River water-surface width retrieval for discharge proxy
High-resolution optical and SAR imagery can extract river channel width at scale, which hydraulic geometry converts to a discharge proxy. SWOT now adds simultaneous water-surface slope, tightening the estimate considerably, though ungauged basins remain genuinely hard.
Sensors
- SWOT KaRIn: Ka-band radar interferometer providing water-surface width, elevation and slope simultaneously at 10–100 m cross-track posting; nominal 21-day global repeat but near-daily coverage at high latitudes during the fast-sampling orbit phase. Detects rivers wider than roughly 100 m reliably; narrower channels fall below the interferometric noise floor.
- Sentinel-1 SAR (C-band): 6-day repeat at mid-latitudes in IW mode, 10 m ground range resolution. Water appears as a low-backscatter surface against rougher land, enabling width extraction on rivers wider than about 30 m. Cloud-independent, but wind-roughened water and emergent vegetation can corrupt the water/land contrast.
- Planet SuperDove: 3–5 m resolution, near-daily revisit over most land areas in eight spectral bands including NIR. NIR absorption by water gives a clean water mask on rivers as narrow as 10–15 m. Dependent on cloud-free conditions; tropical basins can lose weeks of usable imagery in wet seasons, which is precisely when high discharge is most interesting.
- Landsat OLI (8 and 9): 30 m multispectral, 16-day repeat per satellite (8-day when both are operational). The Modified Normalised Difference Water Index (MNDWI) using green and SWIR bands is the workhorse for historical width time series back to 1984. Minimum detectable width is roughly 30 m; narrower channels are missed or underestimated.
What width actually tells you about flow
At-a-station hydraulic geometry, first formalised by Leopold and Maddock in 1953, describes how channel width, depth and velocity each scale as power-law functions of discharge. Width is the most visible of the three from space, so it has become the observable of choice. The relationship takes the form W = aQ^b, where the exponent b typically falls between 0.3 and 0.5 for single-thread channels, though braided and anastomosing rivers behave differently and the coefficients vary by lithology, vegetation and sediment supply.
The practical consequence is that a doubling of discharge does not double the width. You are measuring a compressed signal, which means small errors in width extraction propagate into larger fractional errors in estimated discharge. A 10 % width error in a channel with b = 0.4 translates to roughly a 25 % discharge error. That is acceptable for detecting a major flood pulse; it is not acceptable for water-allocation accounting. Knowing this going in shapes what the data can honestly support.
Extracting width from imagery: the method stack
For optical sensors, the standard approach applies a spectral water index (MNDWI for Landsat and Sentinel-2, a NIR threshold for Planet) to produce a binary water mask, then measures the across-channel distance at regular longitudinal intervals. The GRWL (Global River Widths from Landsat) database, published by Allen and Pavelsky, applied this at global scale using Landsat, cataloguing widths for channels wider than 30 m. It is the most widely used public benchmark and a sensible calibration anchor for any new time-series analysis.
SAR-based extraction follows a similar masking logic but exploits backscatter contrast rather than spectral reflectance. Sentinel-1 IW mode at 10 m resolution can resolve channels that Landsat misses, and it works through cloud. The complication is that calm water looks like smooth bare ground in some geometries, and flooded vegetation can look like open water, so a single threshold rarely works across an entire basin without scene-specific tuning.
SWOT changes the geometry of the problem. Its KaRIn instrument produces a raster of water-surface elevation across a 120 km swath (with a 20 km nadir gap). Width is a by-product of the water mask derived from that raster, but the instrument also delivers slope directly. Slope plus width, fed into a Manning-type equation with an assumed or estimated roughness, gives a discharge estimate that does not depend solely on a width-discharge power law. For gauged rivers, SWOT's slope measurements have shown root-mean-square errors in water-surface elevation of around 10 cm at 1 km averaging, per the mission's published science requirements.
The ungauged basin problem is real and should not be glossed over
Hydraulic geometry coefficients a and b must be calibrated. In a gauged basin, you regress satellite-derived widths against concurrent gauge records and the relationship is reasonably well constrained. In an ungauged basin, you have no such anchor. Regional transfer of coefficients from hydrologically similar gauged basins introduces uncertainty that can exceed a factor of two in discharge estimates. Published studies using Landsat-derived widths in ungauged Himalayan tributaries have reported discharge errors of 40–80 %, which is honest and should be stated clearly to any client expecting water-allocation precision.
SWOT partially addresses this by providing slope, which reduces the number of unknowns in the hydraulic equation. But roughness and cross-sectional shape remain unobserved from orbit, and both matter. Bathymetric surveys or airborne LiDAR can fill the shape gap; roughness is harder. The honest position for an ungauged basin is that satellite width retrieval gives a discharge index useful for anomaly detection and trend monitoring, not an absolute flow volume.
Sensor choice is a trade-off, not a hierarchy
No single sensor dominates across all river types and all use cases. SWOT has the most complete physics but its 21-day revisit (outside the fast-sampling phase) means it will miss short-duration flood peaks on smaller catchments. Sentinel-1's 6-day repeat catches more of the hydrograph but its 10 m resolution limits it to channels wider than about 30 m, and wind effects on large water bodies require manual quality control. Planet SuperDove's near-daily revisit and 3–5 m resolution are attractive for narrow channels, but cloud cover in tropical wet seasons can produce gaps of 10–20 consecutive days exactly when flows are highest.
A practical programme typically fuses sensors: SWOT for slope and absolute elevation reference, Sentinel-1 for cloud-independent width at medium resolution, Planet for high-resolution width on specific reaches of interest. Landsat provides the historical baseline stretching back to 1984, which no other system can match for trend analysis. The fusion logic needs to account for the fact that each sensor's water mask has different edge-detection behaviour, so apparent widths are not directly interchangeable without cross-calibration.
Transboundary rivers and the political value of independent measurement
Many of the world's contested water-sharing agreements involve rivers where one riparian state controls the gauging network and another must trust the reported data. Satellite-derived width time series, produced from open constellations by an independent party, provide a verifiable cross-check that does not require in-country access. The measurement is not as precise as a calibrated gauge, but it is independent, auditable and reproducible.
This is where the combination of Sentinel-1 (free, open, archived) and SWOT (NASA/CNES open data) becomes politically significant. A downstream state can commission a retrospective width analysis on a transboundary reach going back years, compare it with reported upstream releases, and identify anomalies worth investigating. Satellize has built similar independent-monitoring logic for the Tonga crop-estimation programme, where the value of the analysis rests partly on its non-alignment with any national reporting interest. The same principle applies here. Speak to the analytics team about configuring a transboundary monitoring reach.
Typical figures
| Minimum detectable river width (optical) | ~10–15 m (Planet SuperDove, NIR threshold); ~30 m (Landsat OLI, MNDWI) |
| Minimum detectable river width (SAR) | ~30 m (Sentinel-1 IW, 10 m resolution with typical water-mask erosion) |
| SWOT KaRIn reliable detection threshold | ~100 m channel width; water-surface elevation RMSE ~10 cm at 1 km averaging (published mission requirement) |
| Revisit cadence | SWOT: 21-day nominal global; Sentinel-1: 6-day at mid-latitudes; Planet SuperDove: near-daily; Landsat 8+9: 8-day combined |
| Spatial resolution | Planet: 3–5 m; Sentinel-1 IW: 10 m; Landsat OLI: 30 m; SWOT KaRIn: 10–100 m posting across 120 km swath |
| Cloud sensitivity | Optical sensors (Planet, Landsat) fully blocked by cloud; SAR (Sentinel-1, SWOT) cloud-independent |
| Historical archive depth | Landsat: 1984–present; Sentinel-1: 2014–present; SWOT: April 2023–present; Planet: 2016–present (variable coverage) |
| Discharge proxy uncertainty (gauged basin) | Typically 10–30 % depending on channel morphology and calibration density |
| Discharge proxy uncertainty (ungauged basin) | Commonly 40–80 % from published studies; suitable for anomaly detection, not volumetric accounting |
| Delivery formats | GeoTIFF water masks, GeoJSON width transects, CSV width-time series, NetCDF for SWOT reach products |
Analytics Satellize can run
| Width time series per river reach | MNDWI or NIR thresholding on Landsat/Planet; backscatter thresholding on Sentinel-1; cross-channel transect measurement at fixed longitudinal intervals | CSV or GeoJSON time series of mean and percentile widths per reach, with cloud-cover flags and sensor provenance |
| Discharge proxy index | At-a-station hydraulic geometry power-law regression (W = aQ^b) calibrated against available gauge records; regional coefficient transfer for ungauged reaches with stated uncertainty bounds | Discharge index time series with confidence intervals, delivered as CSV with uncertainty column; honest flagging of ungauged-basin caveats |
| SWOT reach-averaged slope and width product | Processing of SWOT Level 2 River Single-Pass product (publicly available via NASA PO.DAAC); reach averaging to reduce noise; Manning-equation discharge estimate using assumed or field-measured roughness | Per-reach GeoJSON with width, slope, elevation and Manning-derived discharge estimate, updated each SWOT overpass |
| Multi-sensor fused width raster | Cross-calibration of Planet, Sentinel-1 and Landsat water masks to a common width definition; gap-filling using SAR during cloudy optical periods | Daily or near-daily fused width raster for target reaches, GeoTIFF, with sensor-of-record metadata per pixel |
| Transboundary anomaly detection report | Comparison of satellite-derived width anomalies against reported upstream release schedules; statistical flagging of divergences exceeding two standard deviations of the historical relationship | Monthly PDF report with annotated time-series plots and reach maps; optional alert feed for threshold exceedances |
| Historical trend analysis (1984–present) | GRWL-anchored Landsat MNDWI time series extended forward with Sentinel-2 and Planet; Mann-Kendall trend test on annual median width per reach | Trend map (GeoTIFF) and summary table of reach-level width change rates, with statistical significance scores |
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.