Satellite data assimilation for flood forecasting

Satellite-derived precipitation, soil moisture and river levels can be fed into hydrological forecast models as state updates, sharpening predictions hours before a flood peak. The physics is well understood; the operational challenge is latency.

Sensors

GPM IMERG: Global Precipitation Measurement constellation; IMERG Late Run product delivers half-hourly rainfall estimates at 0.1-degree (~11 km) spatial resolution with roughly 12-hour latency. Early Run cuts latency to ~4 hours at some accuracy cost. Misses orographic enhancement in steep terrain; underestimates convective cells narrower than the grid cell.
SMAP L-band radiometer: NASA Soil Moisture Active Passive; 1.41 GHz passive radiometry retrieves surface soil moisture (top ~5 cm) at ~36 km resolution with 2–3 day global revisit. Level-3 products carry a reported unbiased RMSE of roughly 0.04 m³/m³ over vegetated land. Canopy water and frozen ground degrade retrievals.
Sentinel-1 SAR (C-band): ESA's C-band (5.405 GHz) synthetic aperture radar; Interferometric Wide swath mode gives 250 km swath at 5 × 20 m ground resolution. Change-detection between pre- and post-event passes can proxy soil moisture anomaly and open-water extent. Revisit at mid-latitudes is 6 days per satellite, 3–4 days with both Sentinel-1A and 1B operational.
SWOT KaRIn: NASA/CNES Surface Water and Ocean Topography satellite, launched December 2022; Ka-band Radar Interferometer measures river water-surface elevation at ~10 m cross-track resolution for rivers wider than roughly 100 m, with a 21-day exact-repeat orbit. Provides altimetric river levels for ungauged reaches, filling a critical gap in data-sparse basins.

Why a forecast model needs satellite observations at all

A hydrological model is a set of differential equations representing water movement through a catchment. It is initialised with a state, driven by meteorological forcing, and it drifts. Every model has structural errors, parameter uncertainty and forcing errors that compound over time. In a gauged basin with dense rain-gauge networks, the drift is corrected continuously by telemetry. In the roughly 60 percent of the global land surface that is poorly gauged, the model runs open-loop, and by the time a flood arrives the initial state can be wrong enough to shift the predicted peak by many hours.

Satellite observations do not replace the model. They provide spatially distributed measurements that can be used to correct the model state at intervals, a process called data assimilation. The correction can apply to soil moisture before the rain event (setting the correct antecedent wetness), to the precipitation forcing during the event, and to river levels during the flood wave's propagation. Each satellite source addresses a different part of the error budget.

Precipitation forcing: what GPM IMERG gives you and where it fails

GPM IMERG merges passive microwave retrievals from the GPM constellation with infrared-based estimates from geostationary satellites, calibrated against ground gauges. The result is a near-global, half-hourly rainfall field at 0.1-degree resolution. For large, synoptic-scale rainfall systems over flat terrain, IMERG performs well enough to drive operational forecasts in data-sparse regions. Several national meteorological services in South and South-East Asia have incorporated it into flood-warning chains.

The limits are real and worth stating plainly. At 0.1 degrees, a single grid cell covers roughly 120 km² at the equator. A convective storm cell of 10 km diameter sits entirely within one pixel; its intensity is averaged away. In mountainous terrain, orographic precipitation on windward slopes is systematically underestimated because the passive microwave signal responds to hydrometeors aloft, not to the actual surface accumulation. Studies in the Himalayas and Andes have documented IMERG underestimates exceeding 50 percent on individual events. For small, steep catchments with response times under six hours, this is often a disqualifying error.

Soil moisture as an antecedent state update

The fraction of rainfall that becomes runoff depends heavily on how wet the soil already is. A dry catchment with 30 percent soil moisture can absorb a 50 mm event; the same catchment at 90 percent saturation will route most of it directly to the channel. Assimilating a satellite soil-moisture retrieval before the rainfall arrives is therefore one of the highest-value interventions a forecaster can make.

SMAP's 36 km retrieval is too coarse to resolve field-scale variability, but it captures basin-mean wetness reliably enough to shift ensemble flood-probability distributions in a statistically meaningful way. Several published studies using ensemble Kalman filter (EnKF) assimilation of SMAP into the Variable Infiltration Capacity (VIC) model have shown reductions in soil-moisture RMSE of 15–30 percent relative to open-loop runs, with corresponding improvements in streamflow peak timing. Sentinel-1 backscatter change can sharpen the spatial pattern at 20 m resolution, though converting backscatter to volumetric moisture requires empirical calibration that varies by soil type and crop cover.

River-level assimilation and the SWOT opportunity

Once a flood wave is in the channel, the most direct correction is to assimilate observed water-surface elevation into a hydraulic routing model such as LISFLOOD-FP or HEC-RAS. Radar altimeters on satellites like Sentinel-3 and the older TOPEX/Jason series have provided along-track water levels for large rivers for decades, but their nadir-only geometry means observations are sparse in time and space.

SWOT's KaRIn instrument changes the geometry. By interferometrically comparing two radar antennas separated by a 10 m boom, it produces a swath of water-surface elevation rather than a single point. Rivers wider than about 100 m can be measured to decimetric accuracy. The 21-day repeat means SWOT does not provide the daily updates a real-time forecast chain needs, but it is invaluable for calibrating the bathymetric and roughness parameters of hydraulic models in ungauged reaches, which in turn improves every subsequent forecast. This is the distinction between state updating and model conditioning: SWOT is primarily the latter.

Latency is the constraint that determines operational usefulness

A satellite product that arrives 18 hours after observation is scientifically interesting but operationally useless for a flash-flood warning on a catchment with a 6-hour response time. Every product in a forecast assimilation chain must be evaluated against the lead time the forecast is trying to provide. GPM IMERG Early Run, at roughly 4-hour latency, is borderline usable for medium-sized basins. IMERG Late Run, at 12 hours, is suitable only for larger, slower systems. SMAP Level-3 composites carry a 1–2 day latency in standard processing, which means they inform antecedent-state initialisation rather than real-time updating.

Variational assimilation methods (3D-Var, 4D-Var) minimise a cost function that balances the departure of the model state from observations against the departure from the model prior, weighted by their respective error covariances. They are computationally heavier than EnKF but handle non-linear model dynamics better. For operational flood forecasting, EnKF variants dominate in practice because they are easier to implement incrementally and their ensemble spread provides a natural uncertainty estimate for probabilistic warnings. Neither method rescues a product with unacceptable latency.

What an analytics service can realistically deliver

A practical satellite-assimilation service for flood forecasting has three layers. First, automated ingestion of IMERG Early Run and SMAP retrievals into a configured hydrological model, with EnKF state updates on each new observation. Second, ensemble output expressed as exceedance probabilities at defined gauge points, not deterministic peak forecasts. Third, honest flagging of the conditions under which the system's skill degrades: orographic basins, frozen-ground seasons, convective events below the IMERG resolution threshold.

Satellize runs analytics on open constellations including Sentinel-1 and GPM products, and can integrate commercial SAR tasking for high-priority basins where Sentinel-1 revisit is insufficient. The company's analytics work, including the Tonga crop-estimation programme, is built on the same open-data pipelines that underpin flood-forecast assimilation. Governments commissioning a sovereign early-warning capability should expect a calibration period of at least one full wet season before operational skill can be honestly assessed.

Typical figures

GPM IMERG spatial resolution	0.1 degree (~11 km at equator); half-hourly temporal resolution
GPM IMERG latency	Early Run ~4 hours; Late Run ~12 hours; Final Run ~3.5 months (gauge-calibrated)
SMAP soil-moisture resolution	~36 km (radiometer); 2–3 day global revisit; sensing depth ~5 cm
Sentinel-1 SAR resolution (IW mode)	5 × 20 m ground range; 250 km swath; 6-day single-satellite revisit, ~3 days with two satellites
SWOT KaRIn river measurement	~10 m cross-track resolution; rivers >~100 m width; 21-day exact repeat; decimetric elevation accuracy
Minimum detectable soil-moisture signal (SMAP)	Unbiased RMSE ~0.04 m³/m³ over vegetated land under standard conditions
GPM archive depth	IMERG V06 back to June 2000 (using TRMM-era data); near-real-time from 2014
Assimilation method classes	Ensemble Kalman Filter (EnKF); 3D-Var; 4D-Var; particle filters (research stage)
Typical forecast lead time enabled	6–72 hours depending on basin size, model configuration and product latency
Key known skill limits	IMERG underestimates orographic precipitation; SMAP blind under dense canopy and frozen ground; SWOT 21-day repeat precludes real-time channel updating

Analytics Satellize can run

Antecedent soil-moisture state map	EnKF assimilation of SMAP L3 retrievals into a distributed hydrological model (e.g. VIC or HEC-HMS); Sentinel-1 backscatter change used to downscale spatial pattern	GIS raster layer of basin soil-moisture percentile, updated on each SMAP overpass; ingested automatically by forecast model
Probabilistic flood-peak forecast	Ensemble hydrological model driven by IMERG Early Run precipitation; EnKF state updates; ensemble spread expressed as exceedance probability at gauge points	JSON feed of exceedance probabilities (10th, 50th, 90th percentile peak flow) at defined locations, updated every 6 hours
Hydraulic model parameter conditioning	SWOT KaRIn water-surface elevation used to calibrate Manning's roughness and channel bathymetry in a 1D/2D hydraulic routing model via variational parameter estimation	Calibrated model parameter set with documented uncertainty bounds; delivered as model configuration files
Precipitation-forcing quality flag	Automated comparison of IMERG grid-cell estimates against any available gauge telemetry; terrain-slope and aspect analysis to flag orographic-underestimate risk zones	Per-timestep quality mask overlaid on IMERG forcing field; alert when flagged cells exceed a defined fraction of the catchment area
Forecast skill audit	Retrospective verification of model-with-assimilation versus open-loop model against historical gauge records; continuous ranked probability score (CRPS) and hit-rate statistics	Seasonal skill report with honest breakdown of performance by event type, basin zone and data-availability condition
SAR-derived inundation extent for model re-initialisation	Sentinel-1 change detection (pre/post backscatter ratio thresholding) to delineate open water; assimilated as a binary wet/dry mask into the hydraulic model state	GeoTIFF inundation mask at 20 m resolution, timestamped to SAR acquisition, with model re-initialisation trigger

Sources

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.