RESEARCH DOSSIER
Insurance Prediction with Images and Remote Sensing
Abstract. Insurance prediction with images and remote sensing integrates the physical observation of insured assets and their environments with actuarial inference about exposure, claim frequency, severity, catastrophic loss, and claims operations. This monograph examines the scientific conditions under which street-level photographs, inspection imagery, aerial surveys, optical satellite observations, synthetic-aperture radar, and multispectral time series contribute valid predictive information. Its central analytical distinction concerns the gap between an observed image, the feature represented by that image, the physical event or condition of an insured asset, and the contractual loss realized by an insurer. A rigorous observation-to-insurance chain requires geospatial entity resolution, acquisition-time control, sensor-specific uncertainty, policy-period alignment, causal and predictive target definition, spatial and temporal validation, and the separation of visual damage proxies from adjusted claims. The synthesis integrates actuarial representation-learning studies, disaster-damage datasets, primary satellite mission documentation, geospatial validation methods, probability scoring, insurance data-quality standards, and jurisdiction-dependent governance. A transparent numerical illustration connects image-based segmentation with frequency, pure premium, and loading arithmetic while explicitly avoiding empirical performance claims. The evidence supports the qualified utility of imagery for selected insurance tasks and establishes testable conditions for determining when it should change a prediction, prompt an inspection, or remain ancillary. No new standalone theory or proprietary empirical discovery is claimed.
Keywords: Insurance prediction; remote sensing; satellite imagery; street-view imagery; image representation learning; property insurance; underwriting; claim frequency; claim severity; expected loss; geospatial entity resolution; exposure data; multispectral imagery; synthetic-aperture radar; catastrophe damage assessment; crop insurance; basis risk; spatial cross-validation; temporal leakage; probabilistic calibration; algorithmic fairness; actuarial data quality; privacy; scientific reproducibility.
Introduction
An insurer observes a policy, an asset, a territory, and eventually a claim. Images introduce a different observation channel: they encode spatially located optical or radar measurements, captured by an apparatus at a particular date and processed into a representation that a researcher can interpret or use in a predictive model. Their value depends on whether the representation improves the estimation of a clearly defined insurance quantity. A photograph of a roof may supply information about observable condition; a satellite time series may reveal vegetation dynamics; a radar acquisition may indicate inundation. None of these observations by itself specifies the insurance policy's covered peril, deductible, limit, causation, adjustment method, or final payment.
Blier-Wong, Lamontagne, and Marceau (2024) provide a directly relevant actuarial demonstration in their insurance pricing study using images. The authors construct street-view representations that can enter generalized linear insurance models and report predictive information for claim frequency for certain home-insurance perils. Their substantive result justifies examining the imagery channel as an observable risk signal. It does not establish that every image improves every insurance target, that the same association transports to other countries, or that image-based pricing automatically satisfies social and regulatory standards.
The research problem extends beyond image classification. Insurance prediction requires an explicit event horizon, exposure denominator, contract definition, and observed outcome. A remote-sensing pipeline must therefore answer several separate questions: what the sensor captured; what the extracted features mean; which insured object the observation concerns; whether the observation was available before the decision; whether the feature predicts the relevant loss; and whether the expected improvement survives comparison with traditional actuarial information and the cost of acquiring, governing, and updating imagery. These questions organize the present monograph.
The method is analytical and comparative. Primary institutional documentation establishes sensor capabilities and constraints; peer-reviewed papers establish particular algorithms and observed empirical findings; professional standards and regulatory materials establish jurisdiction-specific controls. Equations and the worked illustration derive established statistical relationships transparently. Statements about current operational performance are limited to what the cited material actually supports. This distinction protects the analysis from treating plausible uses of technology as already validated insurance results.
Scientific Object, Prediction Targets, and Scope
The unit of analysis may be a building, insured location, parcel, roof, agricultural field, vehicle, commercial facility, or aggregate geographical cell. The insurance observation is usually a policy-exposure unit indexed by asset i, contract c, period t, and peril k. Images may instead be organized by raster cell, camera viewpoint, polygon, satellite tile, or acquisition event. A foundational requirement is to specify the map between the observation units before estimation. A neighborhood image is not a policyholder record, and a building footprint is not necessarily the unit of insured coverage.
For a peril-specific prospective frequency model, let N_{ickt} be the count of qualifying covered claims during a policy period, and let e_{ickt} be earned exposure in years. A model may estimate E[N_{ickt} | X_{ict}, I_{i,\leq t_0}], conditional on traditional attributes X and imagery available at underwriting time t_0. For severity, it may estimate E[Y_{ickt} | N_{ickt}>0, X, I], where Y measures a specified kind of claim cost. Pure premium is expected total covered claim cost per exposure unit. These targets differ because the observation processes differ: claim occurrence, reporting, coverage adjudication, and settlement are distinct stages.
The distinction between risk prediction and damage detection is fundamental. A post-event detector estimates a visual condition D from before-and-after imagery; the insurer's indemnity L also depends on physical damage mechanisms, insured interest, contract terms, excesses, limits, exclusions, valuation, and claim development. Formally, a conceptual chain runs I → observed features Z → estimated condition D → contract-specific L. The successive arrows require separate evidence. The Copernicus damage-assessment manual explicitly describes remotely assessed damage as a proxy, rather than ground-truth damage information. This caveat is consequential for insurance valuation.
The scope of this dossier includes property and catastrophe underwriting, loss-prevention inspections, agricultural and parametric observations, and the assessment of claims-related images. It excludes the invention of a universal satellite-to-premium formula, a proprietary image data set, or an empirical performance benchmark unsupported by actual insured portfolios. It also distinguishes insurance application from the more general computer-vision subject of extracting image representations. The contribution is the integrated interpretation of measurement, insurance target, validation, and governance across these settings.
Images as Measurements of Insured Reality
A digital image is an instrument-mediated sample of a physical scene. The sensor records reflected, emitted, or backscattered energy, and subsequent processing converts raw measurements into pixels or engineered products. Optical visible imagery often makes façades, roofs, and land cover interpretable to human analysts; near-infrared and shortwave-infrared bands are useful for vegetation and moisture-related investigations; synthetic-aperture radar records microwave backscatter and has distinct sensitivity to surface geometry, roughness, dielectric properties, and acquisition geometry. The evidentiary meaning of a pixel depends on its band, resolution, calibration, geometry, and scene conditions.
At the asset level, the observed variable can be written Z_i = g(S_i, Q_i, V_i, T_i, P_i) + ε_i, where S is the physical state of interest, Q the sensor and processing configuration, V viewpoint and visibility, T acquisition time, P scene and atmospheric conditions, and ε measurement error. This is a schematic measurement representation, not a new physical imaging theory. Its usefulness is to make explicit why an image feature that correlates with a roof condition in one imaging regime can fail under altered lighting, season, orientation, construction style, or resolution. Ground observations and controlled inspections remain necessary for validating the meanings assigned to features.
Spatial resolution determines what structures can be resolved, not what perils can be insured. A 10-meter multispectral pixel can be informative for fields, vegetation, land cover, and some hazard surfaces, while being too coarse to diagnose the fastening condition of a single residential roof. A high-resolution aerial image may represent roof geometry but still conceal structural weakness, previous repair history, occupancy, or interior water damage. Conversely, a street-facing image can show wall finishes, floor count, and visible maintenance while omitting the rear roof slope and surrounding terrain. The observation window defines the inference window.
There is also a temporal measurement problem. A recently captured image can show a condition after a repair or hazard event, whereas an older image may show the state that actually existed when a policy began. Both are potentially useful for different targets; using the wrong acquisition date changes the meaning of the estimate. Consequently, source provenance, observation date, seasonal context, geographic footprint, and quality flags are essential model variables rather than administrative afterthoughts.
Street-Level, Aerial, Satellite, and Radar Sources
Street-level and inspection photographs provide a close view of structures and materials. They can support feature identification, the verification of questionnaire fields, site triage, and targeted risk engineering. Their weaknesses include obstruction by trees and vehicles, inconsistent camera orientation, differences in geographic coverage, and privacy and licensing limitations. The Blier-Wong et al. (2024) study demonstrates one practical approach to transforming these images into representations suited to a traditional pricing model, rather than insisting that an image classifier itself should perform all actuarial computations.
Aerial imagery and drone surveys increase geometric detail for roofs, building footprints, site exposure, and post-event visual inspection. Their utility depends on flight permissions, weather, sensor calibration, photogrammetric corrections, geographic referencing, and site access. Higher spatial detail can increase sensitivity to roof or façade changes but can also make model performance dependent on flight altitude, viewing geometry, and the vendor's processing pipeline. Evaluation must compare data acquisition and human interpretation costs with the actual decision improvement.
For broad-area monitoring, Sentinel-2 multispectral optical observations provide 13 bands at nominal 10-, 20-, and 60-meter ground sampling, with a commonly cited two-satellite five-day repeat for a location. The Sentinel-2 instrument specification gives the precise band-resolution distinction. Orbit revisit describes opportunities for acquisition, not guaranteed clear-sky coverage. The constellation's composition and availability change over time, so research datasets should record the actual product and satellite identities instead of treating a nominal interval as an observation guarantee.
Landsat 8 and 9 provide a longer operational observation context with a different resolution and repeat design. The USGS Landsat acquisition schedule describes a 16-day cycle for each satellite and an eight-day offset when combined. Those figures establish temporal sampling opportunities rather than an exact usable-image interval after cloud, snow, or atmospheric exclusions. Landsat can support land-use, vegetation, thermal, and environmental exposure features; at the individual-building level, its resolution generally requires aggregation or other data.
Sentinel-1 uses C-band synthetic-aperture radar and can observe day and night through cloud conditions. The official Sentinel-1 mission documentation identifies its all-weather imaging role. Radar is especially relevant to large-area flood detection, land-surface changes, and repeated environmental monitoring, but it introduces its own interpretation difficulties. Geometric shadow, layover, wind effects on water surfaces, surface roughness, and dense urban fabric can confuse automated classifications. Sensor complementarity should therefore be evaluated at the decision level rather than assumed because more bands or images were added.
A sixth category consists of preprocessed land-cover and hazard products. The CORINE Land Cover inventory offers standardized European land-cover classifications but has spatial and update limitations. Such products are valuable for contextual stratification, while historical categories may be stale for newly developed suburbs or rapidly changed landscapes. Derived product resolution and date must travel with every feature used in insurance pricing.
Data Provenance, Acquisition Times, and Image Quality
A defensible image record identifies provider, acquisition instant or interval, instrument, product level, coordinate reference system, georegistration and orthorectification, spatial footprint, pixel scale, processing history, licensing conditions, and confidence or quality indicators. The crucial question is whether two apparently comparable tiles are comparable observations. Different atmospheric corrections, mosaicking strategies, color adjustments, and resampling operations can produce changes unrelated to insured reality.
For underwriting, a prospective data cut must be defined before a model is trained. If t_q is the quotation time and t_a the acquisition time, permissible predictive information must satisfy t_a ≤ t_q, together with an allowance for the real-world time at which the image became available to the insurer. An image technically acquired before quotation but not published until after the decision is also unavailable as a genuine prospective predictor. Backfilled provider archives are a common source of hidden temporal leakage. An audit should save both acquisition time and effective availability time.
Missingness is structurally informative. A cloudy region, a rural road without street coverage, a building behind trees, and a region lacking commercial high-resolution imagery are missing for different reasons. Omitting these cases may concentrate training data in wealthy, urban, well-imaged areas. Imputing an image feature as though absent imagery represented an average building conceals the distinct selection mechanism. Availability itself may be predictive of region or exposure, creating an apparent gain that should be examined for fairness and transferability.
The professional obligation to examine relevant information is illustrated by the Actuarial Standards Board's ASOP No. 23 on data quality. In its United States professional context, the standard addresses the selection, review, and use of data and relevant limitations. Applied to imagery, those obligations translate into a documented suitability judgment for sensor data, geospatial matches, derived feature validity, and missingness. The standard is an example of professional guidance within its stated jurisdiction, not a global license to use imagery in every insurance decision.
A reliable workflow retains the source product identifier and its transformation lineage. If a roof-condition model is updated, the insurer should be able to determine which imagery, model version, threshold, and reviewer produced an individual signal. Reproducibility can be impaired when vendors overwrite mosaics, replace building polygons, or withdraw access to older panoramas. Preservation therefore includes permissible retention, reproducible identifiers, and versioned derived measurements, subject to data rights.
Geospatial Entity Resolution and Policy Alignment
An image footprint is spatial; a policy record is legal and temporal. Their integration requires resolving a named or geocoded address to an insured asset and checking that the asset is covered by the relevant contract. A building might have several addresses, multiple tenants, condominium units, associated outbuildings, and separate policies. A satellite pixel may overlap several parcels; a parcel may contain multiple buildings; a street-view camera may be located closer to the neighboring property than to the intended structure.
Useful linkage attributes include normalized address, building footprint, parcel geometry, geocode precision, entrance location, occupancy, policy-effective interval, elevation, and site inspection evidence. The join should preserve both the inferred match and its confidence. A one-to-one database key must be earned by evidence, not inferred from nearest-neighbor geometry alone. When a matching polygon overlaps multiple insured buildings, polygon-level context may support a group feature, while asset-specific inference remains unresolved.
Geospatial errors can propagate into loss models. Suppose a vegetation feature is assigned to an adjacent parcel that contains mature trees while the insured roof is exposed and treeless. The model may learn a real association for the neighborhood and yet provide the wrong asset-level explanation. For coastal flood risks, a small elevation or location error can change hazard exposure sharply. It is therefore necessary to distinguish the geographic uncertainty of a feature from the statistical uncertainty of its coefficient.
Policy alignment introduces another layer. The same physical dwelling can carry different limits, deductibles, exclusions, replacement cost conditions, and periods of coverage. If a remote-sensing feature estimates susceptibility to flood, it does not reveal whether flood is covered. Claims must be assigned to the correct contract, peril, earned exposure, and accounting development window. The underlying measurement may be stable while the contractual outcome changes with policy terms.
Entity-resolution audits should report unmatched records, ambiguous matches, duplicate structures, multi-policy buildings, changes in address or boundary conventions, and matching error sampled against ground truth. Their rates belong in the model evaluation because linkage is part of the predictive process. Removing only the most difficult records during validation will overstate real deployment utility.
Representation Learning and Feature Extraction
Computer vision supports both interpretable engineered features and latent representations. Engineered features include roof area, visible wear indicators, tree proximity, building outline, vegetation index, impervious surface share, and water-inundation likelihood. Learned representations map an image to an embedding vector that summarizes patterns discovered by a neural network. A compact embedding can enter a conventional statistical model, giving image information a defined, separately testable place in the actuarial specification.
In the representation-learning actuarial study, Blier-Wong et al. (2024) use information from street-view images within a generalized linear framework. Their design is a useful methodological precedent because it distinguishes extracting representation from relating the representation to insurance outcomes. A practical implementation may fit a vision encoder on permitted imagery with auxiliary labels and then estimate a claim model with encoder outputs and traditional policy features. Whether the encoder remains frozen, is fine-tuned, or is updated by hazard-specific learning must be recorded.
Interpretation requires care. An embedding may encode roof material and vegetation, but it may also encode neighborhood investment, camera equipment, geographic style, season, and provider selection. Good predictive fit does not identify which latent dimension measures a physically causal hazard. Interpretable auxiliary labels and controlled changes can help test a mechanism; they cannot retroactively convert correlation into a causal intervention effect.
Feature extraction should preserve uncertainty. A binary label such as damaged roof discards a large amount of information when the model is only moderately confident. Instead, one may retain a probability distribution over observable conditions, append a quality score, or refer the case to human inspection. For high-cost individual decisions, the relevant output often becomes a decision-support signal with its confidence and limitations rather than an automatic final classification.
Design comparisons should include a baseline using only contract and actuarial variables; a model with conventional geographic factors; a model with transparent image-derived attributes; and an additional model with learned embeddings. The incremental value of an image source is its measured improvement over the best appropriate simpler alternative in a genuinely independent, time-aligned evaluation. Large improvements over a deliberately weak baseline do not prove the economic superiority of an imaging program.
Property Exposure, Underwriting, and Risk Engineering
Property underwriting can use imagery to validate observable fields in an application, prioritize inspections, and investigate changes after policy inception. Examples include roof geometry, exterior material indicators, surrounding vegetation, adjacency, access routes, visible debris, land use, and proximity to mapped water bodies. These are observational covariates with varying reliability. A visible proxy for roof age must be validated against actual age and inspection records; a pixel-level land-cover label should not be presented as a verified individual safety measure.
The insurance purpose determines acceptable error. For a broad inspection queue, a model that reliably ranks locations with probable visible deterioration may have value even when its probability calibration is imperfect. For a material premium change or adverse underwriting decision, the evidence threshold is higher: false positives can impose financial consequences on an asset owner, and model explanations must survive individual review. The relevant benchmark is often the marginal improvement in decisions after inspecting flagged cases.
FEMA's National Risk Index dataset supplies U.S. county- and tract-level information about hazards, annual expected losses, and social vulnerability and resilience. It is useful as a contextual comparator, but its aggregate hazard indicators are not a substitute for a property's policy-specific expected insured loss. A local vulnerability indicator also requires care when used in individual underwriting because geography may stand in for protected or socially sensitive characteristics.
United States government oversight has identified limits in aggregate risk mapping. The Government Accountability Office's 2026 review of FEMA's index discusses methodological and data limitations. This provides an important general lesson: official hazard layers can be valuable, while remaining uncertain and intended for a specific planning scale. Combining many official layers in an underwriting model does not eliminate their individual scale and coverage limitations.
The insurer should maintain a distinction between a signal suitable for asset-owner confirmation, a signal suitable for specialist inspection, and a well-validated input to a portfolio model. Those uses entail different error costs. An imagery system is an operational prediction technology only when its data capture, reviewer workflow, audit record, and downstream decisions have been measured together.
Catastrophe Imagery and Post-Event Damage Assessment
After a flood, earthquake, hurricane, wildfire, or severe storm, satellite and aerial data may rapidly characterize affected areas. Pre-event imagery gives a reference state; post-event imagery supports change detection and apparent damage classification. The peer-reviewed xBD building-damage dataset paper by Gupta and colleagues (2019) describes pre- and post-disaster satellite imagery with building annotations and ordinal damage labels. Such data support method development and benchmarking; they do not automatically yield insurer-specific adjustment outcomes.
Damage labels need an ontology. A model might classify visibly undamaged, possible damage, damaged, or destroyed buildings, while claim payments require legal coverage and valuation. The Copernicus emergency mapping damage-class interpretation expressly includes categories for uncertainty and damage not observable from above. This prevents treating a single overhead view as complete evidence of building condition. Loss-adjustment ground truth must be derived from verified inspections, repair estimates, and coverage determinations where available.
Imagery availability is event-dependent. Smoke, cloud, darkness, ongoing rain, inaccessible airspace, and acquisition schedules affect which observations can be obtained. Copernicus's 2026 explanation of Earth observation limitations warns that damages may be underestimated or missed. The underwriting and claims implications are asymmetric: a confirmed visible destruction signal may be operationally urgent, whereas the absence of a visible damage signal should rarely be interpreted as confirmation that the property is intact.
Radar improves observation under clouds but creates different classification errors. Mason and colleagues (2010), in their urban flood detection study, report performance differences when shadow and layover regions are included. Those estimates belong to a particular TerraSAR-X urban flood study and must not be transported as generic performance metrics for all sensors or countries. They illustrate the value of separately measuring detectable and undetectable scene conditions.
Event-based validation should hold out entire disasters or event-region combinations. If a model trains on nearby buildings and tests on other buildings in the same storm, it may benefit from repeated acquisition style, terrain, damage distribution, and neighborhood context. That experiment evaluates interpolation within a familiar event rather than performance on a new catastrophe. Reporting both outcomes is informative; substituting one for the other is misleading.
Agriculture, Vegetation Time Series, and Index Insurance
Agricultural insurance introduces different units: fields, farm holdings, crop species, growing seasons, agricultural regions, and insured yields or revenue. Remote sensing may observe vegetation greenness, canopy development, moisture-related changes, crop classification, flood extent, drought conditions, and harvest timing. These features can contribute to yield estimation and portfolio monitoring, but satellite spectral values do not directly establish the insured production outcome of a particular farmer.
Temporal feature construction is essential. A vegetation index computed at one date is sensitive to crop species and development stage. A useful time-series model may align observations to phenological phases, planting schedules, irrigation conditions, and regional agronomic practices. A field's spectral response is also affected by atmospheric masking, soil background, mixed pixels, and management. A prediction trained in one climate region can lose validity in another even if the sensor remains unchanged.
Index-based insurance links payment to an observable trigger rather than direct loss adjustment. This creates basis risk: the index can trigger when a farmer incurs little loss or fail to trigger despite substantial crop damage. Carter, de Janvry, Sadoulet, and Sarris (2017) reassess the economic promise and limitations of these instruments in their agricultural index insurance review. Better satellite measurement may reduce some signal error, while contractual basis risk remains where the index and actual loss are imperfectly aligned.
Satellite-trigger design should therefore be evaluated against farmer-level losses and independently observed agricultural outcomes, not only satellite classification accuracy. Researchers need to specify the contract's payout schedule, spatial aggregation, missing-data rules, trigger revisions, and exceptional-event procedures. A low vegetation index can represent drought, crop choice, harvest stage, or management change; the trigger interpretation depends on the agreed insured hazard and observation timing.
The result is a limited but important role for remote sensing: it can make some previously unobservable or expensive-to-observe states measurable at scale, while leaving contractual and economic questions open. An index becomes an insurance instrument through explicit rights and payout rules. An image-derived index remains a scientific observation until its relationship to those rules and to losses has been established.
Claims Operations and Evidentiary Interpretation
Claims applications can receive photographs from policyholders, adjusters, repair contractors, drones, and remote-sensing systems. Their evidentiary roles differ. A submitted close-up may show interior damage unavailable to satellites; a satellite tile may show regional hazard exposure but omit a policyholder's basement flood; an adjuster's photographs may record repairs undertaken after the event. A system should preserve the source and time of every observation to avoid treating inconsistent evidence as an undifferentiated image corpus.
For claims triage, an image model may rank high-urgency cases, support early catastrophe reserve estimates, or identify apparent mismatches warranting inspection. This does not mean the model should adjudicate a claim solely from the image. A defensible claims process connects image-derived findings to causation, policy wording, independent evidence, customer submissions, and appeal or review. Fraud-screening labels particularly require care: a visual inconsistency can be a legitimate signal for investigation and is not proof of dishonest conduct.
The assessment pipeline should distinguish physical damage probability from payment probability and expected payment conditional on verified coverage. A machine can estimate that a roof looks damaged, while the insurer may owe no payment because the damage predates the policy or falls within a contractual exclusion. Conversely, covered internal water damage can occur without visible exterior change. Predictive uncertainty and coverage uncertainty must be reported separately.
Operational measurement should include time-to-inspection, time-to-payment, claim reopenings, reversals on appeal, inspection workload, and customer-level adverse outcomes, alongside accuracy metrics. Speed is valuable when it improves service quality and safety; it is insufficient as a stand-alone validation criterion if it increases incorrect denials or systematically underdetects losses for hard-to-image properties. The objective is a defensible evidentiary decision process rather than image classification for its own sake.
Spatial Dependence, Leakage, and Validation Design
Geospatial observations exhibit dependence. Nearby properties share weather, flood plains, construction practice, maintenance regimes, and socio-economic context. Neighboring pixels in an image overlap in physical content. Multiple views of the same structure may share nearly identical features. Randomly assigning these observations to training and test sets can produce a test set that is nominally independent by row number but substantively dependent on the training data.
Valavi, Elith, Lahoz-Monfort, and Guillera-Arroita (2019) show why spatial separation matters in their spatial cross-validation methodology. Although their application is ecological modeling, the general concern applies to spatially autocorrelated insurance records. Separate geographic blocks, spatial buffers, and environmental strata answer different generalization questions. Block size should reflect correlation scale and deployment purpose rather than a single universal distance.
A rigorous insurance split usually has several levels. First, all images of one asset should remain together across train and test. Second, exposure records for a common insured structure should be managed to prevent recurrent-policy leakage. Third, catastrophe loss experiments should group claims by event; satellite mosaics and acquired tiles can introduce shared acquisition artifacts. Fourth, an out-of-time cohort should reflect actual deployment chronology, including data access delays and claim development. Fifth, a held-out region should test transport to different construction, hazard, and sensor contexts.
Target leakage is more subtle than duplicate images. An image taken after the claim, a repair photograph uploaded during adjustment, an inspection ordered because of a prior high-risk score, or a geocoding correction completed after an event can all encode future information. Leakage audit should examine each variable's earliest availability date and the process that led to observation. Merely dropping an explicit claim-date column leaves these pathways intact.
Validation should report results by peril, geographic region, building type, sensor vendor, acquisition vintage, observation quality, and imagery availability. A model that performs very well where high-resolution data are common and poorly elsewhere should not receive a single unqualified portfolio accuracy statement. A scientific results table must show the population included and excluded from the evaluation.
Calibration, Uncertainty, and Economic Evaluation
Insurance decisions rely on the magnitude of probabilities and expected costs, not just on ranking. A calibrated model predicting a 5% annual claim probability should, in an appropriately defined comparable group and follow-up interval, exhibit approximately the claimed event frequency. Calibration and discrimination answer different questions. An image model can improve ranking yet overstate risk for all houses in a region. Pricing, reserves, and solvency consequences require further assessment.
The statistical foundation for proper probabilistic evaluation is developed by Gneiting and Raftery (2007) in Strictly Proper Scoring Rules, Prediction, and Estimation. For a binary event y_i and forecast probability p_i, the mean Brier loss is the average of (p_i − y_i)^2. For a predictive loss distribution, a suitable proper distributional score may test more than a point estimate. All scores should be computed on independent observations with a clearly specified outcome horizon; insurer-specific weights and accounting targets require separate explanation.
Model uncertainty has several levels: imaging noise, asset match ambiguity, model parameter variance, irreducible event randomness, incomplete claim reporting, missing acquisition, and distribution shift. A single confidence score for a vision classifier does not combine these uncertainties automatically. One can report feature-level intervals, calibrated risk probabilities, sensitivity under alternate matching rules, and scenario-specific loss forecasts. Prediction intervals for aggregate catastrophe loss must also incorporate shared-event dependence.
Economic value is decision-specific. If the use case is inspection prioritization, compare the expected avoided loss or operational gain with the cost of imaging, image processing, skilled reviews, false inspections, data rights, and customer appeals. If the use case is pricing, measure whether out-of-time expected loss is more accurately estimated and whether the resulting rates meet applicable regulatory requirements. A statistically significant reduction in an error metric can have little economic value if the decision threshold remains unchanged.
A transparent evaluation reports the baseline and enhanced model using the same cohort, available-at-decision features, exposure denominator, and period. It reports uncertainty for the difference in performance, ideally with clusters appropriate to space and events. It documents any tuning on the validation sample and reserves a genuinely untouched test for the final claim. Performance must be compared against human inspection and structured risk data, not just a model with no image information.
Sampling, Geographic Transfer, and Fairness
Imagery coverage is unequally distributed. Urban roads are often photographed more frequently than remote settlements; cloud patterns vary with climate; aircraft acquisition favors accessible or commercially valuable regions; inspection images may be collected after an insurer has already selected a high-risk property. Each mechanism affects both training selection and the meaning of missing observations. The deployment population may be systematically different from the observed image population.
Transfer requires examining differences in building materials, roof pitch, local hazards, address quality, agricultural systems, sensor conditions, and insurance terms. A model trained on North American detached houses may be poorly adapted to apartment blocks, informal structures, tropical roofing, or other construction traditions. Satellite imaging can offer broad geographic reach, but global imagery availability does not imply globally exchangeable insurance outcomes or legally permissible use.
Fairness analysis should separate an image's physical content from its social correlations. House size, façade appearance, vegetation maintenance, neighborhood condition, and street-image availability may be correlated with income, historic public investment, or differential access to insurance. An apparent improvement in actuarial discrimination can simultaneously alter access or affordability for groups who are unevenly represented in training data. Whether such differentiation is permitted or reasonable is a substantive jurisdiction-specific question.
Evidence should include coverage and error-rate comparisons across regions and relevant populations, explanations for missing data, adverse-impact analysis where lawful and appropriate, and oversight of model-driven decisions. The NAIC model bulletin announcement records the U.S. insurance-regulatory emphasis on governance, consumer impact, and compliance with existing insurance law. The model bulletin is an adopted NAIC model measure; its application depends on individual jurisdictions and their laws, rather than transforming into a single automatically operative nationwide statute.
A useful analytical conclusion is that geographic transfer and fairness are linked. Where a model is poorly observed, it may be less accurate; where a model is highly observed, it may use additional characteristics with social correlates. The research task is to establish which image features have a stable relationship to insured hazards, which perform as proxies, and which require alternative treatment. These questions demand local empirical evaluation before deployment.
Image Rights, Data Protection, and Regulatory Governance
The technical availability of a street-view or satellite product is separate from the legal right to store it, derive features, train models on it, or use it in commercial insurance decisions. Publicly accessible imagery can be governed by license terms, intellectual property rules, personal-data law, and restrictions on reuse. The provider's actual current contract and any commercial agreement should govern the proposed processing, rather than an inference that public visibility means unrestricted use.
Google's Street View Static API policies describe restrictions on storing, scraping, exporting, and deriving content, together with attribution requirements and contractual qualifications. Those constraints are material when interpreting studies that use images from commercial mapping platforms. A research result showing predictive promise does not by itself authorize a new insurer to replicate the training pipeline or retain historical imagery under the same license. Rights review should precede model design and archive policy.
In the European Union, the General Data Protection Regulation may apply when imagery and its linked predictions involve identifiable persons. Article 22 establishes conditions and safeguards relating to decisions based solely on automated processing that have legal or similarly significant effects; applying it requires examining the actual controller, processing, decision, exceptions, and safeguards. Not every photograph is personal data, and the legal analysis depends on identifiability and use. In an insurance setting, linkage to a named policyholder may make data-protection questions much more salient.
The European AI Act service guidance identifies AI systems used for risk assessment and pricing in life and health insurance as a defined high-risk use case, while distinguishing other insurance operations. Property-insurance imagery models should therefore be assessed under their actual purpose and applicable legal regime instead of being automatically classified in the same manner. Dates of applicability, sector-specific requirements, and national supervisory practice require separate current checks at deployment.
The European Insurance and Occupational Pensions Authority's 2025 AI governance opinion addresses national insurance supervisors and clarifies expectations under insurance-sector rules, including data governance, explainability, fairness, cybersecurity, and human oversight. In the United States, application of NAIC model measures and state insurance requirements is jurisdiction-dependent. NIST's voluntary AI Risk Management Framework can support structured governance, but its voluntary character must not be mistaken for a universal insurance statute.
An insurance imagery audit should therefore include a rights-and-retention register, lawful processing assessment, third-party provider evaluation, human-review controls, feature and decision explanations, correction and challenge procedures, incident records, and model-change logs. These controls enable scientific reproducibility and appropriate governance, but the exact legal obligations belong to each relevant jurisdiction and use case.
Reproducible Numerical Illustration
Consider a hypothetical property portfolio with 1,000 earned policy-year exposures, one specified peril, and 40 qualifying claims during the observation period. Assume every claim has the same illustrative severity of $6,000. Total claim cost is therefore $240,000 and the empirical portfolio frequency is 40/1,000 = 0.04 claims per exposure-year. The corresponding pure premium is $240 per exposure-year. These are constructed arithmetic values and must not be interpreted as an empirical study or a forecast for any real insurer.
Suppose a permitted, retrospectively observed image feature separates 600 policy-year exposures into a lower-signal group with 15 claims and 400 policy-year exposures into a higher-signal group with 25 claims. The observed group frequencies are 15/600 = 0.025 and 25/400 = 0.0625, respectively. At the stipulated common severity, group pure premiums are $150 and $375. The exposure-weighted average, 0.60 × $150 + 0.40 × $375, equals the original $240. The segmentation reallocates estimated costs between groups within the example; it does not create portfolio cost or improve prediction by arithmetic alone.
Now assume a fixed expense provision of $30 per exposure-year, variable expense equal to 25% of written premium, a 5% premium-based contingency allowance, and a separate illustrative 5% premium-based underwriting-profit provision. Under these explicitly assumed conventions, a premium P satisfies P = 240 + 30 + 0.25P + 0.05P + 0.05P. Therefore P = 270/0.65 = $415.384615... per exposure-year. If all per-exposure fixed expenses and loading percentages are held identical across the image groups, their illustrative premiums are (150+30)/0.65 = $276.923077... and (375+30)/0.65 = $623.076923..., giving the same weighted $415.384615... when aggregated. The calculation is a demonstration of the pricing denominator, not an endorsed rate design.
The numerical segmentation is in-sample by construction. It provides no evidence that the imagery feature will predict future claims, that either group's frequency is stable, or that segmentation passes any jurisdiction's insurance rating tests. The appropriate next experiment would fix the image acquisition and feature algorithm before the prospective policy period, estimate model parameters on one cohort, evaluate on a spatially and temporally independent cohort, and compare the enhanced model with a structured-data baseline using identical exposure and claims definitions.
The example illustrates a broader analytical point. A visually plausible split may redistribute expected premiums while leaving aggregate cost unchanged. Scientific value arises only when the split yields reproducible out-of-sample risk information and an acceptable operational decision. If the additional field is mislinked, outdated, or associated mainly with a socioeconomic proxy, the arithmetic remains valid while the insurance interpretation fails.
Reproducible Research and Model Audit
A reproducible project begins with a written estimand: the contract population, peril, exposure period, observation time, event definition, claim amount basis, and decision in which the prediction is used. The cohort and outcome construction should be specified before image feature selection. Policy dates and claims development must be aligned, and exclusions and missingness documented. Without this step, an apparent improvement may reflect changes in the eligible cohort rather than image information.
The source register should then list image provider, permission, geography, license, sensor, resolution, product level, date, cloud or visibility status, processing version, footprint, asset-match confidence, and retained identifier. A feature dictionary should connect each derived variable to the code or algorithm that generated it, the possible physical interpretation, and known limitations. Third-party models deserve the same provenance as in-house models.
Comparative experiments should include at least four arms: structured actuarial data; structured data with established geographic covariates; structured data with inspectable image features; and the same base with learned image embeddings. Models should be fitted to training data only; feature standardization, image encoder adaptation, and threshold tuning must occur inside training and validation boundaries. Test results should be reported on both all eligible records and the subset with usable imagery, so that improvement is not manufactured by selecting the easiest properties.
Error analysis should link failed cases to observation mechanisms. Misclassified damage after a storm may be associated with roof shadows, smoke, sensor angle, structural complexity, or the difference between visible damage and interior loss. A crop index may fail in a field with irrigation, early harvest, or mixed crop types. A geocode error may swap an insured building with its neighbor. Such case analyses identify repairable data problems and limits of the underlying observation method.
Finally, an independent research archive should contain the analytical definitions, source versions where retention is permitted, feature logic, random seeds, split definitions, evaluation scripts, computational environment, sensitivity analyses, and decision audit. If raw image licensing prevents redistribution, document the restriction and preserve reproducible product identifiers and nonrestricted derived statistics. Reproducibility means making the scientific inferences independently assessable within legal access constraints; it does not require redistributing data without permission.
Integrated Scientific Analysis
The examined evidence supports a bounded claim: images can add risk-relevant information to insurance prediction where they measure features associated with insured physical processes and where a suitable actuarial or claims model captures those associations in independent validation. The Blier-Wong et al. result provides direct home-insurance evidence for certain frequency predictions. The xBD corpus provides a separate benchmark for satellite-based apparent building-damage classification. Copernicus and sensor documentation define what is observationally possible. Those findings contribute to different inferential stages and should not be collapsed into a claim that satellite images can directly determine premiums or final indemnities.
Three linked inferential separations govern the synthesis. The first is between physical measurability and insurance materiality: an image may accurately reveal a feature that adds little prediction beyond existing underwriting variables. The second is between predictive correlation and operational portability: a good validation score under shared geography and provider conditions may not survive a new region, sensor, or event. The third is between a predicted loss component and a lawful, appropriate insurance action: pricing, inspection, claims settlement, and index payout carry different contracts, thresholds, and review obligations.
These distinctions generate practical consequences. For property underwriting, image features should be benchmarked against data already collected and should retain an inspection pathway for ambiguous conditions. For catastrophe response, apparent damage classes should support prioritization with a visible uncertainty label, rather than substitute for claim adjustment. For agriculture, satellite observations should be tied to a contract-specific basis-risk analysis. For model governance, acquisition rights and data availability should be treated as inputs to the scientific design rather than documents checked only after deployment.
A scientific research agenda follows directly from unresolved measurement and transport problems. Prospective insurer-linked studies can test incremental frequency and severity prediction across independent countries, perils, and satellite vintages. Cross-event catastrophe experiments can compare flood, wind, fire, and earthquake imagery against adjuster-verified conditions and ultimately paid claims. Agricultural research can test payout reliability against farm-level loss evidence across growing seasons and irrigation regimes. Such studies should predefine target and baseline, quantify missing observations, publish stratified performance, and distinguish the effect of greater image resolution from the effect of new geographic information.
The original authorial contribution is the integration of those established ideas into a contract-aligned scientific account of the observation-to-decision chain, a transparent risk-segmentation illustration, and explicit criteria for interpreting imaging performance in insurance. This is analytical synthesis. It neither constitutes a new named theory nor claims the discovery of empirical effects beyond the verified research.
Conclusion
Images and remote sensing expand the observable evidence available to insurance science by recording the physical structure of assets, their surrounding environment, changes across time, and manifestations of hazardous events. Their scientific contribution arises when a model relates those observations to defined insurance exposures and outcomes under an evaluation design that matches the intended decision. Street-level images may support risk classification and underwriting verification; aerial surveys can support visible asset inspection; multispectral satellite sequences can monitor land cover and agriculture; radar can observe flood-relevant surfaces under cloudy conditions; before-and-after imagery can accelerate catastrophe triage.
The decisive methodological requirements are precise. Image acquisition and availability must precede the predictive decision, geospatial matching must identify the insured asset, derived features must carry quality and provenance information, and validation must separate assets, space, time, and shared catastrophe events. Predictions require calibration and economically meaningful baselines. The apparent physical damage visible in remotely sensed products is a distinct quantity from the insurer's contractual indemnity. Rights, personal-data protection, fairness, human review, and local insurance regulation govern the permissible translation from a model score to a customer-facing decision.
The literature establishes substantive promise for selected uses and leaves important questions about geographic transfer, coverage inequality, environmental change, cost effectiveness, and individual decision reliability. A finished insurance imagery program is therefore a documented sequence of measurement, asset linkage, actuarial inference, prospective validation, governance, and audit. That sequence provides a rigorous basis for deciding which images deserve to influence risk estimates and which conclusions remain for direct inspection or further research.
References
Blier-Wong, Christopher, Luc Lamontagne, and Etienne Marceau. 2024. “A Representation-Learning Approach for Insurance Pricing with Images.” ASTIN Bulletin: The Journal of the IAA 54 (2): 280–309. https://doi.org/10.1017/asb.2024.9.
Carter, Michael R., Alain de Janvry, Elisabeth Sadoulet, and Alexandros Sarris. 2017. “Index Insurance for Developing Country Agriculture: A Reassessment.” Annual Review of Resource Economics 9: 421–438. https://www.annualreviews.org/content/journals/10.1146/annurev-resource-100516-053352.
Copernicus Data Space Ecosystem. 2026-accessed. “Sentinel-1.” European Union Copernicus programme, mission and data-collection information. https://dataspace.copernicus.eu/data-collections/copernicus-sentinel-missions/sentinel-1.
Copernicus Emergency Management Service. 2025. “Detection Methods and Damage Assessment.” On-Demand Rapid Mapping Manual, page updated November 12, 2025. https://mapping.emergency.copernicus.eu/about/rapid-mapping-manual/detection-methods-damage-assessment/.
Copernicus Emergency Management Service. 2026. “Limitations When Using Earth Observation Data.” June 29, 2026. https://mapping.emergency.copernicus.eu/news/limitations-when-using-earth-observation-data/.
Copernicus Land Monitoring Service. 2026-accessed. “CORINE Land Cover.” European Union land-cover product catalog. https://land.copernicus.eu/en/products/corine-land-cover.
European Insurance and Occupational Pensions Authority. 2025. “Opinion on Artificial Intelligence Governance and Risk Management.” August 6, 2025. https://www.eiopa.europa.eu/publications/opinion-artificial-intelligence-governance-and-risk-management_en.
European Union. 2016. Regulation (EU) 2016/679 of the European Parliament and of the Council (General Data Protection Regulation), April 27, 2016, consolidated official text. https://eur-lex.europa.eu/eli/reg/2016/679/oj/eng/.
European Union AI Act Service Desk. 2026-accessed. “Essential Services: Insurance AI Risk Assessment and Pricing.” European Commission implementation guidance concerning Annex III, point 5(c). https://ai-act-service-desk.ec.europa.eu/en/essential-services.
European Union Copernicus Sentinel-2. 2026-accessed. “MSI: Sentinel-2 Auxiliary and Data Product Specification.” Official instrument bands and resolutions. https://dataspace.copernicus.eu/data-collections/copernicus-sentinel-missions/sentinel-2.
Federal Emergency Management Agency. 2025. “National Risk Index Data.” U.S. Department of Homeland Security open-data catalog, metadata updated July 1, 2025. https://catalog.data.gov/dataset/national-risk-index-nri-data.
Gneiting, Tilmann, and Adrian E. Raftery. 2007. “Strictly Proper Scoring Rules, Prediction, and Estimation.” Journal of the American Statistical Association 102 (477): 359–378. https://doi.org/10.1198/016214506000001437.
Google Maps Platform. 2026-accessed. “Policies and Attributions for Street View Static API.” Google for Developers, contractual and product guidance. https://developers.google.com/maps/documentation/streetview/policies.
Gupta, Ritwik, Bryce Goodman, Nirav Patel, Ricky Hosfelt, Sandra Sajeev, Eric Heim, Jigar Doshi, Keane Lucas, Howie Choset, and Matthew Gaston. 2019. “Creating xBD: A Dataset for Assessing Building Damage from Satellite Imagery.” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops: 10–17. https://openaccess.thecvf.com/content_CVPRW_2019/html/cv4gc/Gupta_Creating_xBD_A_Dataset_for_Assessing_Building_Damage_from_Satellite_CVPRW_2019_paper.html.
Mason, David C., Robert Speck, Benjamin Devereux, Guy J.-P. Schumann, Jeffrey C. Neal, and Paul D. Bates. 2010. “Flood Detection in Urban Areas Using TerraSAR-X.” IEEE Transactions on Geoscience and Remote Sensing 48: 882–894. https://doi.org/10.1109/TGRS.2009.2029236.
National Association of Insurance Commissioners. 2023. “NAIC Members Approve Model Bulletin on Use of AI by Insurers.” December 4, 2023, announcement of adopted model bulletin. https://content.naic.org/article/naic-members-approve-model-bulletin-use-ai-insurers.
National Institute of Standards and Technology. 2023. “Artificial Intelligence Risk Management Framework (AI RMF 1.0).” U.S. Department of Commerce, released January 26, 2023. https://www.nist.gov/itl/ai-risk-management-framework.
Actuarial Standards Board. 2016. “Actuarial Standard of Practice No. 23: Data Quality.” Revised December 2016, United States actuarial professional standard. https://www.actuarialstandardsboard.org/asops/data-quality/.
U.S. Geological Survey. 2026-accessed. “What Are the Acquisition Schedules for the Landsat Satellites?” Landsat 8 and 9 orbit and image-acquisition documentation. https://www.usgs.gov/faqs/what-are-acquisition-schedules-landsat-satellites?items_per_page=6&page=1.
U.S. Government Accountability Office. 2026. “Disaster Risk: Improvements Needed to Enhance FEMA’s National Risk Index.” GAO-26-107894. https://files.gao.gov/reports/GAO-26-107894/index.html.
Valavi, Roozbeh, Jane Elith, José J. Lahoz-Monfort, and Gurutzeta Guillera-Arroita. 2019. “blockCV: An R Package for Generating Spatially or Environmentally Separated Folds for k-Fold Cross-Validation of Species Distribution Models.” Methods in Ecology and Evolution 10 (2): 225–232. https://doi.org/10.1111/2041-210X.13107.