State Farm Patents a GAN That Invents the Roof Geometry a Drone Scan Missed

A 3D point cloud built from a drone pass over a roof is almost never complete: obstructed sightlines, reflective shingles and overhanging trees leave gaps a human reviewer would flag as missing data. US Patent 12,705,712, granted to State Farm on August 11, 2026, trains a generative adversarial network to fill those gaps by inventing the missing geometry and labeling roof, window and wall elements in the same pass, then routes the reconstructed surface into both underwriting and claims decisions.

Key Takeaways

• Two operations in one forward pass: the network fills gaps and assigns semantic labels simultaneously, where a conventional pipeline would inpaint first and segment second. That combination is what the independent claim tries to own.
• Claim 5 names roof composition, roof type, window count and story count as extracted characteristics. Those are rating variables, which makes the imputation step an input pipeline for pricing rather than a visualization tool.
• Claim 8 diffs a fresh reconstruction against the customer's prior scan to detect damage, potentially before a homeowner files a claim or an adjuster is dispatched.
• 326 granted AI patents since 2014 put State Farm ahead of USAA's 218 and Allstate's 136, with the three carriers holding 77% of all insurer AI patent filings.

What Claim 1 Actually Describes

The independent method claim is specific about sequence: obtain a 3D point cloud of a structure, augment it with data describing a structural element's characteristics, then generate "a gap-filled semantically segmented 3D point cloud from the 3D point cloud as augmented with the data using the trained generative adversarial network, wherein the trained generative adversarial network fills one or more gaps" (USPTO grant, US 12,705,712 B2).

Two operations happen in one forward pass. The network reconstructs the missing surface and simultaneously assigns a semantic label, roof, window or wall, to every point in that reconstruction. A conventional pipeline runs these separately: patch the point cloud with an interpolation or inpainting method, then segment the patched result.

Dependent claims spell out what the segmentation is meant to extract. Claim 5 covers identifying "a number of windows of the structure, a size of the portion of the structure, a number of stories of the structure, a roof composition of the structure, or a roof type of the structure," and claim 6 extends the same logic to appliances inside the structure. Those are the structural characteristics a homeowners underwriting model uses to set eligibility and rate a policy.

Claim 8 crosses from feature extraction into loss adjustment, describing "comparing the gap-filled semantically segmented 3D point cloud against the historical customer data to detect damage to the structure." A fresh scan gets diffed against the customer's prior scan on file, and a shifted roofline or a missing section of shingles becomes a candidate damage flag.

Imputed Geometry Becomes a Rating Input

Underwriting and reserving models have always run on assumptions, but assumptions conventionally get applied to observed data: a rating factor takes a known roof age and known territory and produces an expected loss cost. This patent changes what counts as the observed input. When the network fills a gap, the roof composition that eventually reaches a rating engine did not come from a measurement of that structure. It came from a statistical prediction conditioned on the visible portion of the scan and on patterns learned from other buildings. The rating input is synthetic before the rating model ever sees it.

That is a different category of model risk than the one pricing actuaries are set up to manage. A rating factor can be audited back to an observed variable and a documented relationship. An imputed structural characteristic needs a second layer of audit: not only whether the rating relationship is correct, but how much of this specific roof was actually scanned.

Nothing in the claim language requires the output to mark which points were observed and which were generated; the claims describe one unified reconstruction, handed to underwriting and claims as though it were a single consistent measurement. A rate filing that documents the roof-composition variable without validating the imputation behind it has audited the wrong layer of the pipeline.

The grant is the sixth branch of one family. Priority runs back to Provisional Application No. 62/967,315, filed January 29, 2020, through continuations filed September 24, 2020 and November 7, 2022, before splitting into applications that apply the identical method to different business domains (USPTO Patent Public Search).

Patent	Application domain
US 11,508,042 / 11,995,805	Imputation of 3D data using GANs (base filing)
US 12,039,706	Facilities management and operations
US 12,051,179	Vehicles and transportation
US 12,056,859	Peril modeling
US 12,243,199	Construction and urban planning
US 11,983,851 / 12,705,712	Underwriting, claim handling and retail operations

Filing the same core method against separate domains forces a competitor to design around one invention repeatedly rather than once. It pairs with US Patent 12,694,432, granted July 28, 2026, covering a pricing model that detects its own miss and deploys the fix (State Farm patent portfolio). Evident's tracker counts 326 granted AI patents for State Farm since 2014 against USAA's 218 and Allstate's 136 (Evident Insights).

The Provenance Gap the Governance Layer Does Not Cover

As of the second quarter of 2026, 25 states and the District of Columbia had adopted the NAIC's AI Model Bulletin in full or substantially similar form (adoption tracker). Its documentation expectations, covering training data provenance, validation testing and disparate-impact review, were written for models that transform observed inputs into a score. A network that generates the input itself sits a step upstream of that scope.

The vendor comparison shows what changes. EagleView, Cape Analytics and Zesty.ai sell structure-level attributes pulled from aerial or satellite capture, with Zesty.ai alone reporting coverage of nearly one million previously uninsurable properties across roughly 200 regulatory approvals (Phidea, 2026). A carrier buying observed imagery can point to the underlying pixels when a decision is challenged.

A Santa Ana homeowner told in early 2026 that her policy would lapse without roughly $20,000 of roof repairs, identified by an unannounced aerial assessment, had that demand reversed on exactly those terms (ABC7 Los Angeles). That case involved observed imagery, not generated geometry.

There is a reserving consequence too. A carrier running claim 8's comparison on a periodic rescan cycle, rather than only after a first notice of loss, opens a claim-reporting channel that does not originate with the insured. Structural deltas surface on the rescan cadence rather than when the loss occurred, which decouples report date from occurrence date. A property triangle that mixes insured-initiated claims with model-flagged ones carries two reporting patterns, and a single tail factor fitted to the historical mix no longer describes both.