Remote Sensing Labs
Browse all docs

Flood Mapping

Flood mapping asks the imagery a time-critical question: where is water sitting now that is not normally there, and how sure are we of the extent unit by unit. The workflow is distinctive because the best sensor depends on the weather — storms that cause floods also fill the sky with cloud, which blinds optical imagery exactly when it is needed, so radar often carries the event while optical imagery carries the calm baseline. This page walks the whole path, from framing the mapping question to interpreting a flood-extent result. Who consumes a flood map, and on what clock, is covered in Disaster Response; how the same water-detection step serves wetland and surface-water records is covered in Environmental Monitoring.

Frame the mapping question first

Decide the area of interest, the event window (the days of peak inundation versus a specific date), and whether you need maximum extent across the event or a date-specific snapshot — a receding flood looks very different depending on when you sample it. Name the reporting units the result must feed: communities, parcels, road segments, cropland, or watersheds. And fix the comparison explicitly: flood water only means something relative to a normal-conditions baseline, because a river channel that is always wet is not a flood. Those choices also decide, up front, which sensing approach the next step should reach for.

Choose a sensing approach and understand the tradeoff

The central decision in flood mapping is optical versus radar, and each detects water differently. Optical sensors use water indices built on the fact that open water is dark in the near- and shortwave-infrared. McFeeters’ NDWI contrasts green with near-infrared, (Green - NIR) / (Green + NIR); the modified MNDWI (Xu, 2006) swaps in the shortwave-infrared band near 1.6 µm, (Green - SWIR1) / (Green + SWIR1), which suppresses built-up and soil noise better. A third index also called NDWI (Gao, 1996) contrasts NIR with SWIR to measure water held in leaves; it is a vegetation-moisture index, not an open-water index, and the shared name is a frequent source of error. All of these rest on the surface responses the Spectral Bands concept describes. Their weakness is fatal during storms: optical imagery cannot see through cloud, so the flood peak is often hidden.

Synthetic aperture radar solves exactly that problem. As the SAR Basics concept explains, radar is an active microwave sensor that supplies its own illumination and is largely unaffected by cloud, day or night — heavy rain can still degrade the signal, more at shorter wavelengths — so it can image a flood while the storm is still overhead. Radar detects open water by its backscatter: a calm water surface acts like a mirror, reflecting the pulse away from the sensor, so open water returns very little and appears dark against rougher, brighter land. Polarization matters: VV gives the strongest contrast for open water, while VH is generally less disturbed by wind-roughened water. Flooded vegetation and flooded urban areas behave the other way. Water beneath a canopy or between buildings forms a corner with trunks or walls that bounces the pulse straight back, so the flood brightens the return rather than darkening it. That double-bounce signal is clearest at longer wavelengths (L-band) in HH polarization, weaker at C-band under dense canopy, and in cities is often read from a loss of interferometric coherence rather than from brightness alone. The honest framing is that neither sensor is “the” flood tool — radar carries cloud-covered peaks but needs different reasoning for flooded vegetation and urban areas, optical maps clear-sky open water cleanly — and many workflows use both.

Acquire baseline and event imagery

With the approach chosen, acquire two comparable observations: a baseline that represents normal conditions and an event image at or near peak inundation. For radar, the open workhorse is Sentinel-1, C-band, imaging land in VV and VH. Sentinel-1B failed in December 2021, leaving a 12-day repeat for several years; Sentinel-1C (launched December 2024) and Sentinel-1D (launched November 2025) restore a two-satellite six-day repeat, but even that can miss a short flood peak, so check what actually acquired over your event before you commit to a date. Prefer scenes from the same relative orbit and viewing geometry so backscatter is comparable date to date. Where the workflow takes the optical route, that pair usually comes from Sentinel-2, with season and sun angle held as close as the archive allows. A long-term permanent-water reference such as the JRC Global Surface Water dataset supplies the “normally wet” layer that the delineation step subtracts. Match the spatial detail to the reporting units, using the tradeoffs the Resolution concept lays out, and use calibrated products, as the Digital Imagery concept advises, so any later difference reflects the water rather than the processing. Finding the two dates is a STAC catalog query rather than a manual browse, which is what makes a narrow event window tractable at all. The deadline constrains that step as much as the weather does: an extent map is wanted within hours, and staging whole scenes across a basin spends most of that budget on transfer, so flood work leans hard on the cloud-native habit of reading only the windows the map needs out of imagery left where it is.

Prepare the imagery before you delineate

Preprocessing differs by sensor, and skipping it is where false floods are born. For optical imagery, mask cloud, cloud shadow, and haze on both dates before computing any index, following the Cloud Masking concept — an unmasked cloud shadow is dark in the infrared and reads as fake water. For radar, use radiometrically calibrated, terrain-corrected backscatter in decibels (dB) so slopes are not mistaken for bright or dark returns, and reduce speckle with care: spatial filters blur narrow channels and small ponds, so multi-temporal filtering across a stack of dates is usually preferred. Either way, confirm both dates share a coordinate reference and grid, as the Coordinate Systems concept requires; a half-pixel misregistration invents change at every shoreline.

Delineate water extent against the baseline

Now identify water on the event image and, crucially, subtract what is normally there. On optical data, threshold NDWI or MNDWI to separate open water from land. On radar, threshold the low-backscatter returns that mark smooth water — but do not borrow a fixed number: open-water backscatter varies with incidence angle, wind, and sensor, so a global cut-off (values around −15 to −20 dB in VV are often quoted) rarely transfers between scenes. Operational methods instead derive the threshold per scene or per tile from a bimodal histogram of water and land, using Otsu’s method or split-based tiling that finds tiles containing both classes (Martinis et al., 2009), or they compare the event against a baseline with a log-ratio (a difference in dB) so that only new darkening is flagged. Remove ground that cannot plausibly flood with a terrain mask such as HAND (Height Above Nearest Drainage) and mask radar shadow and layover; the Copernicus Global Flood Monitoring service applies exactly such an exclusion layer. A one-date threshold alone still confuses flood with permanent water, so compare the event against the baseline: the new water — present at the event, absent in normal conditions — is the flood, and this differencing is the multi-temporal reasoning the Change Detection concept develops. Turning continuous index or backscatter values into a water/not-water label is the labeling step the Classification Basics concept covers, so record the exact thresholds, masks, and baseline definition. Distinguishing flood water from the permanent river, lake, or reservoir is not optional cleanup; it is the core of an honest flood extent.

Separate real flooding from confounders

Not every dark radar pixel or high water index is a flood. Radar mistakes other smooth, low-backscatter surfaces for water — dry sand, tarmac and runways, and radar shadow behind terrain and buildings all return little energy. It also misses water: wind or heavy rain roughens open water until it returns as brightly as land, a leading cause of omitted flood in C-band VV, and water under vegetation or among buildings brightens rather than darkens. Optical water indices are fooled by terrain and cloud shadow and dark soils; MNDWI can confuse some built-up surfaces and snow with water; and turbid, sediment-laden floodwater reflects more in the visible and near-infrared than clear water, lowering index values until real flood water falls below the threshold. Both approaches must exclude the permanent-water baseline, and the Time Series concept matters because a single event scene with a thin baseline can manufacture apparent inundation on its own. Treat a mapped flood patch as a hypothesis and check whether it survives these confounders — and where possible a second sensor or date — before it is escalated for response.

Produce outputs, interpret, and make it reproducible

A flood run typically delivers a flood-extent map, an area table by reporting unit, and an exposure summary for the communities, roads, parcels, or cropland the question named. The confidence statement is part of the product: where cloud forced a gap in the optical record, where radar could not see water under canopy or in the urban core, where wind may have hidden open water, and how far the thresholds were trusted. Record the area of interest, the event and baseline dates and sensors, the polarization, the masking, calibration, and speckle-filter settings, the water index or backscatter threshold and how it was derived, the terrain and permanent-water masks, and the reporting units — enough for the next event to be mapped the same way. Written as a raster processing pipeline, that record becomes the pipeline’s parameters; the optical water indices alone also appear as worked code in the Earth Engine indices lesson.

Sources