Sensors
- Sentinel-1 SAR (C-band, ESA): 10 m ground range resolution in Interferometric Wide Swath mode, 6-day repeat at mid-latitudes (12-day per pass), all-weather day-night imaging. Metallic cage frames and small steel-hulled or aluminium-hulled vessels produce strong radar cross-sections detectable above the sea-clutter background. Vessels shorter than roughly 10–15 m sit near the detection floor in standard IW mode; GRDH processing improves small-target discrimination.
- Sentinel-2 MSI (ESA): 10 m resolution in visible and near-infrared bands (B2–B4, B8), 20 m in red-edge and SWIR bands (B5–B7, B8A, B11, B12), 5-day revisit with two satellites combined. Chlorophyll-a proxies (B4/B3 ratio, NDCI using B5 and B4) and turbidity indices track feeding-induced phytoplankton blooms and suspended solids plumes around cage clusters. Cloud cover is the primary constraint; in persistently overcast coastal regions, usable acquisitions may average fewer than two per month.
- Planet Dove (PlanetScope): 3 m resolution, four-band (RGBNIR) optical imagery with near-daily revisit globally. At 3 m, service vessels of 5 m length or longer are marginally detectable as single-pixel or two-pixel anomalies; reliable vessel shape discrimination requires objects above roughly 10 m. The daily cadence is the primary value: change detection between consecutive clear-sky acquisitions isolates vessel arrival and departure events at individual cage clusters.
- Landsat 9 OLI-2 (USGS/NASA): 30 m multispectral resolution, 16-day repeat. Too coarse for individual vessel detection, but the Operational Land Imager-2's improved signal-to-noise ratio relative to Landsat 8 makes it useful for basin-scale water-quality mapping: chlorophyll, total suspended matter, and coloured dissolved organic matter indices across entire bay or fjord systems. Provides archive continuity back to 1972 via Landsat series for long-term concession-boundary compliance checks.
What cage infrastructure looks like to a radar
Aquaculture cage arrays are, from a radar's perspective, unusually cooperative targets. Steel and aluminium walkways, mooring buoys, and feed pipes form corner-reflector-like geometries that return strong backscatter in Sentinel-1 C-band imagery, often 10–20 dB above the surrounding sea surface. This makes cage positions detectable even in moderate sea states, which is why a concession that appears as featureless water on a nautical chart shows up as a bright cluster of point targets on a SAR amplitude image.
The complication is that small fibreglass or wooden vessels, which are common in artisanal and small-scale aquaculture operations, have much lower radar cross-sections than steel-hulled craft. A 6 m fibreglass skiff may fall below the detection threshold in standard Sentinel-1 IW GRDH processing, particularly when sea clutter is elevated. Distinguishing an aquaculture service vessel from a passing artisanal fishing boat is harder still: both may be similar in size, material, and position relative to the cage boundary. Temporal stacking, looking for vessels that appear repeatedly at the same cage cluster across multiple passes, is more reliable than any single-image classification.
Reading harvest cycles from vessel visit patterns
Harvest events are logistically intensive. They require multiple vessel trips, often involving well-boats, harvest barges, or refrigerated transport vessels that are considerably larger and more radar-reflective than routine feeding skiffs. A spike in vessel detections at a concession, combined with a simultaneous drop in cage-area backscatter intensity (consistent with reduced biomass in nets), is a recognisable SAR signature. The backscatter reduction occurs because dense fish biomass attenuates the radar return from the net structure itself.
Stocking events produce a different pattern: a period of low vessel activity following harvest, then a pulse of small vessel trips consistent with smolt or juvenile stocking, followed by gradual biomass accumulation visible as increasing cage backscatter over subsequent weeks. Neither signature is unambiguous on its own. Cross-referencing with water-quality indices from Sentinel-2 adds confidence: heavy feeding periods correlate with localised chlorophyll increases and dissolved oxygen depletion proxies downwind of the cages, while harvest periods show a rapid clearing of those signals.
Revisit cadence limits the precision of cycle dating. With Sentinel-1 at 6-day repeat and Sentinel-2 at 5-day repeat (cloud permitting), the timing of a harvest event can typically be bracketed to within a few days, not pinpointed to a single day. For regulatory compliance purposes, that precision is often sufficient. For commodity-market timing, it may not be.
Water colour as a feeding and stress indicator
Salmon, sea bass, and shrimp farms discharge nutrient-rich effluent continuously. In enclosed bays with limited tidal exchange, this drives measurable phytoplankton responses detectable in Sentinel-2 band ratios. The Normalised Difference Chlorophyll Index, computed from bands B5 (705 nm) and B4 (665 nm), is sensitive to chlorophyll concentrations in the range of roughly 5–50 µg/L, which covers the eutrophic conditions common near intensive cage sites. Turbidity increases from uneaten feed and faecal matter show up in B4 and B3 reflectance.
These signals are proxies, not direct measurements. Atmospheric correction quality, bottom reflectance in shallow water, and mixing with riverine inputs all introduce uncertainty. In water shallower than approximately 10 m, bottom albedo contaminates the water-leaving reflectance signal and makes chlorophyll retrieval unreliable without explicit bathymetric correction. Analysts should treat Sentinel-2 water-quality outputs as relative indicators of change rather than absolute biogeochemical measurements unless validated against in-situ sampling.
Encroachment on protected zones: where the boundary actually is
Aquaculture concession boundaries are defined in national licensing databases, not always in formats that align neatly with satellite coordinate systems. Checking whether cage infrastructure or vessel activity has expanded beyond a licensed boundary requires co-registering the concession polygon against SAR-detected cage positions, with a realistic understanding of Sentinel-1's geolocation accuracy. In IW mode with precise orbit data applied, geolocation error is typically below 10 m, which is adequate for concession boundaries defined to that precision but not for boundaries drawn to 1–2 m accuracy.
Protected marine area boundaries present a related problem. A vessel detected within a marine protected area is not automatically an encroachment: transit is usually permitted, and the vessel's purpose cannot be inferred from position alone. What satellite analysis can establish is a pattern. A vessel appearing at the same location inside a protected zone on multiple dates, particularly one that correlates spatially with cage-like SAR returns, is a different evidentiary situation from a single transit detection. Regulators using this kind of analysis should treat satellite evidence as a trigger for inspection rather than a standalone enforcement action.
Separating aquaculture vessels from artisanal fishing craft
This is the hardest discrimination problem in the use case, and it is worth being direct about the limits. At Sentinel-1's 10 m resolution, a 12 m aquaculture feed barge and a 12 m artisanal fishing vessel are essentially identical in their SAR point-target signature. Planet Dove at 3 m can resolve hull shape for vessels above roughly 15 m, but most small-scale aquaculture and fishing vessels fall below that threshold. The discrimination therefore has to come from context and behaviour, not from the sensor signature alone.
Contextual rules that improve classification: aquaculture service vessels tend to stop at fixed cage locations and remain stationary for extended periods; artisanal fishing vessels move more continuously and do not repeatedly return to the same sub-hectare position. Spatial proximity to known cage infrastructure, derived from the SAR cage-detection layer, is the strongest single discriminator. Vessels detected within 50 m of a confirmed cage cluster on multiple passes are overwhelmingly more likely to be service craft than fishing boats. This is a probabilistic argument, not a certainty.
Satellize's Tonga crop-estimation programme demonstrated that combining multi-sensor time series with spatial context layers substantially reduces ambiguity in small-target classification problems, even when no individual sensor provides definitive identification. The same logic applies here.
What this analysis cannot do
Satellite observation cannot determine species, fish health, or actual stocking density from orbit. It cannot count individual animals. Harvest tonnage estimates derived from biomass-proxy methods carry wide uncertainty bands and should not be presented as production statistics without ground-truth validation. Night-time operations, common in some harvest workflows, are detectable by Sentinel-1 but not by optical sensors, so a purely optical time series will systematically miss nocturnal activity.
Persistent cloud cover in tropical and sub-polar aquaculture regions (Norway, Chile, Indonesia, the Philippines) can reduce Sentinel-2 clear-sky acquisitions to fewer than one per month during winter or monsoon seasons. SAR fills the gap for vessel and cage detection but cannot substitute for the water-quality information that only optical bands provide. Any operational monitoring programme should account for these data gaps in its reporting cadence rather than presenting incomplete time series as continuous coverage.
Typical figures
| SAR spatial resolution (Sentinel-1 IW GRDH) | 10 m range × 10 m azimuth |
| Optical spatial resolution (Sentinel-2 visible/NIR) | 10 m (B2, B3, B4, B8); 20 m (B5–B7, B8A, B11, B12) |
| High-resolution optical (Planet Dove) | 3 m, 4-band (RGBNIR) |
| Revisit cadence | Sentinel-1: 6 days (two-satellite); Sentinel-2: 5 days (two-satellite); Planet Dove: approximately daily |
| Minimum detectable vessel (SAR) | Approximately 10–15 m steel or aluminium hull in IW GRDH; fibreglass vessels of similar size may fall below detection threshold |
| Water-quality chlorophyll sensitivity (Sentinel-2 NDCI) | Approximately 5–50 µg/L in optically deep, low-turbidity water; unreliable in water shallower than ~10 m without bathymetric correction |
| SAR geolocation accuracy (Sentinel-1, precise orbits) | Typically below 10 m CE90 |
| Archive depth | Sentinel-1: from 2014; Sentinel-2: from 2015; Landsat series: from 1972; Planet Dove: from approximately 2016 |
| Cloud-cover constraint (optical sensors) | Persistent cloud in tropical and sub-polar regions may reduce usable Sentinel-2 acquisitions to fewer than 2 per month |
| Delivery formats | GeoTIFF change-detection layers, GeoJSON vessel-detection events, CSV time-series tables, PDF monitoring reports |
Analytics Satellize can run
| Cage-array position and extent map | SAR amplitude thresholding and CFAR (Constant False Alarm Rate) detection on Sentinel-1 GRDH time stack; persistent bright-target clustering to separate fixed infrastructure from transient vessels | GeoJSON polygon layer of confirmed cage clusters, updated on each new Sentinel-1 acquisition |
| Service vessel visit log per cage cluster | Temporal SAR change detection: vessel-sized point targets appearing within 50 m of confirmed cage positions, flagged per acquisition date | CSV event log with date, cage cluster ID, vessel count, and confidence score; optional GIS point layer |
| Harvest cycle probability index | Multi-sensor fusion: SAR cage backscatter trend (biomass proxy) combined with vessel-visit spike detection and Sentinel-2 chlorophyll clearance signal; rule-based scoring against published aquaculture phenology patterns | Monthly PDF report with harvest-event probability scores per concession and supporting time-series charts |
| Water-quality anomaly alerts | Sentinel-2 NDCI and turbidity index time series; statistical anomaly detection against site-specific baseline; Landsat 9 OLI-2 used for basin-scale context | Automated alert (email or API) when chlorophyll or turbidity index exceeds user-defined threshold at a monitored concession |
| Concession boundary compliance layer | Co-registration of SAR-detected cage and vessel positions against licensed concession polygons and marine protected area boundaries; geolocation uncertainty buffer applied explicitly | GeoJSON compliance layer flagging detections outside licensed boundary, with uncertainty radius annotated per detection |
| Vessel-type probability classification | Contextual Bayesian scoring: vessel position relative to cage infrastructure, dwell time, repeat-visit frequency across SAR time series; Planet Dove shape features where vessel exceeds ~15 m | Per-detection probability score (aquaculture service vs. artisanal fishing vs. unclassified) appended to vessel event log |
| Long-term concession expansion analysis | Landsat archive time series (1972-present) and Sentinel-1 stack to map cage-array footprint changes over multi-year periods; area measurement with stated resolution-limited precision | Annual change report with mapped concession footprint per year and statistical summary of expansion or contraction |
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.