Wetland and rice-fish co-culture system mapping and extent monitoring

SAR backscatter and optical time series map the seasonal flooding dynamics of rice-fish co-culture systems, supporting production estimates and habitat accounting where cloud cover makes optical-only approaches unreliable.

Sensors

Sentinel-1 C-band SAR (ESA): 10 m ground range detected resolution in IW mode, 6-day repeat at the equator (12-day single satellite). C-band double-bounce from vertical rice stems over standing water produces a distinctive backscatter increase of roughly 3–8 dB relative to dry fields, detectable throughout the monsoon cloud season.
ALOS-2 PALSAR-2 L-band SAR (JAXA): L-band (23.6 cm wavelength) penetrates the rice canopy more deeply than C-band, improving sensitivity to water beneath a closed canopy. Spatial resolution 3–10 m depending on mode; 14-day repeat. Particularly useful late in the growing season when C-band double-bounce weakens as the canopy thickens.
Sentinel-2 MSI (ESA): 10 m visible and near-infrared bands, 20 m shortwave infrared. 5-day revisit (two satellites). NDVI and NDWI time series resolve transplanting dates and drainage events during cloud-free windows. Usable fraction drops sharply during monsoon months in humid tropical regions, which is precisely when SAR carries the analysis.
Landsat 8/9 OLI (USGS/NASA): 30 m multispectral, 16-day repeat per satellite (8-day combined). Archive extends to 1984 for Landsat 5, giving multi-decadal baselines for wetland extent change. Thermal band supports evapotranspiration estimation as an ancillary productivity indicator.

What a flooded paddy tells the radar

A rice paddy in early flood stage behaves like a corner reflector. Water beneath the canopy acts as a horizontal mirror; vertical rice stems act as the second face. The C-band pulse from Sentinel-1 bounces off the stem, then off the water surface, and returns almost directly to the sensor. This double-bounce mechanism raises the VV or VH backscatter coefficient by several decibels compared with a dry field of identical geometry. The signal is strong enough to detect inundation through moderate cloud cover and light rain, which is the whole point during a monsoon growing season.

The effect is not constant across the crop calendar. At transplanting, when stems are short and sparse, backscatter is dominated by surface roughness of the water. As the canopy closes, L-band penetrates better than C-band. At grain fill, the canopy attenuates both. A time series of at least eight to twelve Sentinel-1 acquisitions per season is generally needed to track these transitions reliably and separate flooded from drained paddies at the same phenological stage.

The hard problem: rice-fish versus plain flooded rice

Radar backscatter alone cannot distinguish a rice-fish co-culture system from an ordinary flooded paddy. Both are inundated; both carry a rice canopy. The physical signal is the same. Optical imagery can sometimes resolve the slightly deeper water levels and wider bunds characteristic of managed co-culture plots, but at 10 m resolution the difference is marginal and site-dependent.

Reliable classification of co-culture systems requires ancillary data. Cadastral records, agricultural census layers, or field-verified training polygons are the practical inputs. Where these exist, a supervised classifier trained on SAR time-series features (mean backscatter, temporal variance, flood onset date) can separate co-culture plots from simple flooded rice with reasonable accuracy, though published studies report confusion rates that vary considerably by region and plot size. Where cadastral data are absent or outdated, the satellite product maps inundated paddy extent and seasonal flooding duration; the co-culture attribution must be done separately. This is an honest limit of the method, not a gap that more satellite data alone will close.

Building a seasonal flooding time series

The operational workflow fuses SAR and optical acquisitions into a single inundation probability layer updated at each new acquisition. Sentinel-1 IW mode provides the backbone: a 10 m, near-all-weather observation every six days. Sentinel-2 and Landsat acquisitions are ingested when cloud-free pixels are available, adding spectral indices that improve discrimination between open water, flooded vegetation, and moist bare soil.

Flood onset date, flood duration, and drainage date are extracted per pixel from the time series. These three parameters carry most of the agronomic signal. A paddy flooded for 90 days in a single continuous event looks different from one flooded in two shorter pulses, and the difference matters for fish yield estimation. Minimum detectable flood extent is constrained by spatial resolution: at 10 m, plots smaller than roughly 0.1 hectares are unreliable. Sub-hectare smallholder plots in fragmented landscapes are genuinely difficult, and no current open-constellation sensor resolves that fully.

From extent to production estimates

Mapped inundated area is the starting point for production accounting, not the end. Fish yield in co-culture systems depends on flood depth, duration, water temperature, and stocking density, none of which satellite imagery measures directly. What imagery provides is the spatial envelope and its temporal behaviour.

Area estimates derived from Sentinel-1 time series can be combined with government or survey-based yield coefficients to generate district-level production ranges. The uncertainty is substantial: yield coefficients for traditional rice-fish systems in South and Southeast Asia vary by a factor of three or more across published literature, reflecting genuine agronomic variability rather than measurement error. A satellite-derived area figure with a ten percent spatial uncertainty, multiplied by a yield coefficient with a fifty percent agronomic uncertainty, gives a production estimate with wide confidence bounds. Communicating those bounds honestly is part of what makes the product useful to a ministry of fisheries rather than misleading.

Habitat accounting and multi-year change detection

Beyond seasonal production, the same time series supports habitat accounting. Rice-fish co-culture systems in deltaic and floodplain environments are ecologically significant wetlands. They provide seasonal habitat for wild fish populations, waterbirds, and amphibians alongside their managed fishery function. Tracking their extent over years or decades matters for biodiversity accounting frameworks and for assessing the effects of drainage infrastructure investment, land consolidation, or shifting cultivation patterns.

Landsat's archive back to the 1980s makes multi-decadal change detection possible at 30 m resolution. Sentinel-1 and Sentinel-2 provide higher-resolution baselines from 2014 and 2015 onwards respectively. PALSAR-2 adds a complementary L-band record. Together these archives support change attribution at the landscape scale, distinguishing conversion to permanent aquaculture ponds, drainage for upland crops, or urban expansion from interannual variability in flood extent driven by rainfall. Satellize has applied comparable time-series crop-area methods in the Kingdom of Tonga crop-estimation programme; the inundation-mapping workflow shares the same change-detection foundations.

What the data cannot do, and what to ask for instead

Cloud cover during peak monsoon can reduce Sentinel-2 usable acquisitions to near zero for weeks at a time over parts of Bangladesh, the Mekong Delta, and eastern India. SAR fills most of that gap, but heavy rainfall causes surface roughness that can suppress the double-bounce signal, introducing false negatives for inundation at precisely the moments of highest flood extent. Validation with rain-gauge records or reanalysis precipitation data is a sensible quality check.

Commercial very-high-resolution SAR from systems such as ICEYE or Capella (sub-metre to 3 m) can resolve individual bunds and plot boundaries, which is useful for pilot-area validation but not cost-effective at national scale. The practical recommendation for most government clients is to use open constellations for national mapping and reserve commercial tasking for validation strips in representative districts. Ask for a stratified accuracy assessment with confusion matrices rather than a single overall accuracy figure; the latter conceals the classes that matter most.

Typical figures

Primary SAR spatial resolution	10 m (Sentinel-1 IW mode); 3–10 m (PALSAR-2, mode-dependent)
Primary optical spatial resolution	10 m visible/NIR (Sentinel-2); 30 m (Landsat 8/9)
SAR revisit (single satellite)	6 days at equator (Sentinel-1A+B combined); 14 days (PALSAR-2)
Optical revisit	5 days (Sentinel-2A+B combined); 8 days (Landsat 8+9 combined)
SAR frequency / wavelength	C-band 5.4 GHz / 5.6 cm (Sentinel-1); L-band 1.27 GHz / 23.6 cm (PALSAR-2)
Minimum reliably detectable flooded plot	Approximately 0.1 ha at 10 m SAR resolution; sub-hectare smallholder plots are ambiguous
Inundation detection latency	1–3 days after SAR acquisition depending on processing pipeline
Archive depth	Sentinel-1 from 2014; Sentinel-2 from 2015; Landsat from 1984 (Landsat 5)
Typical temporal sampling per season	8–15 SAR acquisitions per 90-day season at equatorial latitudes
Delivery formats	Cloud-optimised GeoTIFF, GeoPackage, WMS/WMTS tile service, tabular area statistics (CSV)

Analytics Satellize can run

Seasonal inundation probability map	Change detection on Sentinel-1 VV/VH backscatter time series using threshold or random-forest classifier trained on reference flood polygons	GeoTIFF layer per season showing per-pixel inundation probability; updated at each new SAR acquisition
Flood onset, duration, and drainage date rasters	Phenological date extraction from SAR backscatter time series (first exceedance of flood threshold, continuous flood duration, last exceedance)	Three-band GeoTIFF per season; tabular summary by administrative unit
Flooded paddy area statistics by district	Pixel counting within administrative boundaries after SAR-optical fusion classification; uncertainty estimated by bootstrapped confusion matrix from validation sample	CSV report with area estimates, confidence intervals, and data-gap flags for cloud-affected periods
Rice-fish co-culture extent map (where cadastral data provided)	Supervised classification combining SAR time-series features with client-supplied cadastral or survey polygons as training labels	GeoPackage of classified co-culture plots with probability score; accuracy report with per-class precision and recall
Multi-year wetland extent change analysis	Annual inundation composites from Landsat and Sentinel-1 archives; change vector analysis between user-defined baseline and current year	Change map (gain, loss, stable) as GeoTIFF; area transition matrix in CSV; narrative summary report
Production-area input layer for yield modelling	Flooded area per crop calendar window, integrated over flood duration, output as tabular input compatible with client yield-coefficient models	Structured CSV or JSON feed by district and dekad, with metadata on data completeness and SAR acquisition gaps

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.