A sector analysis published in 2026 reports two figures that define the problem this domain seeks to address. According to reinsurer data it cites, natural disasters caused about 224 billion dollars in worldwide damage in 2025, with insurers covering about 108 billion.
The same source notes another reinsurer recorded 107 billion dollars in insured losses across 190 events in 2025, with nearly half of total economic losses covered by insurance. In other words, roughly half of the damage is not.
That gap is the economic driver behind the use of Earth observation in this sector. This article describes how satellite observation is used in insurance, what it permits and what it requires in data work. It extends the article on Earth observation, security and sovereignty.
The four uses of Earth observation in the insurance cycle
Their requirements differ markedly depending on when they intervene.
Pre-underwriting risk assessment uses Earth observation to characterise a property and its surroundings, roof condition, flood exposure, proximity of vegetation.
Catastrophe modelling uses the footprints of past events to calibrate risk models, work resting on the archive rather than on recent acquisition.
Claims triage after an event uses rapid acquisition to prioritise interventions and estimate exposure, the security article having described this family as damage assessment.
And parametric insurance uses a satellite measurement as a direct payout trigger, a mechanism deserving separate treatment.
Those four uses share a decisive common feature. They replace or complement expensive ground observation, which is their primary economic justification.
The parametric mechanism
This arrangement is the sector’s deepest transformation.
Its principle deserves stating precisely. A specialist source describes it thus: the core mechanism of parametric insurance is that payouts are dictated by predefined physical parameters, such as a specific flood depth or wind speed, rather than by a subjective financial evaluation of individual property damage.
Three direct consequences follow. The payout delay is measured in days rather than months, since no individual assessment is required. Handling cost collapses, which makes risks of low unit value insurable. And the objectivity of the trigger reduces disputes.
A documented case illustrates the working. The analysis cited above reports that an African risk capacity group announced in October 2025 a combined parametric payout of just over 5.4 million dollars to support Mozambique after the 2024/25 drought season and a tropical cyclone, the payment going to the government and to a food assistance body.
The same source draws the lesson: money can move because an agreed measurement crossed a contract threshold, and not because a loss adjuster completed a slow site-by-site review.
Why Earth observation transformed this market
A specialist analysis formulates the mechanism with useful clarity.
The same source explains the rupture by comparison: before satellites, parametric had to rely mainly on ground-based weather stations, which limited the market to a handful of perils in places where stations happened to be dense enough.
That formulation reveals a direct link with the constellations article in this series. The densification described there did not merely increase data volume; it made insurable risks that were not, for want of an available measurement.
Three families of variable are now measurable and used as triggers: soil moisture and vegetation state, flood extent, and fire severity. Each corresponds to a peril previously hard to cover.
The particular role of radar
This technical capability is decisive in this domain.
A specialist publication describes the difficulty: during active catastrophes, thick blankets of smoke, heavy cloud cover or nighttime conditions can completely blind sensors, delaying damage assessments for days or weeks until the air clears.
The same source describes the response radar provides, its pulses passing through smoke, ash plumes and darkness, ensuring reliable acquisition in near real time.
That property, which the opening article flagged as a general advantage of radar, takes on direct economic value here. A parametric arrangement whose trigger cannot be measured during the event does not work.
One practical consequence follows for a data provider. Training and evaluation corpora on radar imagery are scarcer than those on optical, and radar annotation requires distinct competences few providers hold.
Basis risk, a structural limit of Earth observation triggers
This limit is inherent to the parametric principle itself.
The payout depends on an Earth observation measurement and not on actual damage. Two gaps result, symmetrical to one another and both problematic.
An insured party may suffer damage without the threshold being crossed, and receives nothing despite a real loss.
Conversely, a threshold may be crossed without significant damage, and a payout occurs with no corresponding loss.
Three factors determine the size of that basis gap. The quality of the measurement itself, including its spatial resolution and uncertainty. The relevance of the index chosen, a variable poorly correlated with damage mechanically producing a large gap. And granularity, an index computed over a wide area smoothing very different local situations.
Those three factors share a feature that is absolutely decisive for a provider. All of them depend on the quality of the data used to calibrate the index, which places ground truth at the heart of the product’s economic viability rather than at its periphery.
What these uses require in annotation
The need is greater here than in most other domains.
It concerns four distinct objects. Building footprints are the first of them. A publication in the field notes that satellite-derived footprints, heights and land use classify the assets at risk, which makes them the founding data of any exposure assessment.
Building characteristics are the second: roof type, apparent condition, visible materials. Those attributes feed pre-underwriting risk assessment.
Past event footprints are the third, and they calibrate catastrophe models, retrospective work resting on the archive.
And damage states are the fourth. One source notes that comparing satellite-observed damage patterns against building characteristics refines the damage functions of models.
That fourth object is by far the most demanding. It requires qualifying a damage level from an image, a judgement requiring an explicit convention and a reference, failing which annotators diverge, a difficulty the segmentation cluster documented.
The verification question
This particular use raises a methodological question.
Earth observation allows a claim to be verified through an independent measurement, which the sector presents as a way to cross-check declarations against objective evidence.
Three methodological cautions frame that use. The resolution available may not permit real damage to be established, a damaged roof not always being discernible. The image date must precisely bracket the event, failing which the observation proves nothing. And the absence of visible damage does not prove the absence of damage, a distinction the climate article expressed through the notion of uncertainty.
That third caution is the most important and the most often neglected. An arrangement treating a non-detection as proof of absence will produce unjustified refusals, with the consequences one imagines for the client relationship and for the method’s reputation.
Agriculture, a special case and a major market
It is the most established outlet for parametric Earth observation.
Crop insurance has long rested on indices, and Earth observation has progressively replaced ground measurement there.
Three Earth observation indices dominate that segment. Vegetation indices, computed from spectral bands and correlated with canopy vigour. Soil moisture, measured by radar and correlated with water stress. And rainfall accumulations, from products combining satellite and model.
That segment has three particularities. The risk is diffuse and of low unit value, which makes individual assessment economically impracticable and parametric cover natural. Seasonality structures everything, a trigger having to account for the crop’s development stage. And ground truth partly exists, cultivated area declarations being a source the opening article flagged as the least expensive.
That third particularity deserves emphasis. It makes this segment more accessible than others to a corpus building project, since the reference is not to be created entirely but completed and verified.
The players and their needs
Four families of interlocutor are distinguishable in insurance Earth observation.
Reinsurers hold substantial means and in-house modelling competence. Their need concerns calibration corpora and large-scale exposure data.
Direct insurers have more operational needs, centred on underwriting and claims handling, and they readily use service providers.
Parametric specialists, often young companies, have an acute need and constrained means, which points towards targeted engagements.
And public risk pooling bodies, of which the case cited above gives an example, operate on institutional logics close to those described in the security article.
One observation runs across those four families of interlocutor. The insurance sector is accustomed to reasoning in quantified uncertainty, which makes it demanding on that point and receptive to rigorous documentation, unlike other markets.
What the sector expects of Earth observation exposure data
One deliverable deserves describing because it is this domain’s most volume-intensive demand.
An exposure database lists insured assets and their characteristics, and its quality conditions every model resting on it.
Four attributes are expected. Precise location, the opening article having shown spatial resolution and positional accuracy are distinct notions. Ground footprint, which supplies an area. Height or storey count, which allows a volume and a value to be estimated. And use type, residential, commercial or industrial, which conditions vulnerability.
Three difficulties complicate its production. Distinguishing between adjoining buildings, which a single footprint masks. Outbuildings and annexes, whose inclusion is a convention to be fixed. And recent construction, absent from existing references and detectable only by temporal comparison.
That third difficulty is an opportunity. Updating an exposure database through change detection is a recurring need, of low unit volume and high value, which matches the temporal competence the constellations article identified as scarce.
From demonstration to product
This difficulty explains the gap between announcements and actual deployments in Earth observation.
A convincing demonstration on a few cases does not make an insurance product. Four requirements separate one from the other.
Stability over time. An index must produce consistent results across several seasons, which presupposes retrospective validation on the archive rather than a demonstration on the current year.
Geographic transferability. An index calibrated on one region does not transfer mechanically, a limit the opening article flagged as general and which becomes contractual here.
Guaranteed availability. A parametric product presupposes the measurement will be available when the event occurs, which excludes a source whose continuity is not assured and matches the caveat the market article raised.
And auditability. A contractual trigger must be verifiable by a third party, presupposing a documented and reproducible method rather than an opaque model.
That fourth requirement eliminates the most proposals. It places method documentation on a par with performance, a configuration the medical imaging cluster described in a regulatory context and which appears here under purely contractual constraint.
What this Earth observation market offers a provider
Four characteristics make it interesting.
The need is recurring rather than one-off. An insurance portfolio evolves, models are recalibrated and corpora must be updated.
The documentary requirement is real, which favours a methodical provider, but it is lighter than in regulated domains.
The geographic coverage sought is often broad, which gives value to a provider able to handle poorly served regions, a point the climate article flagged.
And the decision cycle is shorter than in institutional markets, insurance being a commercial sector accustomed to contracting.
One important caveat accompanies that picture. The sector is cautious and it evaluates a method at length before adopting it, which lengthens the first sale while stabilising subsequent ones.
The protection gap as opportunity
That figure indicates where the sector’s growth sits.
With roughly half of economic losses uninsured, extending cover is the principal reservoir. The analysis cited specifies those uncovered losses concentrate in lower-coverage regions and on perils such as flood, wildfire, drought and heat.
Three consequences follow for a data provider.
Demand concerns regions currently poorly served in reference data, which matches the observation made in the climate article on the value of a provider present where observation networks are sparse.
The perils cited are precisely those the satellite measures best, flood by radar, fire by thermal and drought by vegetation indices, which explains the convergence under way.
And the arrangements concerned are often public or mixed, as the case cited above illustrates, which brings this segment closer to the institutional logics described in the security article.
That configuration outlines a favourable position for a provider able to work on poorly documented geographies with a rigorous method, a combination that is not widespread.
Approaching an insurer with an Earth observation offering
These recommendations differ from those applicable to other sectors.
Talk about uncertainty rather than performance. An insurer reasons in distributions of outcomes rather than average scores, and a method presented without uncertainty characterisation will look naive to an interlocutor whose profession it is.
Address basis risk head on. Concealing it does no favours, since the interlocutor knows it and will assess it regardless. Treating it explicitly, indicating how calibration reduces it, places the discussion on methodological ground.
Offer retrospective validation on the archive rather than a demonstration on the current year, the stability requirement set out above and which few proposals satisfy.
And document the method auditably, that requirement being contractual in a parametric arrangement rather than merely desirable.
Those four recommendations share a feature. They mean addressing a sector that masters risk quantification better than most clients of a data provider, which makes rigour pay and approximation immediately visible.
Why insurers make demanding clients
One characteristic of this Earth observation market deserves stating plainly, because it changes what a provider must bring.
Insurance is the business of pricing uncertainty. An insurer does not need to be persuaded that measurement carries error; it needs to know the size and shape of that error to price around it.
Three consequences follow for a data engagement. A performance figure with no confidence interval is less useful than a lower figure with one, which inverts the usual sales instinct. A known failure mode is manageable and an unknown one is not, which makes disclosing limitations a strength rather than a weakness. And a method whose behaviour on edge cases is documented can be underwritten, while one whose behaviour is undocumented cannot be priced at all.
That third point is worth internalising. An insurer confronted with an undocumented method does not reject it because it is bad; it rejects it because no responsible price can be attached to something whose error distribution is unknown.
For a provider, this makes the insurance sector unusually rewarding of methodological honesty. The habits earlier clusters described as regulatory obligations are here simply what allows a commercial conversation to reach a number.
Common errors of reading
These misreadings recur often enough that naming them is usually enough to avoid them.
- Confusing a parametric payout with indemnification of actual damage.
- Neglecting basis risk when designing an index.
- Choosing a variable poorly correlated with the damage it is meant to represent.
- Computing an index over an area too wide for the granularity of the risk.
- Treating a non-detection as proof of absence of damage.
- Using an optical sensor for a trigger that must work during the event.
- Neglecting precise image dating in a claim verification.
- Annotating damage states with no explicit convention and no reference.
- Underestimating the scarcity of radar annotation competence.
- Presenting a method to an insurer without characterising its uncertainty.
- Demonstrating an index on the current year rather than on the archive.
- Proposing a contractual trigger resting on a non-auditable method.
Where the annotation work actually accumulates
A closing observation ties the four annotated objects to the commercial reality of this domain.
Three of them are built once and refreshed periodically: footprints, building characteristics and past event footprints. They represent substantial initial volume and modest recurring work.
The fourth behaves differently. Damage states are annotated after each significant event, which means demand arrives unpredictably, urgently and in concentrated bursts.
That asymmetry has a practical implication a provider should plan for. A relationship built on exposure data is steady and forecastable; one built on post-event assessment is intermittent and requires standby capacity. The two combine well precisely because their rhythms differ, the first funding the readiness the second requires.
A second implication concerns preparation. Post-event work cannot be scoped when the event happens, since the client needs results in days. The conventions, the damage scale and the quality criteria must be agreed in advance, during the calm period, which is another reason the two engagements belong together.
That is probably the most useful commercial structure this domain offers: a steady exposure mandate that funds and prepares an intermittent response capability, with the second being where the differentiation and the margin sit.
What to take away
Roughly half of the economic losses caused by natural disasters is uninsured, and that gap is the economic driver behind the use of satellite observation in this sector.
Three readings emerge. Parametric insurance pays on predefined physical parameters rather than individual damage assessment, which reduces delays from months to days and makes risks of low unit value insurable, one documented case having seen a payout triggered because an agreed measurement crossed a contractual threshold. Every time a satellite gains the ability to measure a new variable at a new resolution and a new frequency, a category of risk becomes insurable, which links constellation densification directly to market extension. And basis risk, the gap between the measurement and actual damage, depends entirely on the quality of the data that calibrated the index, which places ground truth at the heart of the product’s economic viability.
For the open data resource feeding part of these uses, the article on Copernicus describes its exploitation. For security uses with neighbouring requirements, the article on Earth observation, security and sovereignty examines the institutional framework.
To explore delivery arrangements, supported formats and applicable control mechanisms, see our dedicated page on geospatial data processing. And if you need a corpus of building footprints or damage states, let us discuss your project.
