Satellite and geospatial signals

← Web scraping and compliance in alternative data · News and sentiment analytics →

Counting the world from orbit: optical imagery, SAR, aerial and drone data as investment signals — what each can see, revisit-tradeoffs, and how raw pixels become tradeable indicators.

What satellites actually measure

Investment-grade geospatial signals reduce physical world states to counts and measurements: cars in parking lots, crude oil levels inferred from floating-roof tank shadows, ships at anchor off congested ports, construction progress visible as thermal signatures, crop health expressed as spectral indices, and container queues forming outside terminals. Each measurement family has a canonical vendor lineage and a known relationship to fundamentals — parking counts to retail revenue, tank levels to inventory draws, ship positions to commodity flows. The discipline's promise is precisely this mapping between observable physics and economic variables that firms otherwise learn about only quarterly.

Not everything measurable from orbit is economically interesting, and not everything interesting is measurable. Satellites excel where activity leaves a physical trace that is large, outdoors, and persistent: storage, transport, heavy industry, agriculture, and real estate construction. They struggle where value creation happens indoors, in software, or in services. Practitioners therefore start from the question of which physical chokepoints an industry possesses, and only then consider imagery; applying satellite methods to sectors without visible throughput tends to produce expensive confirmation of what filings already disclose.

Origins and development

Geospatial analysis as an investment input has older roots than the small-satellite era. Government programs such as Landsat maintained continuous multi-decade archives of moderate-resolution imagery, and platforms like Google Earth Engine lowered the barrier to analyzing those archives at planetary scale. Early quantitative applications focused on agriculture and commodities, where field sizes justified coarse resolution. The commercial turning point came when venture-backed imaging constellations began launching fleets of small satellites, making frequent revisit of arbitrary locations a purchasable service rather than a government capability.

Analytics firms then abstracted the imagery away from customers entirely. Companies such as Orbital Insight popularized the model of selling counts and indices — cars, tanks, ships — instead of pictures, targeting consumers who had no intention of interpreting pixels themselves. Radio-frequency geolocation and AIS ship-tracking specialists such as Spire Global added signals that do not depend on cameras at all, detecting activity through the emissions that vessels and devices naturally produce. Together these layers shifted the category from imagery licensing toward indicator delivery.

Optical, radar and the revisit tradeoff

Optical imagery is intuitive and comparatively inexpensive per scene but blocked by clouds and confined to daylight — tolerable for desert oil terminals and problematic for equatorial retail networks during monsoon seasons. Synthetic aperture radar sees through cloud cover and darkness, measuring metal objects and surface deformation, at higher cost and lower interpretability for non-specialists. Constellation size determines revisit rate: monitoring one facility weekly is straightforward with almost any provider, whereas monitoring ten thousand stores daily requires either many small satellites or acceptance of statistical sampling rather than a full census.

Aerial and drone layers fill the resolution gap between orbital sensors and ground truth — building-condition intelligence, rooftop assessments for insurers, vegetation encroachment surveys for utilities — trading breadth for detail. Radar specialists such as ICEYE complement optical operators like Planet Labs, and the best geospatial products blend sources deliberately: radar for guaranteed-cadence presence detection, optical for confirmation and context, aerial for inspection-grade detail. Fusion also hedges sensor-specific artifacts, since a detection confirmed across modalities is materially harder to dismiss as noise.

Market structure

The industry divides into vertically integrated operators that own constellations and sell data or analytics downstream, analytics firms that buy capacity across multiple operators and sell indicators, and distribution channels that host both. Operators compete on revisit rate, resolution, and spectral capability; analytics firms compete on detection quality, historical depth, and domain modeling. Much commercial distribution now flows through data marketplaces, including cloud-native venues such as AWS Data Exchange and Snowflake Marketplace, which shorten procurement cycles considerably compared with bespoke contracts.

For buyers, this structure carries a practical implication: exclusivity is rare at the imagery layer, since constellations sell capacity to many intermediaries simultaneously. Durable differentiation concentrates in the processing layer — annotation quality, baseline construction, entity resolution — and in contracts guaranteeing tasking priority for specific sites. When evaluating providers, asking who operates the sensor, who runs the detection pipeline, and whether either role is subcontracted clarifies where quality control actually lives and which failures the vendor can directly remediate.

From pixels to indicators

Raw imagery becomes signal through detection pipelines: car detectors trained on annotated parking lots, tank-diameter estimators derived from shadow geometry, ship classification from hull silhouettes, and change-detection models that flag new construction or disturbed earth. Quality differences between vendors live mostly in these pipelines — detection confidence handling, seasonal baselines, anomaly thresholds — rather than in the underlying photos, which increasingly originate from the same handful of constellations. Two vendors ingesting identical scenes can publish materially different counts, so methodology documentation deserves as much scrutiny as sensor specifications.

Output formats determine consumption cost. Per-site time series with confidence intervals integrate cleanly into factor models and dashboards; raw scenes require in-house computer-vision capability that few investment teams want to maintain, staff, or defend in model validation. Most institutional buyers therefore choose indicator-level products within the broader satellite and geospatial category and reserve scene-level work for bespoke investigations, such as verifying a specific facility claim ahead of an event-driven position. The tradeoff between interpretability and convenience recurs throughout alternative data and is particularly sharp here.

Where geospatial beats other exhaust

Physical verification is the unique asset: satellites observe what actually happened rather than what was reported, searched, or transacted digitally, making them immune to the reporting lags, revisions, and incentives that afflict disclosure-based measures. They dominate in commodities — inventory, flows, production — where supply chain chokepoints are visible; in emerging-market retail where digital exhaust is thin; and in industrial assets whose operators disclose little voluntarily. For agricultural systems, imagery-based indices pair naturally with the datasets covered in weather and agriculture analysis.

They underperform for services, software, and any value chain lacking physical chokepoints, and they face strong competition from complementary exhaust where such exhaust exists. Card panels often beat parking counts for developed-market retail, which is why practitioners frequently read the two together; comparisons with card transaction panels illustrate how each corrects the other's blind spots. Geospatial data's comparative advantage is therefore widest where money trails are absent, delayed, or distorted — precisely the settings where verification is hardest by conventional means.

Caveats: weather, baselines and crowding

Cloud cover creates irregular sampling that naive analyses misread as demand shocks; a fortnight of overcast skies over a retail region is a measurement gap, not a collapse in traffic. Baseline construction — knowing what normal occupancy looks like per site per season, adjusted for holidays, layout changes, and tenant turnover — is where analytical quality is won or lost, and vendors differ enormously in how carefully they maintain baselines. Popular sites also crowd: once a parking-lot signal becomes widely known, its predictive value compresses rapidly.

Crowding pushes vendors toward less obvious facilities and toward fusing imagery with complementary streams such as AIS ship tracks, weather feeds, and flight records, so that residual signal survives wider dissemination. Buyers should ask how a vendor detects regime changes at reference sites, how baselines are recalibrated after known disruptions, and whether published histories are restated when methodologies improve. Restatement policies deserve particular attention, because silently revised histories flatter backtests in ways that live performance will not reproduce.

Limitations and criticism

Serious limitations persist. Interpretation remains inferential: a full parking lot can reflect discounting, staffing cuts, or a neighboring store's closure as easily as organic demand, and disambiguation usually requires auxiliary data. Latency constraints matter for fast strategies, since tasking, downlink, and processing add delay between an event and its observation. Costs are non-trivial for global censuses of small sites, and ethical questions surround pervasive imaging of private property, particularly where individuals or restricted facilities appear in captured scenes.

Critics also caution against narrative excess: compelling aerial anecdotes occasionally outrun the statistical evidence behind them, and a visually striking detection can anchor attention more strongly than its predictive weight justifies. The mature view treats geospatial signals as one verified layer within a broader stack — corroborating or contradicting transactional, textual, and survey inputs — rather than as a standalone oracle. Under that framing, the category's contribution is distinctive because it measures the physical world directly, something no other alternative-data family accomplishes at comparable scale.