Environmental habitat index for cephalopod stock distribution
Satellite-derived sea surface temperature, chlorophyll, sea-level anomaly and mixed-layer depth feed species distribution models that forecast where squid and octopus are likely to concentrate, giving fishing fleets and quota-setting bodies a seasonal environmental picture weeks ahead of survey vessels.
Sensors
- MODIS Aqua/Terra: Daily global ocean colour and SST at 1 km (day/night SST) and 4 km (chlorophyll-a). The two-satellite tandem gives near-daily revisit at most latitudes, though cloud contamination in the western Pacific and Humboldt Current upwelling zones can reduce usable observations to 3-5 clear days per month in some seasons.
- VIIRS SNPP / NOAA-20: 750 m resolution ocean colour and 750 m/375 m SST bands. VIIRS has largely succeeded MODIS for operational fisheries oceanography; its improved signal-to-noise ratio in the near-infrared improves chlorophyll retrieval in turbid coastal margins where octopus habitat overlaps with shelf edges.
- Jason-3 altimeter: Measures sea-surface height anomaly (SSHA) along repeat ground tracks with a 10-day exact-repeat cycle and roughly 300 km inter-track spacing at the equator. SSHA is a proxy for mesoscale eddies that concentrate prey. Multi-mission merged products (CMEMS) reduce the effective sampling gap but cannot resolve features below about 100-150 km.
- GHRSST blended SST: Multi-sensor Level-4 analysis products (e.g. MUR at 1 km, OSTIA at 5 km) fuse infrared and microwave retrievals to fill cloud gaps with optimal interpolation. Microwave SST penetrates cloud but cannot resolve features closer than roughly 25-50 km to the coast, so nearshore octopus habitat near islands or fjords requires careful masking.
- Argo float climatology (mixed-layer depth): Mixed-layer depth (MLD) is not directly observable by passive satellite sensors. Operational products derive MLD from Argo float profiles blended with altimetry. Spatial resolution is coarse (roughly 1-degree grids) and latency can be 5-10 days, which limits MLD as a real-time input but makes it useful as a monthly climatological covariate in species distribution models.
Why cephalopods are unusually tractable from orbit
Most fish habitat modelling struggles because the target species occupies mid-water or demersal zones that satellites cannot see directly. Squid and octopus are different. Both are highly responsive to surface and near-surface thermal structure, and both track prey that aggregate at oceanographic features visible in satellite imagery: chlorophyll fronts, mesoscale eddies, upwelling filaments. The Humboldt squid (Dosidicus gigas) in the eastern Pacific and the Japanese flying squid (Todarodes pacificus) in the Japan Sea have been studied extensively in this context, and published species distribution models for both species use satellite-derived SST, chlorophyll and sea-level anomaly as primary predictors.
The physical logic is straightforward. Squid are ectotherms with narrow thermal tolerance windows. Dosidicus gigas, for instance, is typically associated with SSTs between roughly 15°C and 25°C in the eastern Pacific, though this range shifts with ontogeny and season. Sea-level anomaly identifies mesoscale eddies where vertical mixing concentrates mesopelagic prey. Chlorophyll-a is a proxy for the base of the food web that ultimately supports the myctophids and euphausiids on which squid feed. None of these variables predicts catch directly. Together, they define where conditions are permissive.
Building the habitat suitability index: inputs and architecture
A habitat suitability index (HSI) for cephalopods is typically constructed by normalising each environmental variable to a suitability score between 0 and 1, using response curves derived from historical catch-per-unit-effort data or from published thermal and trophic preference literature. The individual scores are then combined, usually as a geometric or arithmetic mean, to produce a spatial grid of habitat quality. More sophisticated implementations use maximum entropy modelling (MaxEnt) or boosted regression trees trained on observer-programme catch data, with the satellite variables as predictors.
For Todarodes pacificus in the Japan Sea, published work has shown that SST gradient magnitude (the sharpness of fronts, not just the temperature itself) improves model skill over SST alone. This is computationally straightforward: a spatial gradient operator applied to a daily SST raster. For Humboldt squid, sea-level anomaly has been identified as a particularly informative predictor during La Niña years when the thermocline shoals and the oxygen minimum zone expands, compressing the habitable vertical range and concentrating animals near the surface where they become more catchable. The satellite record for these variables extends back to the early 1990s for altimetry (TOPEX/Poseidon) and to 2002 for MODIS, giving a training archive long enough to capture multiple ENSO cycles.
Where the models break down
Honest limits matter here. The most significant failure mode is El Niño. During strong warm events, SST, chlorophyll and SSHA all shift outside the ranges that populated the training data. The 1997-98 and 2015-16 El Niño events caused documented collapses in Humboldt squid availability in the northern part of their range, and models trained on normal-year data gave poor spatial predictions during those periods. Post-hoc, the environmental signal was there in the satellite record; the problem was that the species' response to those conditions had not been adequately characterised in the training catch data.
Cloud cover is a persistent operational problem in the most productive squid grounds. The Kuroshio-Oyashio transition zone in the Japan Sea and the Humboldt Current upwelling system off Peru and Chile both experience extended cloud periods. A 7-day composite of MODIS or VIIRS SST reduces cloud contamination substantially but smooths out the short-lived frontal features that are often the most informative predictors. GHRSST blended products recover some of this through microwave fusion, but the 25-50 km coastal exclusion zone matters for shelf-edge habitats. Mixed-layer depth, as noted, is the weakest input: coarse, lagged, and interpolated rather than observed.
From index grid to fishing ground forecast
The operational output of an HSI model is a spatial probability surface, typically delivered as a GeoTIFF or NetCDF grid at the resolution of the coarsest input (usually 4-9 km, set by the altimetry merge). For seasonal forecasting, this surface is updated weekly or fortnightly as new satellite composites arrive. A fishing ground forecast adds a temporal dimension: the model is run forward using climatological anomaly persistence or a coupled ocean model nudged by satellite observations, projecting likely habitat conditions 2-4 weeks ahead.
Quota-setting applications are more demanding than fleet guidance. A regulator needs not just where squid are likely to be, but an estimate of relative abundance. HSI models do not provide this directly. They identify permissive habitat, not density. Bridging that gap requires coupling the satellite-driven HSI with a population dynamics model or with survey-derived abundance indices. The satellite component constrains the spatial allocation of effort and the identification of core versus marginal habitat, which is genuinely useful for spatial quota management even without an absolute abundance number.
Operational considerations for a government client
A fisheries ministry or regional fisheries management organisation deploying this approach needs to think carefully about three things: the quality of the historical catch data used to train the model, the latency requirements of the operational product, and the communication of uncertainty to quota-setting committees.
On latency: MODIS and VIIRS data are available within a few hours of acquisition through NASA EARTHDATA. GHRSST Level-4 products typically have a 1-2 day latency. The Jason-3 altimetry merge products from CMEMS run at roughly 5-7 days latency for the near-real-time stream. A weekly HSI update is therefore achievable with current infrastructure. On uncertainty: the honest position is that HSI model skill, measured against withheld catch data, is typically moderate, with published studies reporting AUC values in the 0.70-0.85 range for squid SDMs depending on species and region. That is useful but not infallible, and communicating it as a probabilistic map rather than a deterministic forecast boundary is both scientifically appropriate and practically important for managing fisher expectations.
Satellize runs this class of analysis on open constellations including MODIS, VIIRS and GHRSST blended products, with commercial altimetry data added on client licence. The architecture is the same one used in the Tonga crop-estimation programme: open-data ingestion, reproducible processing pipelines, and outputs designed for government decision workflows rather than research publications. A useful starting point for a new engagement is a hindcast validation: running the model over 3-5 years of historical satellite data and comparing the HSI surface against available logbook or observer data to establish baseline skill before any operational commitment.
Typical figures
| SST spatial resolution | 1 km (MODIS/VIIRS infrared, clear sky); 25-50 km (microwave, all-weather); 1-5 km (GHRSST blended L4) |
| Chlorophyll-a resolution | 4 km (MODIS standard product); 750 m (VIIRS); cloud-limited, 7-day composites typically used operationally |
| Sea-level anomaly resolution | ~25 km effective after multi-mission merge (CMEMS); 300 km along-track spacing for Jason-3 alone |
| Mixed-layer depth resolution | ~1-degree grid (Argo-derived climatology); 5-10 day latency for near-real-time products |
| Revisit / composite period | Daily acquisition (MODIS/VIIRS); 7-day composites standard for cloud reduction; altimetry 10-day repeat |
| HSI grid output resolution | Typically 4-9 km, constrained by altimetry merge; finer grids possible but not supported by input data quality |
| Operational latency (weekly update) | 1-2 days for SST/chlorophyll inputs; 5-7 days for altimetry; weekly HSI update achievable |
| Satellite archive depth | SST: 1981 (AVHRR); MODIS ocean colour: 2002; Jason altimetry series: 1992 (TOPEX/Poseidon) |
| Model skill range (published literature) | AUC 0.70-0.85 for squid SDMs depending on species, region and ENSO phase; degrades during strong El Niño |
| Delivery formats | GeoTIFF, NetCDF, GeoJSON contour layers; tabular CSV for time-series covariates; PDF seasonal briefing |
Analytics Satellize can run
| Weekly habitat suitability index grid | Normalised multi-variable HSI with geometric mean combination, response curves from published thermal/trophic preference literature or local CPUE calibration | GeoTIFF probability surface at 4-9 km, updated weekly, covering client-defined EEZ or fishing zone |
| Seasonal fishing ground forecast (2-4 week outlook) | Anomaly persistence or climatological regression applied to current satellite state; HSI projected forward using ENSO phase conditioning | Monthly PDF briefing with map panels and written uncertainty statement for fisheries management committees |
| Core versus marginal habitat classification | HSI threshold segmentation; core defined as upper quartile of suitability score across the seasonal distribution | GeoJSON polygon layer suitable for spatial quota allocation or closed-area boundary setting |
| ENSO phase sensitivity report | Hindcast model runs stratified by Oceanic Niño Index phase; skill metrics (AUC, Brier score) reported per phase | Annual technical report quantifying model reliability under different climate states; informs quota-setting confidence intervals |
| SST front and gradient magnitude layer | Sobel or Roberts gradient operator applied to GHRSST L4 daily SST raster; front persistence mapped over 7-day window | Weekly GeoTIFF gradient magnitude layer; documented as improving Todarodes pacificus model skill in published literature |
| Hindcast validation against logbook data | Retrospective HSI run over 3-5 year archive; spatial correlation and AUC computed against client-supplied CPUE logbook records | Validation report with skill metrics, bias maps and recommendation on model configuration before operational deployment |
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.