Irrigated area mapping using SAR coherence and backscatter
Sentinel-1 SAR coherence loss and seasonal backscatter shifts expose irrigated fields that optical imagery misses under cloud. This page covers the physics, the method, and the honest limits of mapping irrigation at national scale.
Sensors
- Sentinel-1 A/B (C-band SAR, 5.6 cm wavelength): Interferometric Wide Swath mode delivers 250 km swaths at 10 m ground range resolution with 6-day repeat at the equator (12-day with one satellite). The 6-day repeat is the minimum coherence interval; shorter intervals preserve coherence better over stable surfaces, making moisture-induced decorrelation from irrigation events clearly detectable.
- ALOS-2 PALSAR-2 (L-band SAR, 23.6 cm wavelength): L-band penetrates crop canopy and reaches the soil surface more effectively than C-band, giving stronger soil-moisture backscatter contrast. Fine-beam single polarisation offers 3 m resolution; ScanSAR covers 350 km at 100 m. Revisit is 14 days, which limits temporal sampling of fast irrigation cycles.
- RADARSAT Constellation Mission (C-band SAR): Three identical satellites provide 4-day exact repeat and up to daily revisit at high latitudes. Medium-resolution ScanSAR covers 500 km at 50 m; compact polarimetry modes add scattering-mechanism discrimination useful for separating irrigated rice from open water.
- Sentinel-1 archive (2014 to present): A continuous 10-plus year global archive enables multi-year seasonal composites that distinguish persistent irrigation infrastructure from anomalous wet years driven by rainfall. No other freely available SAR archive matches this temporal depth at C-band.
What a wet field tells a radar that a dry one does not
C-band radar interacts with agricultural surfaces through two mechanisms that irrigation disrupts. First, liquid water on or near the soil surface increases dielectric constant sharply, raising backscatter intensity, particularly in VV polarisation which is sensitive to surface roughness and moisture. Second, when soil moisture changes between two SAR acquisitions, the phase relationship between the two passes degrades: interferometric coherence drops. A field that stays dry between passes holds coherence; one that was irrigated in the interval loses it.
The magnitude of coherence loss depends on how much the surface changed, how quickly crops grew, and whether wind moved the canopy. Rain-fed fields in a dry season remain stable and coherent. Irrigated fields in the same period show repeated coherence drops timed to irrigation events. That temporal pattern, not a single image, is the mapping signal. A single pass tells you a field is wet today; a time series tells you it is irrigated systematically.
Building the seasonal composite: from passes to maps
The standard approach stacks coherence values from consecutive 6-day Sentinel-1 pairs across a full growing season, typically 20 to 30 pairs. Pixels that show persistently low coherence during the dry season, when rain-fed land is stable, are candidates for irrigation. A threshold is applied, or a supervised classifier trained on ground-truth points, to produce a binary irrigated or non-irrigated layer. Backscatter time series from the same stack add a second discriminant: irrigated fields show a characteristic wet-season-like backscatter signature even when rainfall records show drought.
At national scale this requires careful orbit-track mosaicking, since ascending and descending geometries produce different layover and shadow patterns over undulating terrain. Flat plains are straightforward. Terraced hill agriculture, common in South and Southeast Asia, requires terrain-corrected backscatter and coherence computed separately per track before mosaicking. Processing pipelines built on ESA's SNAP toolbox or equivalent handle this, but computation time scales with area. A country the size of Ethiopia requires meaningful cloud infrastructure to process a full season in reasonable time.
Where the method fails, and what to do about it
Natural wetlands, seasonally flooded grasslands, and rice paddies in their flooded establishment phase all produce coherence loss and elevated VV backscatter that resembles irrigated cropland. Separating them requires ancillary data: land-cover masks, soil maps, or optical imagery from cloud-free windows. L-band from PALSAR-2 helps because the double-bounce signal from flooded rice (water surface plus vertical stem) is spectrally distinct from the surface scattering of irrigated wheat or cotton.
Minimum field size is a hard constraint. At Sentinel-1's 10 m resolution, fields below roughly 0.5 hectares are unreliably detected because mixed pixels at field edges contaminate coherence estimates. Smallholder systems in sub-Saharan Africa and South Asia, where plot sizes routinely fall below 0.2 hectares, are substantially undercounted by C-band methods alone. PALSAR-2 fine-beam at 3 m improves this, but its 14-day revisit and limited free data access make it a supplement rather than a backbone. Honest mapping reports should state a minimum mapping unit explicitly.
Dense vegetation canopies in mature maize or sugarcane attenuate C-band signals enough to mask the soil moisture signal below. This is less a problem for mapping irrigated area, which is typically done at the start of the season when canopies are thin, than for in-season monitoring of irrigation events.
National-scale programmes: what the public record shows
FAO and national statistical agencies in India, Pakistan, and several African countries have used Sentinel-1 coherence methods to cross-check census-reported irrigated area, which is often politically sensitive and unreliably reported. Published studies in journals such as Remote Sensing (MDPI) have demonstrated country-scale mapping with overall accuracies in the 80 to 90 percent range for large-field systems, dropping to 65 to 75 percent in smallholder landscapes. Those figures are worth quoting to a buyer: good enough for policy planning, not good enough for field-level subsidy verification without validation sampling.
The Global Food and Agriculture Organisation's AQUASTAT database provides reference irrigated-area statistics that SAR-derived maps can be benchmarked against. Discrepancies between satellite estimates and AQUASTAT are themselves informative: they often reveal informal or undeclared irrigation expansion, which is exactly what a government water-resource ministry wants to know.
From map to decision: what a government actually receives
A raw irrigated-area polygon layer is rarely what a client needs. The useful outputs are change maps showing expansion or contraction of irrigated area year on year, disaggregated by river basin or administrative district, with uncertainty estimates attached. Water-resource ministries use these to allocate abstraction licences and detect unlicensed extraction. Agricultural ministries use them to project food production potential ahead of harvest. Both applications require the map to be reproducible annually on a consistent methodology so that trends are comparable.
Satellize structures analytics engagements around this kind of repeatable annual cycle. The Tonga crop-estimation programme illustrates the general approach: open-constellation data processed through a defined method, delivered as a structured output against which the client can act. For irrigation mapping, the equivalent output is a seasonal irrigated-area GIS layer with a documented minimum mapping unit, a confusion matrix from validation sampling, and a basin-level summary table. Clients who want to commission a scoping study for their country or basin can request a methodology note and indicative processing scope from the team.
Typical figures
| Primary sensor spatial resolution | Sentinel-1 IW mode: 10 m (range) × 10 m (azimuth) after multi-looking; PALSAR-2 fine-beam: 3 m; RCM medium: 50 m |
| Coherence pair interval | 6 days (Sentinel-1, single satellite); 12 days (Sentinel-1, one satellite operational); 14 days (PALSAR-2); 4 days (RCM) |
| Swath width | Sentinel-1 IW: 250 km; PALSAR-2 ScanSAR: 350 km; RCM ScanSAR: 500 km |
| Radar frequency and polarisation | Sentinel-1 and RCM: C-band (5.4 GHz), VV+VH dual-pol; PALSAR-2: L-band (1.27 GHz), full-pol and compact-pol modes available |
| Minimum mapping unit (practical) | ~0.5 ha at C-band 10 m; ~0.1 ha at L-band 3 m fine-beam; smaller fields substantially undercounted |
| Archive depth | Sentinel-1: 2014 to present (global); PALSAR-2: 2014 to present; ALOS PALSAR-1 precursor archive from 2006 |
| Cloud sensitivity | None: SAR penetrates cloud cover entirely, which is the primary advantage over optical methods in monsoon-affected regions |
| Typical seasonal stack size | 20 to 30 coherence pairs per growing season for Sentinel-1 at 6-day repeat; 8 to 12 pairs for PALSAR-2 |
| Reported mapping accuracy (large-field systems) | 80 to 90% overall accuracy in published studies; 65 to 75% in smallholder landscapes below 0.5 ha plot size |
| Deliverable formats | GeoTIFF (irrigated/non-irrigated binary or probability), GeoPackage polygons, basin-level CSV summary, QGIS-compatible project |
Analytics Satellize can run
| Seasonal irrigated-area map | Interferometric coherence stack thresholding or supervised classification over 6-day Sentinel-1 pairs across a defined growing season | GeoTIFF binary layer and polygon GeoPackage with minimum mapping unit documented; confusion matrix from validation sample |
| Year-on-year irrigation expansion or contraction | Consistent annual reprocessing of coherence composites; change detection between seasonal layers | Annual change GIS layer; basin-level summary table showing net hectares gained or lost per administrative unit |
| Informal irrigation detection | Comparison of SAR-derived irrigated area against AQUASTAT or national census reference; flagging of statistically significant discrepancies | District-level discrepancy report with satellite-derived estimate, reference figure, and percentage deviation |
| Irrigated area probability surface | Random forest or logistic regression classifier trained on coherence, VV/VH backscatter time-series features, and terrain covariates; outputs per-pixel probability rather than binary class | Continuous probability GeoTIFF; configurable threshold for downstream binary mapping |
| Wetland and rice paddy separation layer | Multi-sensor fusion of C-band coherence with L-band double-bounce index from PALSAR-2 and optical NDWI from cloud-free Sentinel-2 windows | Three-class GIS layer (irrigated cropland, flooded rice, natural wetland) with per-class accuracy statement |
| Irrigation infrastructure footprint | Canal and field-boundary delineation from high-coherence linear features in SAR amplitude; cross-referenced with OpenStreetMap water-infrastructure layer | Canal network polyline layer; command-area polygon delineating the area served by each identified infrastructure element |
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.