Flood extent mapping with optical imagery
Multispectral satellites can delineate inundated surfaces within hours of a flood peak, but only when skies clear. This page explains the spectral physics, the indices, and the honest limits of optical flood mapping.
Sensors
- Sentinel-2 MSI (ESA): 10 m resolution in visible and near-infrared bands; 20 m in red-edge and shortwave infrared. Five-day revisit at the equator with both satellites combined. The 20 m SWIR band (Band 11, 1610 nm) is the workhorse for MNDWI computation. Free, open archive from 2015.
- Planet SuperDove: 3 m resolution, eight spectral bands including NIR. Revisit varies by latitude and tasking priority but can reach daily in high-demand areas. Commercially licensed. Useful for mapping inundation at field and street scale, though SWIR is absent, limiting MNDWI to NDWI variants.
- MODIS Terra/Aqua (NASA): 250 m to 500 m resolution depending on band. Near-daily global coverage from two satellites. Spatial detail is coarse, but the revisit means a cloud-free observation often arrives within 24 to 48 hours of a storm clearing. The published MODIS Near Real-Time Flood Product (MNDWI-based) has been operational since 2012.
Why water goes dark in the near-infrared
Open water absorbs near-infrared radiation almost completely. Vegetation and dry soil reflect it strongly. That contrast is the physical foundation of every spectral water index used in flood mapping, and it is not an artefact of processing: it follows directly from the absorption spectrum of liquid water around 800 to 900 nm.
The Normalised Difference Water Index (NDWI), proposed by McFeeters in 1996, exploits this by computing (Green minus NIR) / (Green plus NIR). Positive values indicate water; vegetated land returns negative values. The Modified NDWI (MNDWI) replaces the NIR band with shortwave infrared (around 1610 nm), which suppresses the signal from built-up surfaces and sediment-laden water more effectively. For flood mapping in urban or agricultural settings, MNDWI generally outperforms NDWI, which is why Sentinel-2's Band 11 matters so much.
Before-and-after differencing: what it adds and what it costs
A single post-event water index image identifies all water bodies, including rivers, lakes and reservoirs that were there before the flood. The standard way to isolate inundation is image differencing: subtract the pre-event index from the post-event index, then threshold the result. Pixels that cross the threshold are classified as newly flooded.
The method is straightforward but introduces dependencies. The pre-event image must be acquired under comparable phenological conditions; a pre-event scene taken in winter when fields are bare will produce false positives against a post-event summer image where crops would normally raise NIR reflectance. Seasonal archive depth matters. Sentinel-2's archive from 2015 and Landsat's from the 1970s provide enough history to select seasonally matched baselines, but analysts must choose deliberately rather than automatically.
Threshold selection is the other decision point. A fixed global threshold for MNDWI performs poorly across soil types and turbidity levels. Otsu's method, which finds the threshold that minimises intra-class variance, is widely used and works well when water and land are both present in the scene. In images where flooded area is a small fraction of the total, it can underperform, and manual adjustment or training data improves results.
The cloud problem has no optical solution
Active flood events almost always arrive beneath cloud. Convective storms, tropical cyclones, and monsoon systems that produce the worst inundation also produce the densest cloud cover. Optical sensors cannot see through cloud. This is not a processing limitation; it is physics.
In practice, the first usable Sentinel-2 or SuperDove acquisition after a major flood event may arrive two to five days after the peak, by which time water has already begun to recede in well-drained areas. The mapped extent will understate peak inundation. MODIS's near-daily revisit improves the odds of catching a partial clear, and its global flood product has been used in operational response, but at 250 m resolution it misses street-level detail that matters for evacuation routing or damage assessment.
The honest position is that optical flood mapping is most reliable in the recession phase, not at peak. For peak-inundation mapping during active storms, synthetic aperture radar is the appropriate sensor class. That use case is covered separately in the sibling page on SAR flood mapping.
Sentinel-2 in practice: resolution, latency, and what gets missed
A Sentinel-2 scene covers 290 km by 290 km in a single pass. Processing to Level-2A (surface reflectance) through the Copernicus Data Space is typically available within a few hours of acquisition. For a disaster-declared event, the International Charter on Space and Major Disasters can task priority acquisitions and accelerate delivery, though the cloud physics still apply.
At 10 m resolution in the visible bands and 20 m in SWIR, Sentinel-2 resolves individual fields and road segments. It will not reliably detect inundation in narrow urban streets, culverts, or beneath dense tree canopy. Water under forest is systematically underestimated by optical indices because the canopy signal dominates. Flooded paddy fields with standing crop present a similar ambiguity: the water is real but the index reads the vegetation above it.
SuperDove at 3 m fills some of the urban gap, but its lack of a SWIR band means analysts rely on NIR-based NDWI, which is more susceptible to confusion with shadow and dark impervious surfaces. Careful masking of shadow zones is required in dense urban scenes.
From index to deliverable: the analytic chain
A flood extent map is not an end product for most clients. Emergency managers need it georeferenced and overlaid with population grids, road networks, and critical infrastructure. The standard workflow runs from atmospherically corrected imagery to index computation, to binary water mask, to vector polygon, to intersection with exposure datasets such as WorldPop or OpenStreetMap infrastructure layers.
Accuracy assessment requires reference data, which is rarely available in real time. Post-event validation typically uses high-resolution commercial imagery, field reports, or SAR-derived extents as cross-checks. Reported overall accuracies for MNDWI-based flood mapping in peer-reviewed studies typically range from 85 to 95 percent in open agricultural landscapes, dropping in urban and forested areas.
Satellize runs this analytic chain on open Sentinel-2 and MODIS imagery and can add commercial SuperDove tasking where street-level resolution is required. The same processing pipeline underpins the crop-area estimation work done for the Kingdom of Tonga, where accurate water and land discrimination is a prerequisite for yield modelling.
Typical figures
| Best spatial resolution (optical flood mapping) | 3 m (Planet SuperDove); 10 m visible / 20 m SWIR (Sentinel-2 MSI); 250 m (MODIS) |
| Revisit frequency | Daily (MODIS, global); 1 to 2 days (SuperDove, latitude-dependent); 5 days at equator (Sentinel-2 two-satellite pair) |
| Typical latency to analysis-ready data | 2 to 6 hours post-acquisition for Sentinel-2 L2A via Copernicus Data Space; MODIS NRT products within 3 hours |
| Key spectral bands | Green (560 nm), NIR (842 nm), SWIR (1610 nm) for MNDWI; Green + NIR for NDWI |
| Minimum detectable inundated area | Approximately 0.01 ha at 10 m resolution (Sentinel-2); approximately 6 ha at 250 m (MODIS) |
| Hard limit | Cloud cover: zero signal beneath cloud, regardless of sensor |
| Archive depth | Sentinel-2 from 2015; MODIS from 2000; Landsat (for baseline construction) from 1972 |
| Delivery formats | GeoTIFF (index rasters, binary masks), GeoJSON / Shapefile (flood polygons), COG for web delivery |
| Flood product latency (operational) | First optical map typically 24 to 120 hours after event peak, cloud-dependent |
Analytics Satellize can run
| Binary flood extent mask | MNDWI thresholding (Otsu or supervised) on Sentinel-2 L2A before/after image pair | GeoTIFF raster and vector polygon layer, georeferenced to WGS84 |
| Inundated area by administrative unit | Zonal statistics: flood mask intersected with official administrative boundary polygons | Tabular report (CSV) and choropleth GIS layer by district or commune |
| Flood depth proxy (where DEM available) | Flood extent boundary intersected with high-resolution DEM (SRTM, Copernicus DEM 30 m) to estimate water surface elevation and infer depth | Depth raster and classified depth zones (0 to 0.5 m, 0.5 to 1 m, >1 m) |
| Exposed population count | Flood mask overlaid on WorldPop or LandScan population grid; pixel-level summation | Situation report figure and disaggregated table by administrative level |
| Critical infrastructure exposure | Flood polygon intersection with road, hospital, school and utility layers from OpenStreetMap or national GIS | Infrastructure impact GIS layer and summary table for emergency coordinators |
| Flood recession time series | Multi-date MNDWI stack across post-event acquisitions; per-pixel first-dry-date extraction | Animated GIF or multi-band GeoTIFF stack; recession rate table by zone |
| Near-real-time coarse alert (MODIS-based) | MODIS 250 m MNDWI anomaly against seasonal baseline; threshold exceedance flagging | Daily alert feed (GeoJSON) covering affected river basin, with cloud-fraction metadata |
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.