Recent changes to Arctic river discharge

The study addresses the need for a comprehensive, accurate understanding of Arctic river discharge changes under rapid climate warming and Arctic amplification. Arctic rivers, draining 22.1 million km² into the Arctic Ocean, Bering Strait, and Hudson, James, and Ungava Bays, influence ecosystems, societies, and global climate. Publicly available gauge records have declined since the mid-1980s and are spatially biased toward North America, limiting prior assessments which largely relied on gauges. The authors hypothesize that combining satellite remote sensing with hydrologic modeling can deliver a spatially and temporally complete reanalysis of pan-Arctic discharge, improving estimates of total freshwater export, its trends, and dynamics such as spring freshet timing and stream intermittency.

Previous work documented Arctic hydrologic sensitivity to climate change and reported increasing river discharge to the Arctic Ocean, but analyses largely depended on sparse and biased gauge networks with limited temporal completeness and geographic coverage (especially in Eurasia and small streams). Gauge-based discharge estimation is prone to errors during ice conditions and shifting hydraulics, and public access has diminished since the mid-1980s. Recent advances include remote sensing discharge inversion (AMHG/BAM/geoBAM), global hydrography and runoff datasets (MERIT Hydro, GRADES, GloFAS), and increased data/computational resources. Modeled global river discharge products show promise, but fusing models with remote sensing can better constrain dynamics and propagate observational information through space and time. Prior pan-Arctic estimates of total export and its acceleration (e.g., 3.47–10 km³/yr/yr) may be conservative due to incomplete spatial sampling and lack of small-stream representation.

The authors produced the Remotely-sensed Arctic Discharge Reanalysis (RADR), a daily discharge product for 486,493 river reaches across the pan-Arctic (1984–2018), by assimilating satellite-derived discharge into hydrologic model simulations. Data inputs: MERIT DEM for terrain; MERIT Hydro and MERIT-Basin for hydrography and network topology; GRWL for widths; HydroLAKES to mask lake-influenced reaches; gauge data from USGS, ECCC, CEHQ, GRDC, R-ArcticNET, and ArcticGRO for calibration/validation; Landsat 5/7/8 Tier 1 surface reflectance imagery (1984–2018) via Google Earth Engine; TEOW for ecoregions; NSIDC permafrost maps; GRADES runoff/discharge; GloFAS discharge reanalysis; DOR and GOODD for regulation/dams. Width extraction: Reaches narrower than 90 m (Landsat resolution constraints) were excluded from width measurement but included in discharge simulation. Reaches connected to lakes were excluded from inversion. Cross-sections per reach were dynamically spaced based on mean width and kept away from confluences. Using RivWidthCloud in GEE and the revised DSWE classifier, multi-temporal widths (April–November) were extracted from 155,710 Landsat images, yielding 227 million width observations across 2.93 million cross-sections in 131,153 reaches. Mean widths agreed with GRWL with 0.6% relative bias. Remote sensing discharge (geoBAM): The Bayesian AMHG-based geoBAM algorithm inverted discharge from multi-temporal widths at up to 40 cross-sections per reach. Priors (min/mean/max discharge, channel slope) were derived from hydrologic simulations and in situ data where available, updated every five years and aligned to months with satellite data (typically April–November). This produced 9.18 million discharge estimates at 131,153 reaches, with explicit uncertainties. Hydrologic modeling: Two baseline simulations were generated on the MERIT-Basin network using the Hillslope River Routing (HRR) model to route runoff. For GRADES, 0.25° runoff (VIC forced by MSWEP V2 and reanalysis) was area-weighted to sub-catchments then routed. For GloFAS, gridded discharge was converted to runoff, mapped to sub-catchments, and routed. Bankfull widths were estimated from Landsat/GRWL/MERIT where available, otherwise from width–area relationships; reference flow derived from width–discharge scaling. Plane roughness in HRR was calibrated against 1,079 daily gauges (1984–1998) and validated (1999–2018). Reservoir operations were not explicitly simulated; lakes/reservoirs treated as low-slope reaches. Data assimilation: A Local Ensemble Transform Kalman Filter (LETKF) assimilated geoBAM discharge into each baseline (GRADES- and GloFAS-based) using 20-member multiplicative ensembles (coefficients 0.1–2.5) to represent model uncertainty and geoBAM’s posterior uncertainties as observation error. A centered 7-day smoother propagated analysis weights within local hydrologic patches (≤5,000 km²) subject to observation availability. Two assimilated products (from GRADES and GloFAS baselines) were evaluated against gauges; the better-performing assimilated baseline at each gauge was selected and that choice propagated upstream until the next gauge. In ungauged basins, the nearest gauged basin’s better-performing product was selected. This optimal blend formed RADR. Performance was assessed with daily NSE and KGE versus 1,079 gauges; monthly comparisons used an additional 1,076 gauges. Definitions: Total water export equals the sum of annual average discharge at all watershed outlets (NextDownID=0). TCSF is the temporal centroid of discharge March–July. Intermittent reaches are those with average open-water (April–November) zero-flow days (ZFD) between 7 and 230; ZFD defined as days with Q<0.001 m³/s. Regulated reaches have DOR>0 or lie within 500 m of a GOODD dam.

• Total freshwater export: 5,169 km³/yr on average (range 4,656–6,073 km³/yr), with North America contributing 1,768 km³/yr (34.2%) and Eurasia 3,401 km³/yr (65.8%). Average runoff to the Arctic Ocean is 234 mm/yr, 3–17% higher than prior estimates.
• Acceleration (trend in export): Pan-Arctic increase of 11.6±4.6 km³/yr/yr (0.22%/yr; PWMK p=0.03), a factor 1.2–3.3 greater than prior literature (3.47–10 km³/yr/yr). North America: 3.4±1.4 km³/yr/yr (0.19%/yr; p=0.04). Eurasia: 8.2±4.4 km³/yr/yr (0.24%/yr; p=0.13).
• Spatial heterogeneity: 22.7% of 486,493 reaches show significant discharge trends (15.8% with field significance). 82.5% of significant-trend reaches are small streams (drainage area <1,000 km²). Decreases observed in upstream/middle Yukon and Mackenzie; stronger increases in rivers draining to Hudson Bay; decreased discharge in upper Yenisey; increased in central Lena. Ecoregions: deserts/semi-deserts and continuous permafrost show more increases, steppes show decreases.
• Spring freshet timing (TCSF): Eurasia advanced by −0.11±0.03 days/yr (~1.1 days/decade; p=0.01). North America showed no significant trend (0.05±0.03 days/yr; p=0.30).
• Summer intermittency: Open-water zero-flow days (ZFD) in intermittent reaches decreased by −0.31±0.05 days/yr (~3.1 days/decade; p<0.001), indicating intermittent reaches are getting wetter (less frequent drying).
• Model–satellite fusion performance: Assimilating Landsat-derived discharge improved daily skill at 1,079 validation gauges; median NSE and KGE increased by 0.16 and 0.09, respectively, with larger improvements on regulated reaches. RADR accuracy increases with stream order, and improvements over baselines are greatest in smaller rivers. Aggregate satellite adjustment decreased baseline model discharge by ~4%; median relative bias of RADR vs gauges is −2.5% (daily/monthly).

By fusing remote sensing discharge estimates with hydrologic reanalyses, the study delivers a spatially and temporally complete view of pan-Arctic river discharge dynamics, addressing deficiencies of gauge-only or model-only approaches. Findings indicate that both total freshwater export and its acceleration have been larger than previously reported, with strong spatial heterogeneity across basins, ecoregions, and permafrost regimes. Earlier Eurasian spring freshet is consistent with reservoir operations and earlier snowmelt, and the reduction in open-water intermittency suggests wetter conditions in non-perennial streams. The improved dynamics from assimilation highlight the value of satellite data in capturing human regulation effects and small-stream behavior, which dominate Arctic hydrography yet are poorly gauged. These refined estimates have implications for Arctic Ocean freshening, ecosystem productivity, and water resource planning, and they provide more reliable inputs for climate and biogeochemical assessments.

The study presents RADR, a publicly available daily discharge reanalysis for 486,493 Arctic river reaches (1984–2018) created by assimilating 9.18 million satellite-derived discharge estimates into two global hydrologic baselines. RADR reveals higher total freshwater export and stronger acceleration than previously recognized, documents heterogeneous regional trends, and quantifies shifts in freshet timing and reductions in stream intermittency. Assimilation substantially improves model skill, particularly in regulated and small-stream reaches. The approach, leveraging open satellite and model datasets, is scalable to other regions and can enhance global hydrologic understanding. Future work could incorporate explicit reservoir/lake operations, extend assimilation to under-ice periods as feasible with additional sensors, evaluate alternative models/forcings, and apply the framework globally to refine the water cycle’s representation in Earth system models.

• RADR’s water balance derives from model baselines; assimilation primarily improves dynamics rather than correcting all biases, and some reaches retain low absolute skill and sizable biases, especially in regulated systems.
• Reservoir and lake operations were not explicitly simulated; lakes/reservoirs were represented as low-slope reaches, which can affect timing and magnitude.
• Landsat-based width extraction is limited to reaches wider than ~90 m and to open-water months (April–November), excluding under-ice conditions; multi-temporal widths cannot be directly validated at scale.
• geoBAM relies on priors from models and gauges, which can introduce dependencies and potential biases; discharge estimates are sparse in time and uneven across reaches.
• The choice of baseline models/forcings (GRADES, GloFAS) influences assimilated outcomes; different model ensembles could yield different results.
• Spatial selection heuristics (local patch size, smoother window) and gauge-based optimal blending may propagate local choices upstream, potentially misrepresenting ungauged areas.