Inside Swiss Re's Drone Flood Patent: A Vulnerability Curve for Every Building

A drone flies a user-drawn polygon, photographs every point inside it from at least two positions, and the system triangulates a bare-earth terrain model with a ground sampling distance of 5 cm or less. Ground control points hold the root mean square error to 15 to 40 mm. That measurement chain is the independent claim of US Patent 12,705,882 B2, issued to Swiss Reinsurance Company Ltd on August 11, 2026 (USPTO, August 2026).

The actuarial payload sits at the far end of that chain. Claim 18 asks for separate vulnerability curves, one per building and one per sub-unit inside the polygon, each indexed to inundation height or return period. Flood ratemaking has run for fifty years on zone assignment and, more recently, on property-level proxies. This claims the sub-building level, on terrain the reinsurer measured itself.

Executive Summary

Swiss Re filed the priority application in Switzerland on October 5, 2022, carried it through PCT/EP2023/077532 a year later, and entered the United States on May 6, 2024 as application 18/655,995. The USPTO issued 18 claims on August 11, 2026. The single independent claim is a system claim, reciting drones, sensors, a calibration module and a filter, not a bare method.

That drafting choice is doing work. Insurance analytics patents have been the softest target under Section 101 since Alice, and the Federal Circuit's 2025 Recentive Analytics decision narrowed the space further for machine-learning claims. A claim that recites a measured error tolerance of 15 to 40 mm at physical ground control points is a materially harder abstract-idea rejection to write.

The output layer is where the pricing content lives. A flood hazard aggregator generates vulnerability measures for each unit and sub-unit inside the polygon, keyed to forecasted or user-selected return periods and inundation heights, and the specification runs those through to monetary property damage and business interruption estimates. Each sub-unit carries its own damage curve, with a dependency scheme for cascading failures.

Claim 7 supplies the parametric hook: the aggregator determines damage by matching structures against a digital data warehouse of historically measured flood damage at corresponding inundation heights and return periods. That is a payout basis. It relocates residual basis risk from zone assignment to depth measurement, which is a different and smaller error term.

Patent Details

The Mechanism

From Overlapping Photographs to a Bare-Earth Model

The operator draws a polygon on a topographic map through a geo-tagging interface. Drones then image the enclosed area so that every possible location is captured in at least two photographs taken from different positions, with an image overlap the specification puts at 70 percent or more. Common points are identified across image pairs, a ray is generated from the camera to each measured ground point, and the intersection of those rays fixes a three-dimensional coordinate.

Georeferencing comes from ground control points, physical marks on the site tied to surveyed coordinates. The claim language sets the tolerance directly: "to achieve an accuracy of 5 cm or less for elevation and surface inclination drone-based measurements, root mean square error values at the at least two ground control points is in a range of 15-40 mm" (US 12,705,882 B2, August 2026). Dependent claims add post-processing and real-time kinematic GPS correction, and claim 14 escalates the number of control points until the tolerance is met.

The last step is a progressive morphological filter, tuned by parameter calibration, that strips vegetation and structures out of the surface model to leave a digital terrain model. That distinction between the digital surface model and the terrain model is not cosmetic. Water routes over bare earth, and a canopy-contaminated surface model will put flow paths in the wrong place.

Where the Hydrology Enters

The specification is explicit that prior-art flood mapping leans on one-dimensional unsteady-flow tools such as HEC-RAS and HEC-HMS coupled to GIS, and that the industry has generally accepted coarse inputs in exchange for tractable computation. The forecast simulation module here is distributed rather than sectional: modelling parameters are assigned per cell and classified into four families, climate-related, topography-related, vegetation and land-use, and soil, with cells sharing terrain properties sharing parameter values.

That parameter-sharing scheme is the practical answer to the obvious objection. A 5 cm grid over an industrial site produces an enormous number of cells, and calibrating a free parameter per cell is not feasible. Tying parameters to terrain class collapses the estimation problem to a manageable number of independent parameters while preserving the fine geometry that determines where water actually goes.

Output is a set of pre-generated inundation extent layers at defined return periods, with the specification naming 10, 20, 30, 40, 50, 75 and 100 years, plus a custom return period entry. Claim 4 adds selectable earth warming scenarios, so the same terrain model can be re-run under alternative climate assumptions without a new flight.

The Vulnerability Curve, One Per Sub-Unit

Claim 15 lets the operator segregate the site into functional components and assign functions, interactions and dependencies between them. Claim 18 then gives each of those components its own vulnerability curve against inundation height or return period. A slider moves the water level and the system returns property damage and business interruption loss measures in monetary terms.

The specification's worked example makes the granularity concrete. At a flood high water mark of 1.1 m, roughly 3.6 ft, no buildings on the illustrated site are impacted. Raise the mark to 2.7 m, about 9 ft, and the boiler room, an LPG tank farm, the engineering block, workshop, link warehouse and production building are all impacted. A protection measure fitted to the boiler room lifts that unit's own threshold without moving anyone else's.

Two consequences follow. First, the site loss function against depth is a step function with as many risers as there are protected thresholds, not the smooth curve a zonal average implies. Second, the dependency scheme means a low-value unit can drive a high-value loss: the specification notes that a power supply breakdown can halt an entire plant while a storage unit failure may not.

Actuarial Implications

Dispersion, Not the Mean

Replacing a zonal average with per-structure curves does relatively little to the expected loss at a well-modelled site. It does a great deal to the variance. A zonal factor implicitly assumes every exposure in the zone shares one depth-damage relationship, which smooths the aggregate severity distribution toward the average building. Resolving thresholds unit by unit restores the discontinuities that the smoothing removed.

For a primary writer that mostly changes the risk load. For a reinsurer pricing an excess layer over a single industrial site, it changes the attachment analysis directly, because the probability mass between two adjacent protection thresholds is exactly the material that decides whether a layer is exposed at a given return period. A step function with a riser just above the attachment point prices very differently from a smooth curve through the same mean.

Parametric Triggers and Where the Basis Risk Moves

Parametric flood cover has historically struggled because the natural index, a gauge reading or a declared flood extent, correlates loosely with what happened at a specific insured building. Claim 7's damage warehouse is an attempt to shorten that gap by matching the insured structure to comparable historical structures at comparable depths, so the trigger pays against a modelled site loss rather than a regional index.

The residual basis risk does not vanish. It changes character. Under a zonal trigger the dominant error is classification, whether the building belonged in the zone at all. Under this construction the dominant error is measurement, whether the depth at that structure on that day matched the depth the index assigned. The patent measures terrain to 5 cm; it does not measure the water. Depth still arrives from a forecast module, a gauge, or post-event imagery, and that is where the remaining uncertainty concentrates.

That is a favorable trade for a cedent, because measurement error is estimable from the instrument and largely symmetric, while classification error is neither. It is also the reason the specification frames the system as providing "inputs for parametric risk-transfer structures" rather than as a rating engine in its own right.

Build Versus Buy for Primary Flood Writers

Private flood remains a small and concentrated line. Private insurers held about 7.1 percent of the US flood market in 2022, up from 3.6 percent in 2018, with more than 140 carriers offering some form of flood coverage and total flood direct premiums of roughly $4.09 billion (Triple-I, August 2023). A book that size does not fund a proprietary photogrammetry and hydrology stack.

So the build-versus-buy question resolves toward buy, and the counterparty holding the patent is also the counterparty selling the reinsurance. Swiss Re has been signalling the underlying economics for two years: global insured flood losses were $3.4 billion in 2025 against $107 billion of total insured natural catastrophe losses and $220 billion of economic losses, and insured flood losses across Asia have been growing at an estimated 12 percent per year against a regional flood protection gap of 83 percent (Swiss Re Institute, sigma 1/2026, March 2026).

The gap figure is the commercial argument. Uninsured natural catastrophe losses reached $424 billion in 2025, up 7 percent, with $140 billion of that in North America (Swiss Re Institute, June 2026). Closing any part of a flood gap requires pricing risks nobody has priced before, and pricing them requires exactly the structure-level data this patent describes acquiring.

Against Risk Rating 2.0 and First Street

The comparison the trade coverage skipped is the one that matters for scope. Two systems already rate US flood below the zone. Neither operates at the resolution or the unit of analysis this patent claims, and neither is trying to.

Risk Rating 2.0 already replaced binary zone rating with per-property variables including distance to a water source, flood type and frequency, ground elevation and rebuild cost (FEMA, accessed August 2026). It is a genuine improvement and it remains constrained by what the program can observe remotely across 4.8 million policies. The affordability machinery blunts it further: the GAO put the median NFIP premium at $689 against a $1,288 full-risk target, a gap the statutory 18 percent annual cap will take a decade to close, as covered in our analysis of the reauthorization cliff.

First Street's national model attacked the mapping gap from the other direction, identifying 14.6 million US properties at substantial flood risk, about 1.7 times the FEMA special flood hazard area count, of which 5.9 million sit outside the SFHA and are therefore unaware of it (First Street Foundation, June 2020). That is breadth without depth: excellent for portfolio triage and disclosure, insufficient for deciding whether a boiler room floods at 1.6 m or 2.25 m.

Which is the real boundary on the Swiss Re system. A drone survey with surveyed ground control is a per-site cost, not a per-policy cost. It will never rate a homeowners book. It fits large commercial and industrial single-location risks where business interruption dominates property damage, where the insured will pay for the survey, and where a parametric wrapper is already commercially plausible. Reading this as a personal-lines flood story misreads the economics of the flight.

What to Check Before Relying on Output Like This

1. Was the vulnerability curve calibrated on structures genuinely comparable to the insured one, or on a warehouse match that shares occupancy code and little else? Claim 7 leaves "corresponding or closely similar" undefined.
2. Does the depth estimate at settlement come from the same distributed model that priced the risk, and if so, what is the correlation between pricing error and settlement error?
3. Are protection measures in the digital twin current? A flood barrier modelled at survey date and absent at loss date is a silent overstatement of the threshold.
4. How is the business interruption dependency graph validated? The specification says it can be built "through feedback with site personnel", which is an expert-judgment input carrying no stated error bound.
5. Does the terrain model predate any site earthworks, paving, or drainage change? The 5 cm tolerance is meaningless against a two-year-old survey of a site that has since regraded.

Balz Grollimund, Swiss Re's head of Catastrophe Perils, framed the loss environment this patent is built for plainly: "If losses return to normal long-term levels, they would total US$148 billion in 2026" (Insurance Journal, March 2026), with modelled peak scenarios reaching about $320 billion. Against that backdrop the interesting fact about US 12,705,882 is not the drone. It is that a reinsurer now holds the intellectual property covering how a flood loss at a specific building gets measured, and the primary carriers who need that measurement will be renting it from the party on the other side of the treaty.