Crop loss estimation from flood recession timing
Multi-temporal SAR and optical imagery can pinpoint the day floodwater leaves each agricultural parcel, then cross-reference that duration against crop submergence thresholds to estimate yield loss before harvest.
Sensors
- Sentinel-1 SAR (C-band, ESA): 10 m ground range resolution in IW mode, six-day exact repeat at mid-latitudes (12-day per orbit direction). Detects open water as low backscatter against brighter vegetated surrounds. Cloud-penetrating, so recession timing is not lost to monsoon overcast. Known ambiguity: saturated bare soil and shallow standing water can produce similar backscatter signatures in C-band.
- PlanetScope (Planet Labs): 3 m resolution, near-daily global revisit in four optical bands (blue, green, red, NIR). Enables high-spatial-resolution NDVI recovery curves once cloud clears, confirming or refining SAR-derived recession dates at field scale.
- Sentinel-2 MSI (ESA): 10 m in visible and NIR bands, 20 m in red-edge and SWIR, five-day revisit with two satellites. SWIR bands (1610 nm, 2190 nm) improve discrimination between standing water and moist soil, partially resolving the ambiguity that troubles C-band SAR alone.
- MODIS Terra/Aqua: 250 m to 500 m resolution, twice-daily overpass. Too coarse for parcel-level work but useful for regional flood duration mapping and for anchoring time-series baselines across full river basins. Archive extends to 2000, supporting historical loss calibration.
Why the clock starts the moment water arrives
Crop submergence tolerance is not a binary threshold. It is a function of duration, growth stage, temperature and species. Agronomic research on rice, the world's most flood-prone staple, documents that complete submergence for more than three days during the tillering or booting stage can reduce yield by 30 to 70 percent, depending on variety. Maize is less forgiving: even two days of waterlogging at the silking stage can cut yields by half. Wheat root systems begin to suffer within 24 to 48 hours of soil saturation. These figures come from controlled and field studies published in peer-reviewed agronomy literature and are the physical foundation on which satellite-derived timing data becomes economically meaningful.
The satellite's job is therefore not simply to map where flooding occurred. It is to answer a more precise question: on which calendar date did water leave each parcel, and how does that date compare to the crop's growth stage on that same date? A field flooded for four days during vegetative growth may recover fully. The same field flooded for four days during grain filling may not. Getting the recession date right, at field scale, is what separates a useful loss estimate from a coarse approximation.
Reading recession from radar: what the backscatter tells you and what it hides
Sentinel-1's C-band SAR (5.4 GHz) reads open water as very low backscatter, typically below minus 15 dB in VV polarisation over calm surfaces, because the specular reflection returns almost nothing to the sensor. As water recedes and soil or vegetation reappears, backscatter rises. A time series of Sentinel-1 acquisitions, processed to gamma-nought backscatter, produces a characteristic signature: a dip during inundation followed by a recovery. The inflection point, fitted with a change-point algorithm or a simple threshold crossing, gives the recession date to within the six-day repeat interval.
The honest limitation is that saturated bare soil can produce backscatter values close to those of shallow standing water in C-band. A paddy field that has just drained but remains waterlogged at the root zone may look, to the radar, much like one still inundated. Incorporating Sentinel-2 SWIR reflectance, which is more sensitive to surface moisture than NIR alone, reduces this ambiguity when cloud cover permits. PlanetScope NDVI recovery provides an independent confirmation signal: live crop tissue begins reflecting NIR within days of stress relief, so a rising NDVI curve constrains the recession date from the biological side rather than the physical.
Neither sensor alone is sufficient. The operational approach uses SAR to maintain temporal continuity through cloud cover, then uses optical data to sharpen and validate the recession date estimate when the sky clears. Expect recession date uncertainty of roughly three to six days under persistent monsoon cloud, and one to two days where optical and SAR data overlap.
Translating inundation duration into a yield-loss surface
Once recession dates are assigned to individual parcels, the analysis cross-references three inputs: the inundation duration in days, the crop calendar for the region (planting date, growth stage sequence, expected harvest window) and the species-specific submergence tolerance curves from agronomic literature. The output is a parcel-level yield-loss fraction, typically expressed as a percentage of expected yield under normal conditions.
Crop calendars are the largest source of uncertainty in this step. Where planting dates are known from government records or from SAR-derived transplanting detection (flooded paddies show a distinctive backscatter signature at transplanting), the growth-stage assignment is reasonably precise. Where planting dates must be inferred from regional averages, the uncertainty in growth-stage assignment can easily exceed the uncertainty in the recession date itself. Communicating this honestly to clients, and flagging parcels where planting-date confidence is low, is part of responsible product design.
Aggregating parcel-level fractions to administrative units gives district or province-level loss estimates that can feed food-security assessments, insurance triggers or government compensation schemes. The aggregation step also smooths some of the parcel-level noise, though it can mask localised losses in heterogeneous landscapes.
What the archive makes possible that field surveys cannot
Sentinel-1 data is free and open, with a consistent global archive back to 2014. MODIS extends flood-duration records to 2000. This means that for any given river basin, it is possible to reconstruct flood recession timing and estimate crop loss for past events, calibrate the model against actual reported yields where those records exist, and build a regional loss function that improves with each season of data.
Field surveys are expensive, slow and, in active flood conditions, often impossible. A satellite-derived loss estimate can be produced within days of recession, covering an entire river basin, at a spatial resolution that distinguishes individual field parcels. That speed matters for anticipatory humanitarian action, insurance payouts and government procurement decisions. A survey that arrives three months after harvest is too late to influence any of those decisions.
Where the method works well and where it does not
The approach performs best on rice in flat deltaic landscapes: large, geometrically regular paddies, well-documented crop calendars, and a flood-recession dynamic that is slow enough for six-day SAR revisit to capture reliably. It also works on wheat and maize in temperate floodplains, though parcel sizes in some regions are small enough that 10 m SAR resolution begins to mix field and boundary signals.
It performs poorly in three situations. First, under dense forest or tall crop canopies, C-band SAR cannot see the soil surface at all; the double-bounce from flooded vegetation can actually increase backscatter, inverting the expected signal. Second, in hilly or terraced landscapes, SAR layover and shadow artefacts corrupt the backscatter record in ways that are difficult to correct without high-quality DEMs. Third, where multiple flood pulses occur within a single growing season, disentangling cumulative stress from individual events requires careful time-series decomposition that adds analytical complexity and uncertainty.
Satellize's crop-estimation work in the Kingdom of Tonga has demonstrated that combining open-constellation SAR with high-resolution optical revisit produces parcel-level estimates with enough precision to support government food-security planning. The same methodological stack applies to flood-loss estimation wherever crop calendars are available to anchor the growth-stage assignment.
Typical figures
| Primary SAR resolution (Sentinel-1 IW mode) | 10 m (range) × 10 m (azimuth), ground range detected |
| SAR revisit interval | 6 days at mid-latitudes (single satellite); 12 days per orbit direction |
| Optical resolution (PlanetScope) | 3 m, near-daily revisit |
| Optical resolution (Sentinel-2 MSI) | 10 m visible/NIR, 20 m red-edge/SWIR, 5-day revisit (two satellites) |
| Recession date uncertainty (SAR only) | 3 to 6 days under persistent cloud; 1 to 2 days with optical confirmation |
| Minimum parcel size reliably resolved | Approximately 0.5 ha with Sentinel-1 at 10 m; smaller parcels suffer mixed-pixel effects |
| Spectral bands used | C-band VV/VH (SAR); NIR, SWIR (Sentinel-2); Red, NIR (PlanetScope NDVI) |
| Archive depth | Sentinel-1 from 2014; MODIS from 2000; PlanetScope from approximately 2016 |
| Latency from flood recession to initial estimate | 2 to 5 days after next cloud-free or SAR acquisition |
| Delivery formats | GeoTIFF parcel-loss rasters, GeoJSON/Shapefile with per-parcel attributes, tabular CSV by administrative unit, PDF summary report |
Analytics Satellize can run
| Flood recession date map | Change-point detection on multi-temporal Sentinel-1 gamma-nought backscatter time series (VV and VH polarisation); validated against Sentinel-2 SWIR where cloud-free | GeoTIFF raster assigning recession date (day of year) to each 10 m pixel; per-parcel median date in GeoJSON |
| Inundation duration surface | Difference between flood onset date (from SAR or optical) and recession date; onset from companion flood-extent analysis | GeoTIFF raster of inundation duration in days, clipped to agricultural land mask |
| Growth-stage-at-inundation classification | Cross-reference of recession date with regional crop calendar (transplanting to harvest phenology); planting dates inferred from SAR transplanting signal or government records | Parcel-level attribute table classifying each parcel by growth stage at peak inundation (vegetative, reproductive, grain fill) |
| Parcel-level yield-loss fraction | Species-specific submergence tolerance functions from published agronomy literature applied to inundation duration and growth stage; uncertainty range reported per parcel | GeoJSON with per-parcel yield-loss fraction (0 to 1) and confidence band; GeoTIFF for GIS integration |
| District and province loss aggregation | Area-weighted aggregation of parcel-level fractions to administrative boundaries; expressed as estimated production loss in tonnes where crop-area statistics are available | CSV and PDF table by administrative unit; suitable for food-security situation reports or insurance trigger documentation |
| NDVI recovery monitoring (post-flood) | Time series of NDVI from PlanetScope or Sentinel-2 over affected parcels; recovery rate compared to unaffected reference parcels in same agro-ecological zone | Weekly GeoTIFF NDVI anomaly layers and tabular recovery-rate summary for up to 60 days post-recession |
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.