Sensors
- Sentinel-1 C-band SAR (ESA): 10 m ground range resolution in Interferometric Wide Swath mode, 250 km swath, 6-day repeat at the equator with both satellites operating (12-day with one). The primary workhorse for paddy flood mapping: freely available, consistent since 2014, and well-characterised in peer-reviewed rice studies across South and Southeast Asia.
- ALOS-2 PALSAR-2 L-band SAR (JAXA): L-band (1.27 GHz) penetrates the rice canopy more deeply than C-band, preserving a water-surface double-bounce signal even after tillering when C-band loses sensitivity. Spatial resolution 3–10 m depending on mode; 14-day repeat. Useful for later growth stages where C-band becomes ambiguous, though tasking is not free.
- RADARSAT-2 C-band SAR (MDA): Flexible polarisation modes including quad-pol, which improves discrimination between flooded vegetation and open water. Resolutions from 3 m (Fine mode) to 100 m (ScanSAR). Commercial tasking allows targeted acquisition over a specific delta at a specific time, valuable when Sentinel-1 revisit does not align with a fast-moving flood event.
- Sentinel-2 MSI (ESA): 10–20 m multispectral imagery used as a complement, not a substitute: NDVI and NDWI time series help confirm crop growth stage and identify cloud-free windows for cross-validation. 5-day revisit with both satellites. Cloud cover in monsoon season is severe and limits optical utility precisely when flood risk is highest.
Why radar, and why rice specifically
When a rice paddy is flooded and the crop is young, the water surface acts as a near-perfect specular reflector. Microwave energy from a SAR satellite bounces away from the sensor rather than back to it, producing a sharp drop in backscatter. This is the same physics that makes a calm lake look dark on radar imagery. The signal is strong, consistent, and does not depend on sunlight or cloud-free skies.
Rice is unusual among staple crops because deliberate flooding is part of the cultivation cycle. Farmers flood fields at transplanting, maintain shallow water through vegetative growth, then drain before harvest. That structured calendar means a SAR time series does not just detect disaster; it maps the entire agronomic sequence. A lender or insurer who knows the transplanting date knows the growth stage at any subsequent point in the season, and therefore knows what was at risk when a cyclone or monsoon surge arrived.
Reading the backscatter calendar
Sentinel-1's 6-day repeat (with both satellites) produces a time series dense enough to resolve the three or four distinct flood phases of a paddy season. At transplanting, backscatter in VV polarisation typically falls to values around -15 to -20 dB over open flooded fields, well below the -8 to -12 dB range of moist bare soil. As the canopy develops through tillering and heading, backscatter rises again because the crop itself scatters energy back toward the sensor. This trajectory is the fingerprint of a healthy paddy season.
By fitting a phenological model to the backscatter time series, analysts can estimate transplanting date, peak vegetative stage, and approximate harvest window at field parcel resolution. Several published studies using Sentinel-1 over the Mekong Delta, the Irrawaddy Delta, and the Ganges-Brahmaputra plain have demonstrated this approach with overall accuracy figures above 85 percent for transplanting-date estimation, though performance varies with field size and local agronomic practice. The method is documented in the peer-reviewed literature; it is not proprietary.
The credit-exposure application follows directly. If a lender holds a portfolio of paddy loans across a district, and a flood event is detected by SAR on a specific date, the growth-stage model tells the analyst what fraction of the portfolio was at early vegetative stage (low standing crop value, high replanting cost), heading (maximum value at risk), or post-harvest (loss already realised or avoided). That is a materially different risk picture from simply knowing that water was present.
Where the method fails, and what to do about it
Two physical effects degrade the signal in ways that matter for operational use. First, after tillering, the rice canopy becomes dense enough to scatter C-band energy before it reaches the water surface. Backscatter rises and the flooded-field signature weakens or disappears. This is not a processing artefact; it is a fundamental limit of 5.4 GHz radiation interacting with vegetation of that density. L-band SAR from ALOS-2 partially mitigates this because longer wavelengths penetrate the canopy further, but ALOS-2 is a tasked commercial asset with a 14-day revisit and coverage gaps.
Second, wind roughening of flooded surfaces raises backscatter and can cause inundated fields to be misclassified as dry land. A wind speed of roughly 3–5 m/s is enough to produce measurable roughening at C-band. Acquisitions during or immediately after storms, precisely when flood mapping is most urgent, carry elevated false-negative risk. Analysts should flag acquisition-time wind speed from reanalysis products and treat high-wind overpasses with appropriate scepticism.
Optical data from Sentinel-2 helps when cloud cover permits, providing an independent check on crop stage and surface water extent. In practice, during the South Asian monsoon, cloud-free optical imagery over flooded paddies is rare for weeks at a time. The honest position is that SAR is the primary sensor for this application, and its late-season canopy limitation is real and should be disclosed to any insurer or lender using the output.
From flood map to credit-exposure layer
The analytic product a lender actually needs is not a flood map. It is a table: for each loan or insured unit in the portfolio, what growth stage was the crop at on the date of the event, and what area was inundated. Producing that table requires three inputs: the SAR-derived inundation extent raster, the phenological calendar derived from the same time series, and a loan or policy register that can be spatially joined to field parcels.
Field parcel boundaries are often the binding constraint. Where cadastral data exists and is current, the join is straightforward. Where it does not, parcels can be estimated from the SAR time series itself using change-detection segmentation, though the resulting boundaries carry positional uncertainty of one to two pixels (10–20 m for Sentinel-1) and should be treated as approximate.
Satellize has applied related phenological calendar methods in its Kingdom of Tonga crop-estimation programme, where Sentinel-1 and Sentinel-2 time series are used to characterise agricultural cycles in a small-island context. The paddy flood-calendar problem is technically similar, scaled to delta geographies with far larger parcel counts.
Practical parameters for a lender or insurer commissioning this work
Archive depth matters. Sentinel-1 data runs back to 2014 for most coverage zones, providing eight to ten full rice seasons for baseline calibration. That is enough to characterise normal transplanting-date distributions and identify anomalous seasons, which is the comparison a parametric insurance product requires.
Latency from SAR acquisition to a processed inundation layer is typically 24–72 hours for a standard Sentinel-1 product pipeline, depending on data downlink scheduling and processing queue. For post-event loss assessment, that is fast enough to be operationally useful. For real-time flood monitoring during an active event, commercial tasking of RADARSAT-2 or a similar asset can reduce latency further, at cost.
Coverage is not a limitation for the major rice-growing deltas. Sentinel-1 provides systematic global coverage. The Mekong Delta, the Irrawaddy, the Chao Phraya basin, the Ganges-Brahmaputra-Meghna system, and the major Indonesian rice bowls are all within routine acquisition zones. A lender with exposure across multiple countries can run a single consistent methodology across the entire portfolio.
Typical figures
| Primary sensor | Sentinel-1 C-band SAR, 5.405 GHz |
| Spatial resolution | 10 m (Sentinel-1 IW mode); 3–10 m (RADARSAT-2 Fine/ScanSAR); 6–10 m (ALOS-2 PALSAR-2) |
| Revisit interval | 6 days (Sentinel-1, dual satellite); 14 days (ALOS-2); variable (RADARSAT-2, tasked) |
| Swath width | 250 km (Sentinel-1 IW); 50–500 km (RADARSAT-2 mode-dependent) |
| Processing latency (post-acquisition) | 24–72 hours for standard inundation layer; faster with priority tasking pipelines |
| Archive depth | Sentinel-1: 2014 to present (8–10 full rice seasons for most Asian coverage zones) |
| Minimum detectable inundated area | Approximately one Sentinel-1 pixel (100 m²) in open water; larger parcels required for reliable phenological modelling (typically >0.5 ha) |
| Known signal degradation | Canopy closure after tillering (C-band); wind roughening above ~3–5 m/s; steep terrain shadow |
| Delivery formats | GeoTIFF inundation rasters, GeoPackage or Shapefile parcel-level exposure tables, CSV loss-stage summaries, PDF seasonal reports |
| Optical cross-validation | Sentinel-2 MSI, 10–20 m, 5-day revisit; cloud-limited during monsoon season |
Analytics Satellize can run
| Seasonal transplanting-date map | Backscatter time-series phenological modelling (VV/VH Sentinel-1); threshold-based detection of initial inundation drawdown followed by vegetative backscatter rise | GeoTIFF raster of estimated transplanting date per field segment, delivered at season onset; updated as new acquisitions arrive |
| Flood-event inundation extent layer | Change-detection against pre-event backscatter baseline; thresholding in VV polarisation with wind-speed quality flag from ERA5 reanalysis | GeoTIFF inundation mask with confidence classification (high/medium/low) and acquisition-time wind-speed annotation |
| Growth-stage-at-event table | Spatial join of inundation mask to phenological calendar; growth stage assigned per parcel using days-since-transplanting model | CSV or GeoPackage table: parcel ID, inundated area (ha), growth stage at event date, estimated standing crop value band |
| Portfolio credit-exposure summary | Aggregation of parcel-level growth-stage table against lender loan register; exposure bucketed by growth stage and administrative unit | PDF report and accompanying spreadsheet: district-level exposure by growth stage, ranked by estimated loss severity |
| Historical season baseline (calibration dataset) | Retrospective Sentinel-1 time-series analysis over 2014–present archive; normal transplanting-date distribution and flood-frequency statistics per district | GeoPackage historical flood-frequency layer and season-calendar statistics table, used to set parametric trigger thresholds |
| Near-real-time flood alert | Automated backscatter anomaly detection on each new Sentinel-1 acquisition; alert triggered when inundation exceeds user-defined area threshold within a defined portfolio boundary | Email or API alert with inundation area estimate, acquisition time, and wind-speed quality flag; updated at each overpass |
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.