual atmospheric correction errors from LEDAPS (the algorithm used for producing Landsat surface reflectance), we adopted a correction approach following (92), whereby the shortwave infrared (SWIR) reflectance is subtracted from the blue, green, red, and NIR bands for each Landsat image. This approach is based on the principle that water-leaving radiance in the shortwave infrared band is expected to be negligible, thus allowing for the reduction of atmospheric residuals in the visible and NIR bands.

We selected high-quality matching pairs of satellite-in situ measurements to calibrate the coefficients using the following criteria: (i) The time difference between the two independent observations was within ±1 day (a sensitivity analysis using only same-day data yielded comparable accuracy; however, we adopted a ±1-day window to increase sample size and representativeness). (ii) The Landsat observations were of high quality as determined by the associated QA flag (see table S1). In addition, to ensure that the pixel represented water coverage, the modified normalized difference water index [ $\text{MNDWI}=({\text{SR}}_{\text{green}}-{\text{SR}}_{\text{swir}})/({\text{SR}}_{\text{green}}+{\text{SR}}_{\text{swir}})$ ] was required to be greater than 0.05, and the probability of water presence GSWO was required to be higher than 50% (MNDWI >0.05 and GSWO >50%). (iii) The Landsat 3 by 3–pixel windows centered on the field monitoring sites were homogeneous (coefficient of variation <0.15). These tests identified 648 high-quality matching pairs, which were selected to calibrate the algorithm coefficients. The C_POC of these pairs ranged from 31 to 32,060 μg/liter, and their spatial distributions are shown in fig. S1A. To train the algorithm, we applied the Lasso regression procedure; this method automatically selects important features by shrinking the coefficients of less relevant variables, reduces the risk of overfitting by penalizing large coefficients, and enhances the algorithm’s ability to generalize to unseen data (93). The fivefold cross-validation strategy was used to validate the algorithm performance. The calibrated algorithm coefficients are as follows: c₁ = −1.3324, c₂ = 2.1928, c₃ = 4.6253, c₄ = 0.1276, c₅ = 0.1837, c₆ = 0.00004, c₇ = 0.8652, c₈ = 0.1132, and Intercept = 1.4951.

The validation results demonstrated high algorithm performance across four orders of magnitude. The satellite-derived and in situ values of C_POC agreed well, with a median absolute percentage error (MAPE) of 40.21% and an absolute relative error (ARE) of 0.003 (fig. S1B and Eqs. 2 and 3). Such uncertainty levels are comparable to current state-of-the-art algorithms for quantitatively retrieving various water constituents using remote sensing (94, 95)

$$\text{MAPE}=\text{Median}\left(\frac{\mid C_{\text{POC}\_\text{predi}}-C_{\text{POC}\_\text{insitu}}\mid }{C_{\text{POC}\_\text{insitu}}}\right)$$ (2)

$$A\text{RE}={10}^{\text{Median}\left[{\text{log}}_{10}\left(\frac{C_{\text{POC}\_\text{predi}}}{C_{\text{POC}\_\text{insitu}}}\right)\right]}-1$$ (3)

where $C_{\text{POC}\_\text{predi}}$ and $C_{\text{POC}\_\text{insitu}}$ represent the POC concentrations from algorithm prediction and in situ measurement, respectively.

To assess the algorithm’s robustness for long-term POC monitoring, we compared C_POC time series data from satellite retrievals with in situ measurements from hydrological stations on six major rivers: the Kolyma, Lena, Ob’, Yenisey, Mackenzie, and Yukon rivers. These stations provide more than 20 years of POC records (2003 to present) through the Arctic-GRO project, with measurements taken approximately once per month. The comparison of long-term in situ POC measurements with remote sensing–derived data shows high consistency (fig. S2). Remote sensing results effectively capture both the seasonal fluctuations and long-term trends in river POC dynamics. We acknowledge that satellite-derived POC data are only available during the ice-free season, while field measurements can obtain data in ice-covered seasons through ice-hole sampling. However, we believe that this limitation does not significantly affect satellite-derived C_POC and flux evaluations for two reasons: (i) During the frozen period, river C_POC levels are very low (~10 times lower than in summer), and discharge is near zero; (ii) the seasonal Mann-Kendall (MK) test (see later) is not sensitive to systematic data gaps across all studied years, ensuring reliable trend detection. Moreover, we verified that our algorithm demonstrated consistent performance across other rivers, regardless of variations in soil organic carbon content or river width (fig. S1, C and D).

Exploring spatial patterns and long-term trends

We applied the POC retrieval algorithm developed for the study to 1.1 million Landsat-4/5/7 images dating from 1985 to 2022 that cover the pan-Arctic region. We did not include Landsat-8/9 data in this analysis because of significant differences in the relative spectral responses between these newer sensors and their predecessors—especially in the red and NIR bands, which are critical for POC retrieval (see Eq. 1). These spectral discrepancies can introduce substantial inconsistencies in the retrieved POC values and consequently affect the reliability of long-term trend detection.

Our study defined the pan-Arctic region similarly to (9, 30), with an expansion that includes neighboring river basins (such as the Amur River) with extensive permafrost coverage (31). We used the QA flag (see table S1) and two masks to identify the river pixels within each image: (i) To ensure the water presence, we only included pixels with MNDWI >0.05 and GSWO >50%, similar to the method used to select satellite-in situ matching pairs, and (ii) considering potential land adjacency effects (96), we only selected pixels located on the river central lines and excluded rivers with a width <120 m. The river central lines and river widths were determined from the GRWL dataset (29).

We retrieved C_POC values for all river centerline pixels within each river reach. A river reach is defined as the segment between two neighboring tributary confluences, following the convention of (29). We tested the spatial consistency of C_POC within individual reaches and found that the median coefficient of variation was relatively low (9%), likely due to the absence of major lateral inputs or outputs within these segments. On the basis of this, we calculated the mean C_POC for each river reach and subsequently derived monthly mean C_POC time series from 1985 to 2022 (Fig. 1). We selected river reaches with more than 222 valid satellite observations for further long-term change trend analysis; this number represents six observations per year or one observation per month, given the duration of the frozen season over the study region. The mean observational frequency is 11 times per year for the river reaches examined (fig. S3). A total of 10,470 river reaches were examined, and their length ranged from 1.4 to 305.9 km. We further estimated the length-weighted mean C_POC within a drainage basin (i.e., the longer the river reach, the greater the weight in the basin) and conducted similar calculations for length-weighted mean C_POC at the continental (Eurasia and North America) and entire pan-Arctic scales using the following formula

$$C_{\text{POC}\_\text{mean}}=\frac{{\displaystyle \sum _{i=1}^N}{C}_{\text{POC}\_i}\times {\text{Length}}_i}{{\displaystyle \sum _{i=1}^N}{\text{Length}}_i}$$ (4)

where N represents the number of reaches within the basin or pan-Arctic, ${\text{Length}}_i$ represents the length (km) of the ith reach, and $C_{\text{POC}\_i}$ represents the POC concentration in the ith reach.

We estimated the annual POC flux (F_POC, Tg/year) into the Arctic Ocean from rivers and the Greenland ice sheet. We matched the river mouth discharge stations and Greenland ice sheet outlets (33, 97) to the nearest river reaches on the basis of two key criteria: (i) hydrological representativeness—the reach corresponds to the location of the terminal hydrological control station at the lowermost basin outlet, ensuring that it captures the integrated discharge from the entire upstream watershed; and (ii) data availability—sufficient POC observations are available to support reliable flux estimation. A total of 151 river stations and 49 outlets were successfully matched (see Fig. 2). The total discharge from these rivers and outlets represents 85% of that for the pan-Arctic region. The monthly mean POC concentration of the matched river reach (C_POCi) was multiplied by the concurrent monthly discharge (Q_i) and was then accumulated across each year to evaluate the annual F_POC, which can be expressed as follows

$$F_{\text{POC}}={\displaystyle \sum _{i=1}^{12}}{C}_{\text{POC}i}·Q_i$$ (5)

Considering the occasional missing data in the monthly POC time series in certain years, we filled the gap using the two nearest available monthly values from the preceding and subsequent years. Note that some rivers are located within our defined pan-Arctic region but do not flow into the Arctic Ocean (such as the Columbia River in North America); we did not calculate F_POC for these rivers, and our analysis focused solely on examining their C_POC.

We adopted the seasonal MK test (98) to determine the trend of pan-Arctic fluvial C_POC and F_POC over the past four decades. The seasonal MK test performs a nonparametric MK test for individual months, which minimizes the potential impacts of seasonality, missing data, and other artifacts on the multidecadal trends of fluvial POC. The seasonal MK test generates a change rate (Sen’s slope) and P value to determine the trend and its significance. Note that owing to the considerable interference caused by the Mt. Pinatubo eruption on satellite signals between 1991 and 1995, we excluded data from this period in our trend analysis (99).

To examine abrupt changes in pan-Arctic fluvial POC, we applied a change detection algorithm, BEAST (Bayesian Estimator of Abrupt Change, Seasonal change, and Trend) (100), to the time series of monthly mean POC. BEAST first decomposes the time series into two components, trend and seasonality, and then uses the Bayesian method to estimate the probability of abrupt change occurrence in any year (see fig. S4A). We recorded the year associated with the probability peak and considered the data with a probability peak of <0.1 as invalid and excluded it from further analysis. We detected valid abrupt changes for 5915 river reaches, and the histogram of occurrence years is shown in fig. S4B. As most (65%) of the abrupt changes occurred after 2005, we further compared the POC trends across two periods: 1985 to 2005 and 2006 to 2022 (Fig. 1D and fig. S9). Note that the Landsat coverage over the Siberia area is limited during the first period (see the area outlined in red in fig. S9). However, such data limitation has only a minor impact on the overall changes of C_POC within the pan-Arctic region. We performed a sensitivity analysis by excluding the Siberian dataset entirely. The resulting pan-Arctic C_POC trend was +4.31 μg/liter per year (+0.34%/year), closely matching the trend derived from the full dataset (+3.79 μg/liter per year, +0.31%/year). These differences are statistically insignificant, indicating that the absence of Siberian observations does not materially influence the main conclusions.

Uncertainty analysis

We further conducted a sensitivity analysis to ensure the viability of the chosen observational frequency (averaging 11 times/year) in generating reliable trends. We identified 80 hydrological gauging stations distributed across the pan-Arctic region, where extensive long-term datasets (spanning from 7 to 35 years) of SSC were available at varying intervals (ranging from daily to submonthly). We used SSC data instead of POC because of the rarity of high-frequency POC observations and with the added understanding that POC is highly correlated with SSC (39, 86). For each station, we randomly selected 11 observations within each year. Data selections were made to ensure that the time intervals mirrored multiples of 8 days, corresponding to the observation frequency of Landsat satellites. We then compared the temporal trend derived from the original full dataset with that derived from the 11 randomly selected values. The rationale behind this comparison was to capture any differences between these trends as a representation of the uncertainty arising from infrequent Landsat observations. The outcome of this comparison indicated that the trend slopes exhibited a strong agreement between the full dataset and the subsampled dataset (r² = 0.85 and slope = 1.05; fig. S3B). Such an analysis confirmed that an observational frequency of 11 times per year is adequate for characterizing trends in riverine SSC and POC.

To further evaluate whether the uncertainties associated with our POC retrieval algorithm could affect the long-term trends of C_POC and F_POC, we performed the following sensitivity analysis: First, we generated Gaussian-distributed noise, which had a median value of 40.21% (i.e., the same as the MAPE of the POC retrieval algorithm). Second, we added the noise randomly to the C_POC retrievals of all river reaches and estimated the associated F_POC for all examined river mouths and ice sheet outlets. Third, we compared the trends of the original and noise-added data for C_POC and F_POC, with their absolute differences (±error bar) computed as the specified error bounds. The results showed that the data with added noise had the same change directions (increasing or decreasing) as the original data for both C_POC and F_POC, and there were no large differences in the magnitude of the trends (fig. S10, A and B). These results suggest that the MAPE of our POC retrieval algorithm only represents random differences between each in situ and satellite matching pair and could be reduced through spatial and temporal aggregation. We reported the error bars for these key statistics by calculating the absolute differences in statistical results between the noise-added and original datasets. Specifically, the length-weighted mean C_POC for pan-Arctic rivers is 1221 ± 33 μg/liter, the length-weighted mean C_POC for the pan-Arctic increased by 11.5 ± 0.8% (140 ± 6 μg/liter) during the past four decades, and the mean F_POC into the Arctic Ocean during the study period was 4.13 ± 0.01 Tg/year. We also compared our F_POC estimates with those of previous studies of the six major rivers (fig. S8), and these also appeared consistent with some of those studies (22, 23, 78, 101–104). We note that the relatively large differences from some previous estimates could result from the different sample sizes (mean C_POC for the river reach versus point-based samples) or periods. Within this context, we further emphasize the necessity for sustained and consistent remote sensing observations over the historical and future periods in examining the POC dynamics. Also noted is the fact that these compared studies have used different methods to consider ice-covered periods when estimating river POC fluxes, which could explain slight result variations. However, the impact of these periods on POC flux is generally limited because of low river discharge and POC concentrations during such times (22).

Our study excluded the analysis during ice-covered periods because of the limitations of satellite retrievals. During these periods, only a fraction of sediment and POC is transported, primarily because both the discharge and POC concentration are typically low. A previous study indicated that the discharge during ice-covered periods, which span from November to April, accounts for less than 10% of the annual total (105). Consequently, the cumulative export of POC during these ice-covered periods tends to be relatively low because of the lower POC concentration during drier seasons (20, 22). Furthermore, under the influence of global warming, there is a projected increase in ice-free days within pan-Arctic rivers, estimated at ~+6 days/°C (106). This results in a greater availability of open water in recent years. However, we believe that the potential impact of this change on our observed trends is likely to be minimal for two reasons: First, the rates of increase in mean air temperature were +0.06°C/year in the Eurasian Arctic and +0.03°C/year in the North American Arctic. Consequently, the alterations in the number of ice-free days over the past four decades are expected to be 6 to 12 days. This magnitude of change is relatively small, especially when considering the frequency of our satellite observations, which occur every 8 days. Second, the POC flux during the extended ice-free periods is anticipated to be minimal because of the low POC concentration and river discharge during dry seasons.

We also acknowledge that the atmospheric correction algorithm for satellite images is challenging in high-latitude regions. Low solar elevation angles increase atmospheric path lengths and can introduce greater uncertainty into surface reflectance retrievals. To address this, our study used the LEDAPS algorithm, which has been validated and widely applied in high-latitude and inland water studies (69, 107, 108). Furthermore, we applied strict quality control procedures, including cloud masking and filtering of low-illumination scenes, to ensure the reliability of the reflectance data. In addition, our POC inversion algorithm was developed using these reflectance products, meaning that any errors from atmospheric correction are implicitly incorporated into the algorithm coefficients and reflected in the overall algorithm uncertainties. Sensitivity analyses further show that after spatial and temporal averaging, these uncertainty levels do not affect the robustness of our long-term trend analysis (fig. S10).

Analyzing the impacts of climate change

To investigate the climate drivers behind variations in fluvial POC, we evaluated the potential relationships between POC and temperature and precipitation at the subbasin scale. Subbasin boundaries were obtained from the HydroBASINS database (109), which defines global river basins at levels 1 through 12. For our analysis, we used the level 4 basins, with a total number of 194 basins examined. First, we generated annual mean time series datasets for length-weighted C_POC, precipitation, and air temperature for each subbasin from 1985 to 2022. In cases where C_POC data were missing for a given month, we interpolated the missing data by using the two nearest available data from the preceding and following years. We then established climate elasticity models to quantify the POC sensitivity to temperature and precipitation for each basin. Atmospheric warming and changes in precipitation have been shown to be the two major factors that modulate fluvial sediment and POC patterns in the unpopulated study region (10, 47, 50). The climate elasticity (e) of POC is defined as the change percentage in POC divided by the change percentage in the climate variables (110, 111). For example, the elasticity for precipitation (e_P) can be expressed as

$$e_P=\frac{d{C}_{\text{POC}}/C_{\text{POC}}}{\mathit{dP}/P}$$ (6)

When long-term observations are available, a nonparametric approach can be applied to estimate the climate elasticity (110), and e_P can be rewritten as

$$e_P=\frac{∆C_{\text{POC}i}/\overline{C_{\text{POC}}}}{∆P_i/\overline{P}}=\frac{(C_{\text{POC}i}-\overline{C_{\text{POC}}})/\overline{C_{\text{POC}}}}{(P_i-\overline{P})/\overline{P}}$$ (7)

where $∆C_{\text{POC}i}$ and $∆P_i$ are the differences of C_POC and precipitation in the ith year as compared to their long-term mean of values (i.e., $\overline{C_{\text{POC}}}$ and $\overline{P}$ ), respectively. Note that the climate elasticity model is statistically meaningful only when a significant correlation (P < 0.05) was found between the long-term $∆C_{\text{POC}i}/\overline{C_{\text{POC}}}$ and $∆P_i/\overline{P}$ . Then, e_P was estimated as the linear regression slope between the $∆C_{\text{POC}i}/\overline{C_{\text{POC}}}$ and $∆P_i/\overline{P}$ , following the same method as (110, 111). The use of linear regression overcomes the potential problem of numerical instability when the denominator is near zero (i.e., $P_i$ is close to $\overline{P}$ ). We interpreted $e_P$ as the proportional change of POC in response to a 1% increase in precipitation.

Similar to precipitation, we established the elasticity model for air temperature (e_T) as follows

$$e_T=\frac{∆C_{\text{POC}i}/\overline{C_{\text{POC}}}}{∆T_i}=\frac{(C_{\text{POC}i}-\overline{C_{\text{POC}}})/\overline{C_{\text{POC}}}}{T_i-\overline{T}}$$ (8)

where $∆T_i$ is the temperature anomaly in the ith year compared to the long-term mean temperature. Then, e_T was estimated as the linear slope between $∆C_{\text{POC}i}/\overline{C_{\text{POC}}}$ and $∆T_i$ , representing the proportional change of POC in response to a 1°C increase in air temperature.

On the basis of the annual mean time series datasets during 1985 to 2022 for length-weighted C_POC, precipitation, and air temperature, a total of 162 models were successfully established, with 97 for temperature and 65 for precipitation, respectively. More specifically, 33 river basins showed sensitive responses to both air temperature and precipitation, and air temperature or precipitation alone was sensitive in 60 and 35 river basins, respectively (Fig. 3, A and B).

To further investigate the impacts of permafrost degradation on fluvial POC transport, we analyzed changes in active layer thickness across the pan-Arctic region. Using annual active layer thickness maps from 1997 to 2021, we determined the change trends (Sen’s slope, cm/decade) at each 0.01° by 0.01° pixel (Fig. 3E). In addition, we compared basin-averaged rates of active layer thickness change (%/decade) with the corresponding length-weighted C_POC change rates (%/decade) at the basin scale (Fig. 3F). Linear regression was used to assess the relationship between these two variables, excluding basins affected by dam-induced trapping effects from the analysis.

Supplementary Materials

This PDF file includes:

Figs. S1 to S10

Table S1

References