Snowpack water-equivalent monitoring for hydro-power forecasting

Passive microwave and optical sensors together estimate the water stored in mountain snowpacks weeks before it reaches a turbine. Neither sensor alone is sufficient, and honest fusion of both is what makes operational hydro forecasting work.

Sensors

AMSR2 on JAXA GCOM-W1: Passive microwave radiometer at 6.9 to 89 GHz. SWE retrieved from brightness-temperature differences between 18.7 and 36.5 GHz channels. Native spatial resolution roughly 25 km for SWE products. Daily global coverage. Retrieval accuracy degrades sharply above roughly 150 mm SWE due to microwave signal saturation; also sensitive to wet snow, forest canopy and ice crusts.
MODIS Terra and Aqua (MOD10A1 / MYD10A1): 500 m daily fractional snow-cover product using the Normalised Difference Snow Index (NDSI) derived from visible and shortwave-infrared bands. Two overpasses per day per sensor reduce cloud gaps. Gaps from persistent cloud cover can reach 10 to 20 consecutive days in maritime mountain ranges, requiring temporal or spatial gap-filling.
VIIRS I-band snow products (VNP10A1): 375 m daily snow-cover product from Suomi-NPP and NOAA-20, extending the MODIS snow-cover record with finer spatial detail. Same NDSI approach; same cloud-masking limitation. Useful for smaller catchments where 500 m MODIS pixels blend snow-free and snow-covered ground.
Sentinel-2 MSI: 10 m visible and 20 m shortwave-infrared bands. NDSI at 20 m resolves individual glacier tongues, avalanche paths and wind-scoured ridges that MODIS cannot separate. Revisit is 5 days at mid-latitudes with both satellites, but cloud persistence in alpine regions means usable scenes can be sparse in winter. Used primarily to calibrate and validate coarser daily products, not as the operational backbone.

What the microwave actually measures, and where it stops

Passive microwave radiometers infer SWE from the scattering effect that snow grains have on upwelling microwave radiation from the ground. Larger grains and deeper snowpacks scatter more strongly at 36.5 GHz than at 18.7 GHz, widening the brightness-temperature difference between channels. AMSR2 exploits this relationship to produce daily, all-weather SWE estimates across entire mountain ranges at roughly 25 km resolution.

The physics imposes a hard ceiling. Once SWE exceeds approximately 150 mm, the snowpack becomes opaque to the relevant microwave frequencies and the brightness-temperature difference saturates. The retrieval reports a number, but it stops tracking reality. In the Himalayas, the Andes and the western North American ranges, seasonal SWE routinely reaches 400 to 800 mm at elevation. This is not a minor edge case; it is the majority of the water that matters most to a hydro operator. Wet snow, which occurs during mid-winter melt events, further confuses the retrieval by mimicking a shallower, denser pack. Ice crusts formed after rain-on-snow events add a third source of bias. Any operational system that relies on passive microwave alone will underestimate peak storage and produce optimistic spring-inflow forecasts.

Optical snow cover: high resolution, conditional on cloud

MODIS and VIIRS measure fractional snow-covered area (SCA) using NDSI, the ratio of green-band to shortwave-infrared reflectance. Fresh snow has very high green reflectance and very low SWIR reflectance; bare ground and vegetation reverse this. MODIS MOD10A1 maps SCA daily at 500 m, resolving the patchwork of snow-covered and snow-free terrain that characterises a catchment in early autumn or late spring. Sentinel-2 pushes this to 20 m, which matters when a catchment contains steep south-facing slopes that lose snow weeks before north-facing ones.

Cloud is the practical constraint. A persistent frontal system can block optical observation for a week or more across an entire mountain range, precisely when the snowpack is changing fastest. The standard operational response is to combine observations from multiple satellites (Terra, Aqua, NOAA-20, Suomi-NPP) to reduce the probability that all sensors are simultaneously clouded, then apply temporal gap-filling: propagating the most recent clear-sky observation forward, weighted by a climatological melt or accumulation rate. Spatial interpolation using terrain aspect and elevation adds a second layer. Neither method is perfect; a gap-filled pixel carries more uncertainty than a directly observed one, and any gap-fill product should report that uncertainty explicitly.

Fusion: where the two signals compensate for each other

The operational standard, used by agencies including the US National Operational Hydrologic Remote Sensing Center and several European hydro utilities, fuses MODIS fractional SCA with AMSR2 SWE. The logic is complementary. MODIS resolves spatial heterogeneity and detects snow presence or absence accurately in shallow packs, where microwave retrievals are also most reliable. AMSR2 provides daily, cloud-free estimates that anchor the total basin-wide water volume, even though those estimates saturate at depth. Together, SCA constrains the spatial extent of the snowpack while SWE constrains the volumetric total.

In practice, the fusion is implemented through empirical regression or, increasingly, through data-assimilation schemes that ingest both data streams into a land-surface model. The model carries a prior state, updates it with each new satellite observation weighted by its known uncertainty, and produces a posterior SWE field at sub-kilometre resolution with quantified confidence intervals. The output is not a single number but a probability distribution over possible spring inflow volumes, which is exactly what a reservoir operator needs to make hedged decisions about drawdown and power scheduling.

Turning a snowpack estimate into a generation forecast

SWE is a reservoir of water, not a flow rate. Converting it to a forecast inflow requires a runoff model that accounts for temperature-driven melt timing, soil infiltration, rain-on-snow events and routing through the catchment. Degree-day models are the simplest approach: accumulated positive temperature drives melt at a fixed rate per degree. Energy-balance models are more accurate but require wind, humidity and radiation inputs that are themselves derived from reanalysis or numerical weather prediction.

The satellite contribution is in the initial condition. A well-constrained SWE field at the start of the melt season reduces forecast uncertainty by a larger margin than any improvement to the melt model itself, because the dominant source of error in spring-inflow forecasts is not knowing how much water is stored in the snowpack. Published studies of western US basins have shown that SWE-initialised statistical models can produce skilful inflow forecasts three to four months ahead, with skill declining as the melt season progresses and the snowpack signal is consumed. Forecast skill is highest for basins where snowmelt dominates over rainfall runoff and where the catchment has a long melt season that spreads the inflow over weeks.

Satellize applies this fusion approach operationally, running MODIS and AMSR2 data streams through gap-filling and assimilation pipelines for client catchments. The Tonga crop-estimation programme uses a related multi-sensor fusion logic for agricultural monitoring; the hydro snowpack workflow shares the same data-engineering architecture.

Honest limits and the cases where satellite data is not enough

Forested catchments are a persistent problem. The forest canopy intercepts and sublimates a significant fraction of snowfall, and it also absorbs and re-emits microwave radiation in ways that mimic a shallower snowpack. Correction algorithms exist but introduce additional uncertainty. In heavily forested boreal basins, passive microwave SWE errors can reach 50 mm or more, which is large relative to the total pack.

Very high-elevation, high-SWE catchments, including the main stems of Himalayan and Andean hydro systems, remain genuinely difficult. The saturation problem is worst precisely where the water volumes are largest. Spaceborne lidar from ICESat-2 can measure snow depth along ground tracks at centimetre vertical precision, but its spatial sampling is too sparse to characterise a full catchment without interpolation. Airborne lidar surveys fill this gap operationally in some North American basins, but they are expensive and infrequent. Satellite-only solutions for these catchments should be presented as best-available estimates with honest uncertainty bounds, not as definitive measurements.

Typical figures

SWE spatial resolution (AMSR2)	~25 km (standard product); downscaled products to ~5 km using terrain covariates
Snow-cover spatial resolution	500 m (MODIS), 375 m (VIIRS), 20 m (Sentinel-2 NDSI)
Revisit (passive microwave)	Daily global coverage, all weather
Revisit (optical, cloud-free)	1–2 days (MODIS + VIIRS combined); 5 days (Sentinel-2 at mid-latitudes, cloud-permitting)
SWE retrieval valid range	~0–150 mm; retrieval saturates above this threshold
Passive microwave frequency channels used	18.7 GHz and 36.5 GHz (primary SWE channels on AMSR2)
NDSI threshold for snow classification	NDSI ≥ 0.4 (MODIS/VIIRS standard; adjustable for forested or shallow-snow conditions)
Archive depth	MODIS from 2000; AMSR-E/AMSR2 from 2002/2012; VIIRS from 2012; Sentinel-2 from 2015
Product latency	Near-real-time MODIS/VIIRS: ~6 hours after overpass; AMSR2 SWE: ~24 hours
Delivery formats	GeoTIFF, NetCDF-4, COG; basin-aggregated CSV time series; API feed

Analytics Satellize can run

Daily basin-mean SWE map with uncertainty bounds	AMSR2 brightness-temperature retrieval fused with MODIS SCA via regression or data assimilation; forest-cover correction applied	GeoTIFF raster per catchment, daily update, with per-pixel confidence layer
Cloud-gap-filled fractional snow-cover time series	Multi-sensor temporal compositing (MODIS Terra + Aqua + VIIRS) with terrain-aspect-weighted spatial interpolation for residual gaps	Daily 500 m SCA raster; basin-aggregated percentage snow-covered area as CSV
Seasonal SWE accumulation and anomaly report	Comparison of current-year SWE trajectory against 20-year MODIS/AMSR climatological baseline; percentile ranking per catchment	Monthly PDF bulletin with time-series plots and percentile maps; GIS layer
Spring inflow volume probability forecast	SWE-initialised degree-day or energy-balance runoff model; ensemble output driven by ECMWF seasonal temperature scenarios	Probabilistic inflow forecast (10th, 50th, 90th percentile) at 1-week to 4-month horizon; updated weekly
Melt-onset detection alert	Passive microwave wet-snow signature (brightness-temperature diurnal variation at 36.5 GHz) combined with MODIS SCA decline rate	Email or API alert when melt onset detected across >20% of catchment area
Sentinel-2 calibration scenes for SWE downscaling	NDSI at 20 m resolution used to train sub-pixel snow-fraction model; applied to coarsen-resolution operational products	Seasonal archive of clear-sky Sentinel-2 NDSI scenes per catchment; calibration coefficients

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.