River discharge estimation from satellite observations
Volumetric river flow can be inferred from space using width, surface velocity, slope and elevation as hydraulic proxies. Accuracy depends heavily on calibration data and unknown bathymetry, but the SWOT mission is reshaping what is possible in ungauged basins.
Sensors
- SWOT KaRIn (Ka-band Radar Interferometer): Simultaneously measures water-surface elevation, width and slope at 10–100 m posting across swaths up to 120 km wide. Nominal revisit 21 days; elevation precision around 10 cm for reaches longer than 10 km. The first sensor capable of deriving slope directly from space at river scale.
- Sentinel-1 SAR (C-band, 5.6 cm wavelength): 6-day revisit over Europe, 12 days globally. IW mode provides 10 m ground range resolution. Used for water-surface width delineation and, via along-track interferometry or image-pair offset tracking, surface velocity estimation. Cloud-independent, day-night capable.
- Sentinel-2 MSI (multispectral optical): 10 m resolution in visible and near-infrared bands. NDWI or MNDWI indices delineate bankfull and low-flow widths clearly on rivers wider than roughly 30 m. 5-day revisit at mid-latitudes with both satellites. Blocked by cloud, so width time series have irregular gaps in tropical basins.
- ICESat-2 ATLAS (photon-counting lidar): Measures water-surface elevation along discrete ground tracks with a vertical precision of a few centimetres. Repeat cycle 91 days, so it is not a monitoring tool in isolation, but it provides high-quality absolute elevation tie-points useful for calibrating SWOT or altimetry-based slope estimates.
What a satellite actually measures, and what it does not
Discharge is a volume per unit time: width multiplied by depth multiplied by velocity, integrated across a cross-section. Satellites can observe width and, with some effort, surface velocity and water-surface elevation. Depth is invisible from orbit. That single fact shapes every method described here.
Width is the most accessible proxy. On rivers wider than about 30 m, Sentinel-2 MNDWI mapping reliably tracks the wetted channel boundary. Narrower rivers fall below the pixel floor and require SAR or very-high-resolution commercial imagery. Width alone is a weak predictor of discharge unless the channel geometry is known, because the same width can correspond to very different depths depending on whether the bed is sand, gravel or bedrock.
Rating curves: turning proxies into flow estimates
The standard approach is a hydraulic rating curve, an empirical or physics-informed relationship between an observable (width, elevation or their combination) and discharge. At gauged stations, the curve is fitted to concurrent in-situ measurements. At ungauged sites, it must be transferred from nearby gauges, inferred from hydraulic geometry theory, or derived from a calibrated hydrodynamic model.
At-many-stations (AMS) methods exploit the fact that discharge is conserved along a reach. If SWOT or Sentinel-1 observes width and slope simultaneously at multiple cross-sections, the Manning equation constrains the unknown roughness and bathymetry through the consistency requirement that all sections must carry the same flow. Published studies using simulated SWOT data have reported discharge errors in the range of 20–35% for large rivers under this framework, though performance degrades sharply on rivers narrower than roughly 100 m or where floodplain storage is significant.
Velocity adds information. Along-track SAR Doppler shift, exploited by systems such as Sentinel-1 in specific acquisition modes, can detect surface velocities above roughly 0.3 m/s on wide, smooth reaches. Image-pair correlation (tracking surface features between two acquisitions separated by seconds to days) works on braided or debris-laden rivers where texture is high. Neither method penetrates to the bed, so depth remains the residual unknown.
SWOT: simultaneous width, elevation and slope from one pass
Launched in December 2022, the Surface Water and Ocean Topography mission is the first satellite designed explicitly to measure rivers as hydraulic systems rather than as incidental targets. Its Ka-band radar interferometer produces water-surface elevation maps at 10–100 m posting across a 120 km swath, with a nominal vertical precision of about 10 cm for river reaches exceeding 10 km in length. Crucially, it measures slope directly, not just point elevation.
Slope matters because it drives flow. A reach with a steep slope and narrow width may carry more discharge than a wide, flat reach. Prior altimetry missions (Envisat, ICESat, CryoSat-2) provided point elevations along fixed tracks; deriving slope required interpolating between tracks separated by tens to hundreds of kilometres, introducing large errors. SWOT collapses that problem within a single overpass.
The residual uncertainty is bathymetry. SWOT sees the water surface, not the channel bed. The effective hydraulic depth must be estimated from ancillary data: field surveys, acoustic Doppler measurements, or modelled cross-sections. In data-scarce basins this is the dominant error source, and no orbital sensor currently resolves it. Airborne lidar during low-flow periods can map dry-bed topography, but that is a costly field campaign, not a satellite product.
Where the method works and where it breaks
Performance is strongly correlated with river size. For rivers wider than 100 m, the combination of Sentinel-1 width time series and SWOT elevation offers genuinely useful discharge estimates, particularly for detecting relative change (flood peaks, seasonal cycles, inter-annual anomalies) even when absolute accuracy is moderate. For rivers between 30 and 100 m, Sentinel-2 width mapping degrades and SAR speckle becomes a larger fraction of the signal. Below 30 m, satellite-only discharge estimation is not operationally reliable with current open-access sensors.
Cloud cover breaks optical time series. In the Congo, Amazon and Mekong basins, cloud-free Sentinel-2 observations can be separated by weeks during the wet season, which is precisely when discharge variability is highest. SAR fills some of that gap but does not fully substitute for optical width mapping in vegetated floodplains where forest canopy obscures the channel edge.
Ephemeral and flashy rivers in arid regions pose a different problem: the event may peak and recede within hours, faster than any current satellite revisit. SWOT's 21-day cycle is irrelevant to a Sahelian wadi in spate. For those systems, satellite data is more useful for mapping channel planform and estimating bankfull capacity than for tracking individual events.
Calibration, validation and the sparse-gauge problem
The global network of river gauges has been contracting since the 1980s. The Global Runoff Data Centre (GRDC) holds records for roughly 10,000 stations worldwide, but active stations are concentrated in North America, Europe and parts of Asia. Large portions of central Africa, the Amazon headwaters and Central Asian dryland basins have gauge densities below one station per 50,000 km² of catchment.
This is precisely where satellite discharge estimation is most needed, and also where calibration data to validate it is scarcest. The result is a circular dependency: the method needs ground truth to be trusted, but the basins that most need the method have the least ground truth. Partial solutions include using historical gauge records (even discontinued ones) to fit rating curves, transferring hydraulic geometry relationships from physically similar gauged basins, and using citizen-science water-level observations. None of these is fully satisfying.
Satellize applies open-constellation analytics to hydrological monitoring in data-sparse environments. For clients managing infrastructure or water rights in ungauged basins, we scope the achievable uncertainty honestly before committing to a monitoring design, and we flag where a targeted field campaign would reduce satellite-derived error more cost-effectively than additional imagery.
Practical outputs and what to expect from them
A satellite-derived discharge product for a gauged large river can achieve relative errors of 15–25% at daily to weekly timescales when a good rating curve exists. That is sufficient to detect drought onset, track seasonal filling of downstream reservoirs, or identify anomalous abstractions upstream. It is not sufficient for flood-warning thresholds that require accuracy within a few percent.
For ungauged rivers, the honest expectation is detection of relative change at the 30–50% level, useful for trend analysis and anomaly flagging rather than absolute water accounting. SWOT is expected to improve this as its data archive grows and the research community refines AMS inversion methods, but the mission is still in its early operational phase and published validation results for diverse river types remain limited.
Deliverables from a satellite discharge monitoring programme typically take the form of reach-averaged discharge time series (CSV or GIS-ready format), anomaly alerts when flow departs from a seasonal climatology by a defined threshold, and periodic summary reports covering hydrological conditions across a defined basin. The time series can be extended backwards using Sentinel-1 and Landsat archives, which in some cases reach to the early 1990s for width-based proxies.
Typical figures
| Spatial resolution (width mapping, optical) | 10 m (Sentinel-2); reliable on rivers wider than ~30 m |
| Spatial resolution (width mapping, SAR) | 10 m IW mode (Sentinel-1); detects edges on rivers wider than ~20–30 m |
| SWOT elevation precision | ~10 cm for reaches >10 km; degrades on narrower or rougher channels |
| Revisit (Sentinel-1) | 6 days over Europe, 12 days globally (single satellite) |
| Revisit (Sentinel-2) | 5 days at mid-latitudes (both satellites combined); cloud-limited in tropics |
| Revisit (SWOT) | 21-day exact repeat; fast-sampling phase provided ~1-day revisit during commissioning |
| Minimum river width for reliable discharge proxy | ~100 m for good accuracy; 30–100 m marginal; below 30 m not reliable with open sensors |
| Discharge error range (gauged, large rivers) | 15–25% typical; 20–35% for AMS methods on ungauged reaches (published simulation studies) |
| Archive depth | Sentinel-1 from 2014; Sentinel-2 from 2015; Landsat width proxies back to early 1990s |
| Frequency / band | Ka-band (SWOT, ~35 GHz); C-band SAR (Sentinel-1, 5.405 GHz); VNIR/SWIR optical (Sentinel-2) |
Analytics Satellize can run
| River width time series | MNDWI or NDWI thresholding on Sentinel-2; active-contour edge detection on Sentinel-1 SAR | Monthly GIS layer of bankfull and low-flow channel width per reach; CSV time series per cross-section |
| Relative discharge index | Empirical power-law rating curve fitted between satellite-derived width and available gauge records | Weekly discharge index time series with uncertainty bounds; anomaly alert when index exceeds seasonal envelope |
| Water-surface slope | SWOT KaRIn elevation differencing along reach; ICESat-2 tie-point calibration where tracks intersect | Reach-averaged slope map updated per SWOT overpass; input layer for Manning-based discharge inversion |
| Surface velocity field | SAR image-pair offset tracking (intensity cross-correlation) or along-track Doppler analysis on Sentinel-1 | Surface velocity raster per acquisition pair; integrated with width to produce continuity-based discharge estimate |
| At-many-stations discharge inversion | Manning-equation AMS framework using simultaneous SWOT width, elevation and slope observations across multiple cross-sections | Discharge estimate with documented uncertainty per reach; PDF summary report per basin per SWOT cycle |
| Historical flow reconstruction | Landsat-derived width proxy back-cast using established rating curve; gap-filling with Sentinel-1 during cloud periods | 30-year discharge proxy time series for selected reaches; trend analysis and low-flow/high-flow percentile report |
Who does the work
We can get this done for you. Satellize runs its own analyst desk and a strong science team. You do not buy a data feed and work out what it means; our people source the imagery, run the analysis described on this page, and hand you the answer with its confidence limits stated. Discuss this requirement.