Snowmelt runoff timing and volume forecasting from satellite snow cover
Spring flood peaks and hydropower revenues hinge on when the snowpack releases. Satellite snow-cover area, degree-day melt models and SMAP soil-moisture state can constrain runoff timing and volume weeks in advance, but forest canopy masking and cloud cover impose real limits that any honest forecast must account for.
Sensors
- MODIS Terra/Aqua: 500 m spatial resolution, daily global coverage (Terra and Aqua combined give two overpasses per day). The MOD10A1/MYD10A1 daily snow-cover products and MOD10A2 eight-day composites are the operational backbone of basin-scale fractional snow-cover area (fSCA) retrieval. Cloud obscures roughly 30–50 % of daily scenes at high latitudes in spring; multi-day compositing reduces but does not eliminate the gap.
- Sentinel-2 MSI: 10 m (visible/NIR) and 20 m (SWIR) resolution, five-day revisit at the equator with both satellites, longer at high latitudes. Band 3 (green), Band 8 (NIR) and Band 11 (SWIR at 1.6 µm) support the Normalised Difference Snow Index and sub-pixel fractional mapping. Cloud cover is the dominant operational constraint; a single usable scene per week is realistic for many boreal catchments in April and May.
- Sentinel-1 SAR (C-band): C-band backscatter drops sharply when the snowpack becomes wet (liquid water fraction above roughly 1–3 % by volume). This wet-snow signature is detectable at 10–20 m IW mode resolution regardless of cloud or darkness, giving a binary melt-onset flag that optical sensors cannot provide under overcast conditions. It does not measure snow depth or SWE directly.
- SMAP radiometer: L-band (1.4 GHz) passive radiometer at 36 km effective resolution, with a two-to-three day revisit. The Level-3 soil-moisture product indicates whether the ground beneath a melting snowpack is frozen, partially thawed or saturated, which controls how much meltwater runs off versus infiltrates. Coarse resolution means it characterises sub-basin soil state rather than field-scale variation.
Why timing matters more than total volume
A catchment that holds 200 mm of snow water equivalent will produce very different flood hydrographs depending on whether it melts over three weeks or ten days. Hydropower operators scheduling reservoir drawdown and flood-risk managers issuing evacuation warnings both need the timing, not just the total. That is the forecasting problem: total SWE is a boundary condition, but melt rate is the variable that determines peak discharge.
Degree-day models remain the workhorses of operational snowmelt forecasting because they require only air temperature and a calibrated melt factor, both of which are available from numerical weather prediction. The satellite contribution is the fractional snow-cover area that tells the model how much of the catchment is still snow-covered on any given day. As the snow line retreats, the effective contributing area shrinks and the model adjusts its melt-rate estimate accordingly. Without that spatial information, a lumped model assumes uniform cover and systematically overestimates late-season runoff.
What a floating roof gives away: reading fSCA from MODIS and Sentinel-2
MODIS fSCA products use the NDSI (Band 4 minus Band 6, normalised) with an empirical correction for sub-pixel mixing. At 500 m, a pixel reported as 40 % snow-covered is a statistical statement about the mixture of bare ground, vegetation and snow within that half-kilometre cell. The SNOWMAP algorithm underlying MOD10A1 applies a forest-correction factor derived from NDVI, but the correction is imperfect and consistently underestimates cover under dense canopy.
Sentinel-2 at 10 m resolves individual forest gaps. Where MODIS sees a blended signal, Sentinel-2 can distinguish a snow-filled clearing from a snow-free one a hundred metres away. The practical workflow is to use MODIS for daily basin-wide fSCA tracking and Sentinel-2 for periodic high-resolution calibration of the MODIS retrieval, particularly in the transition zone where fSCA is falling rapidly and model sensitivity to that number is highest. Cloud permitting, one clear Sentinel-2 scene per fortnight is usually enough to recalibrate the MODIS-based fSCA time series.
The canopy problem: boreal forest hides the snow
Optical sensors see the canopy top, not the ground beneath it. In a closed boreal spruce stand with a canopy closure of 70–80 %, the NDSI signal is dominated by dark needles even when the forest floor holds a substantial snowpack. Published studies comparing MODIS fSCA against ground-based measurements in Scandinavian and Siberian forests report underestimates of 15–30 percentage points in densely forested pixels. For a catchment that is 60 % forested, that error propagates directly into the runoff forecast.
SAR wet-snow detection partially compensates. When the snowpack becomes isothermal at 0 °C and liquid water appears, C-band backscatter from Sentinel-1 drops by roughly 2–4 dB relative to the dry-snow or frozen-ground reference. This signal penetrates the canopy (C-band interacts with branches but passes through better than optical wavelengths interact with needles) and provides a melt-onset date that is independent of cloud and canopy masking. The limitation is binary: SAR tells you melt has begun, not how much snow remains or at what rate it is melting. Combined with the optical fSCA time series, it constrains the start of the melt window even when clouds block every Sentinel-2 acquisition for two weeks.
Soil state as the hidden switch
Two catchments with identical SWE and identical air temperature trajectories can produce peak flows that differ by a factor of two or more if one has frozen ground and the other does not. Frozen soil is effectively impermeable: meltwater that would infiltrate under thawed conditions becomes direct runoff instead. SMAP's L-band brightness temperature is sensitive to the freeze-thaw state of the top 5 cm of soil. The Level-3 freeze-thaw product, at 36 km resolution, provides a daily flag that the runoff model can use to switch between high-runoff and high-infiltration parameterisations.
The coarse resolution of SMAP is a genuine constraint. A 36 km grid cell in a mountainous catchment will average across elevation bands with very different freeze-thaw states. The practical workaround is to use SMAP as a basin-average prior and apply a temperature-lapse-rate correction to distribute freeze-thaw state by elevation band. It is an approximation, but it is better than assuming uniform soil state across a 5,000 km² catchment.
Honest limits of the satellite-only approach
Cloud cover is the most persistent operational problem. In the Norwegian and Finnish spring, a MODIS daily scene is cloud-free perhaps 40 % of the time. Eight-day composites help but introduce a temporal smear exactly when fSCA is changing fastest. Sentinel-2's higher resolution does not help if the sensor is looking at cloud tops.
The approach also does not measure SWE directly. Passive microwave retrievals of SWE (AMSR2, SSMI) saturate above roughly 150 mm in forested terrain and are not covered here. What the satellite-based fSCA approach provides is a spatial constraint on the melt model, not an independent volume measurement. Forecast skill degrades when the initial SWE estimate (from in situ gauges, reanalysis or separate passive microwave retrieval) is poor. In ungauged basins with no ground truth, the uncertainty in peak-flow timing is typically plus or minus several days, and volume uncertainty can exceed 20 %.
Satellize runs operational fSCA time-series extraction on MODIS and Sentinel-2 archives, integrates Sentinel-1 wet-snow onset dates and ingests SMAP freeze-thaw state to drive degree-day runoff models for client catchments. The analytics workflow draws on the same open-constellation approach used in the Tonga crop-estimation programme, adapted for cold-region hydrology.
Putting the forecast to work: what operators actually receive
A useful snowmelt runoff forecast for a hydropower operator is not a single number. It is a probabilistic hydrograph: a family of curves reflecting uncertainty in air temperature over the coming two to four weeks, uncertainty in the initial fSCA estimate, and uncertainty in soil state. The satellite data constrains the spatial distribution of snow cover; the meteorological ensemble constrains the temperature forcing. Together they produce a forecast that can be updated daily as new MODIS scenes arrive and weekly as Sentinel-2 and Sentinel-1 acquisitions refine the picture.
For flood-risk managers, the most actionable output is a melt-onset date with a confidence interval, derived from the first Sentinel-1 wet-snow detection combined with the forecast temperature trajectory. That date, issued two to three weeks before peak flow, is enough lead time for reservoir pre-release decisions and public warning systems in most mid-latitude and sub-Arctic basins. The satellite contribution is not a replacement for gauges and weather models; it is the spatial layer that those models have historically lacked.
Typical figures
| MODIS fSCA spatial resolution | 500 m (MOD10A1/MYD10A1) |
| Sentinel-2 snow mapping resolution | 10 m (NDSI using Bands 3, 8, 11) |
| Sentinel-1 wet-snow detection resolution | 10–20 m (IW mode GRD) |
| SMAP soil freeze-thaw resolution | 36 km effective (L-band radiometer) |
| MODIS revisit | Daily (Terra + Aqua combined); 8-day cloud-composited product also available |
| Sentinel-2 revisit | 5 days at equator; longer at high latitudes; cloud-free scenes may be 1–2 per fortnight in boreal spring |
| Sentinel-1 revisit | 6–12 days per track at mid-latitudes; near-daily coverage possible at high latitudes with multiple tracks |
| SMAP revisit | 2–3 days global |
| Wet-snow detection threshold (SAR) | Liquid water fraction approximately 1–3 % by volume; backscatter drop of roughly 2–4 dB relative to dry reference |
| Archive depth | MODIS from 2000; Sentinel-2 from 2015; Sentinel-1 from 2014; SMAP from 2015 |
Analytics Satellize can run
| Daily basin fSCA time series | MODIS MOD10A1/MYD10A1 NDSI retrieval with cloud-gap filling by temporal interpolation and Sentinel-2 periodic recalibration | GeoTIFF raster stack and basin-average CSV, updated daily during melt season |
| Melt-onset date and confidence interval | Sentinel-1 C-band change detection against dry-season backscatter reference; threshold at 2 dB drop sustained over two consecutive acquisitions | Alert report with onset date, spatial extent of wet-snow area, and uncertainty range |
| Soil freeze-thaw state by elevation band | SMAP Level-3 freeze-thaw product disaggregated by elevation using DEM lapse-rate correction | Daily raster and tabular summary per elevation band, ingested as model forcing |
| Probabilistic runoff hydrograph | Degree-day melt model driven by satellite fSCA and NWP temperature ensemble; Monte Carlo uncertainty propagation over temperature and fSCA error terms | 10th/50th/90th percentile hydrograph as CSV and PDF report, updated daily |
| Canopy-corrected fSCA for forested catchments | Forest-fraction mask from Sentinel-2 NDVI applied to MODIS fSCA; empirical upward correction based on published boreal forest underestimation coefficients | Corrected fSCA raster with per-pixel uncertainty flag, delivered as GeoTIFF |
| Season-end snow-line recession animation | Time-composited Sentinel-2 fSCA scenes stacked into temporal sequence; cloud-masked frames dropped or interpolated | Annotated GIF or MP4 for operational briefings, plus underlying GeoTIFF stack |
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.