Aquaculture farm extent mapping for fisheries-sector credit exposure

Coastal aquaculture ponds have a distinctive spectral and geometric signature detectable from orbit. Mapping their area, fill status and abandonment rate gives lenders and insurers a ground-truth check on fisheries-sector exposure in Southeast Asia and China that no borrower-supplied figure can replicate.

Sensors

Sentinel-2 MSI: 10 m resolution in visible and NIR bands, 5-day revisit at the equator with both satellites. The NIR band (Band 8, 842 nm) is the primary discriminator: pond water absorbs strongly in NIR, giving a sharp contrast with vegetation and soil. Free archive from 2015.
Planet SuperDove: 3 m resolution, 8 spectral bands including red-edge and NIR, daily revisit over most of Southeast Asia. Resolves individual small ponds (0.5 ha and above) that Sentinel-2 undersamples, and captures within-pond density gradients from turbidity.
WorldView-2/3: 0.3–0.5 m panchromatic, 1.2–1.8 m multispectral, tasked on demand. Used for ground-truth validation of pond boundaries and detection of internal pond infrastructure (aerators, feeding platforms) that correlates with stocking density and farm value.
Landsat 8/9 OLI: 30 m resolution, 16-day revisit, free archive from 1972 (Landsat 5/7/8/9 combined). Coarser than Sentinel-2 but indispensable for multi-decade trend analysis: tracking pond expansion into mangrove, long-run abandonment rates and regional exposure growth.

What a rectangular pond tells a balance sheet

Aquaculture is among the fastest-expanding forms of coastal land use in Southeast Asia and southern China, yet the physical asset base underpinning fisheries loans and insurance policies is rarely verified independently. Borrowers self-report pond area. Insurers price policies on declared stocking density. Neither figure is audited from the air.

Satellite imagery changes that arithmetic. A fish or shrimp pond has three properties that make it detectable from orbit: it is inundated (low NIR reflectance), it is geometrically regular (rectangular bunds, consistent aspect ratios), and it is spectrally distinct from natural water bodies because pond managers maintain turbidity and algal load at levels that shift the blue-green colour balance measurably. These properties persist across cloud-free acquisitions, making pond mapping a tractable classification problem rather than an interpretive one.

The physics of pond water versus the sea

Natural coastal water and managed pond water occupy different positions in spectral space. In the NIR (roughly 750–900 nm), both absorb strongly, but the ratio of green-to-NIR and blue-to-NIR reflectance differs because pond water carries higher concentrations of phytoplankton, suspended sediment and dissolved organic matter. Sentinel-2's Band 3 (green, 560 nm) and Band 8 (NIR, 842 nm) together produce a Modified Normalised Difference Water Index (MNDWI) signal that separates managed ponds from tidal channels and estuaries with reasonable reliability in clear conditions.

Pond geometry reinforces the spectral signal. Object-based image analysis algorithms that combine spectral thresholds with shape metrics (rectangularity, compactness, bund-width regularity) outperform pure pixel classifiers, particularly where turbid coastal water reduces spectral contrast. At 10 m resolution, ponds above roughly 0.25 ha are consistently detectable in Sentinel-2 data. Below that threshold, Planet SuperDove's 3 m imagery is needed. WorldView-2/3 is reserved for validation and for high-value individual farm assessments where internal infrastructure matters.

Fill cycles and abandonment: the signals lenders miss

A static pond-area map answers one question: how much collateral exists? A time series answers a more useful one: is it operating? Shrimp ponds in Vietnam, Thailand and Indonesia typically run two to three production cycles per year, each involving a fill, a grow-out period of 60–120 days, and a harvest drain. This cycle is visible in multi-temporal MNDWI stacks. A pond that is consistently full across a 12-month Sentinel-2 archive is in production. One that is dry or in irregular fill patterns may be fallow, diseased or abandoned.

Abandonment is a material credit risk. White-spot syndrome and early mortality syndrome have caused rapid, large-scale pond abandonment in several Southeast Asian countries over the past two decades. A lender with a portfolio of aquaculture loans in affected provinces has no way to detect this from financial statements alone until the borrower stops servicing debt. A quarterly satellite-derived fill-status layer, run over the loan portfolio's geographic footprint, provides an early-warning signal that is independent of borrower reporting.

Mangrove encroachment complicates the picture at pond edges. Where mangroves have recolonised abandoned bunds, the pond boundary becomes ambiguous in optical imagery. This is an honest limit of the method: edge-detection accuracy degrades in transitional zones, and a conservative area estimate is the appropriate response rather than an aggressive one.

Cloud cover, turbidity and the honest limits of optical mapping

Coastal Southeast Asia sits under persistent cloud for much of the year. The Mekong Delta, the Gulf of Thailand coast and the Sundarbans all experience multi-week cloud cover during monsoon. This is not a fatal constraint but it is a real one. Sentinel-2's 5-day revisit means that over a 90-day window, most areas accumulate enough cloud-free observations for a reliable composite. Planet's daily tasking provides denser temporal sampling but at higher data cost.

Turbid coastal water is a harder problem. In the Mekong Delta and parts of the Riau Islands, suspended sediment from river discharge raises the reflectance of natural water in the green and red bands, compressing the spectral gap between natural water and pond water. Classification accuracy drops measurably in these conditions. Published studies using Sentinel-2 data in the Mekong Delta report overall accuracies of roughly 85–92% for pond detection, with the lower end occurring in high-turbidity coastal zones. Lenders should treat pond-area estimates in these regions as lower bounds, not precise figures.

SAR (Synthetic Aperture Radar) can detect inundated surfaces through cloud, but it does not cleanly separate managed ponds from natural water bodies without the geometric and spectral context that optical imagery provides. The two methods are complementary, not interchangeable. This page covers the optical approach; SAR-based inundation mapping is addressed separately in the sibling page on parametric flood insurance trigger verification.

From map to credit metric

The analytic output a lender needs is not a classified image. It is a structured table: farm identifier, centroid coordinates, mapped area in hectares, fill-status classification (active, fallow, abandoned), and a change flag relative to the previous assessment period. That table can be joined to a loan register by geographic proximity, producing a portfolio-level exposure summary that distinguishes performing collateral from degraded or missing assets.

Insurance underwriters have a related but distinct need. For parametric aquaculture products, the trigger is typically a physical event (typhoon, flood, disease outbreak) rather than a yield shortfall, but the exposed area still needs to be verified at policy inception and at renewal. A satellite-derived pond-area layer, updated annually or after a significant event, provides an auditable basis for sum-insured calculations that is independent of the insured party.

Satellize runs this class of analytics on Sentinel-2 and Planet time series for clients with coastal agricultural exposure, drawing on the same classification pipeline used in its Tonga crop-estimation programme. Engagements typically begin with a baseline extent map for a defined administrative region, followed by a quarterly update layer and an alert protocol for rapid abandonment events.

Archive depth and what history reveals

Landsat's archive extends to the early 1970s, with consistent OLI data from Landsat 8 available from 2013 and Sentinel-2 from 2015. For a lender assessing a coastal province in Vietnam or Fujian, a decadal time series reveals not just current pond extent but the trajectory: how rapidly the sector expanded, where it is now contracting, and which sub-regions show chronic instability. That historical context is relevant to portfolio-level credit decisions in ways that a single-date map is not.

The practical archive depth for high-confidence pond mapping is roughly 2015 to present using Sentinel-2, with Landsat extending the trend line back another decade at coarser resolution. For individual farm-level due diligence, Planet's commercial archive provides sub-annual detail from approximately 2016 onwards in most of Southeast Asia.

Typical figures

Spatial resolution (mapping)	10 m (Sentinel-2), 3 m (Planet SuperDove), 30 m (Landsat 8/9)
Spatial resolution (validation)	0.3–1.8 m (WorldView-2/3, tasked)
Revisit frequency	5 days (Sentinel-2, dual satellite); daily (Planet SuperDove); 16 days (Landsat 8/9)
Minimum detectable pond area	~0.25 ha at 10 m (Sentinel-2); ~0.05 ha at 3 m (Planet SuperDove)
Key spectral bands	Green (560 nm), NIR (842 nm) for MNDWI; red-edge (705–740 nm) for vegetation/bund separation
Cloud impact	Single-date accuracy degrades significantly; 90-day composites restore reliability in most regions
Turbidity impact	Classification accuracy ~85–92% in Mekong Delta conditions; lower in high-turbidity coastal zones
Archive depth	2015–present (Sentinel-2); 2013–present (Landsat 8); ~2016–present (Planet, most of SE Asia)
Geographic coverage	Global; primary focus Vietnam, Thailand, Indonesia, Philippines, southern China
Delivery format	GeoTIFF classified raster, GeoJSON/Shapefile polygon layer, CSV portfolio-join table, PDF assessment report

Analytics Satellize can run

Baseline pond-extent map	Object-based image analysis combining MNDWI thresholding with shape metrics (rectangularity, compactness) on Sentinel-2 or Planet composites	GeoJSON polygon layer with area (ha) and centroid per farm unit; PDF summary by administrative district
Quarterly fill-status classification	Multi-temporal MNDWI stack analysis; ponds classified as active, fallow or abandoned based on inundation consistency across the period	Updated GeoJSON layer with fill-status attribute; change flags vs. prior quarter; CSV for portfolio join
Abandonment early-warning alert	Threshold-based anomaly detection on rolling 90-day MNDWI time series; alert triggered when pond transitions from consistently filled to dry across two or more consecutive acquisitions	Alert report with affected pond IDs, coordinates, area and first-detection date; delivered within 5 days of trigger
Decadal expansion and contraction trend	Annual pond-area classification on Landsat 8/9 OLI archive (2013–present), extended with Sentinel-2 from 2015; area change computed per administrative unit	Time-series chart and GeoTIFF stack; tabular area-by-year summary for lender portfolio context
Individual farm due-diligence report	High-resolution WorldView-2/3 tasking combined with Planet time series; internal infrastructure detection (aerators, feeding platforms) used as stocking-density proxy	PDF report with annotated imagery, mapped area, fill history, infrastructure inventory and confidence assessment
Portfolio exposure summary	Spatial join of satellite-derived pond polygons to lender loan register by geographic proximity; active area and abandonment rate aggregated per loan	CSV/Excel table with loan ID, mapped collateral area, fill status and change flag; ready for credit-risk workflow ingestion

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.