Same-Day Flood Claims: The Maps Exist, the Evidence for Paying on Them Does Not

The problem

Starting point. The anchor paper covers software that reads photographs of a damaged car and proposes parts, repair operations and cost. It runs at named insurers in the United States and France, but no independent study has measured how close its estimates come to the final repair bill, and photographs cannot show structural damage hidden behind a panel¹. This paper adds a different setting: a flood that produces hundreds of thousands of claims at once, where the question is whether images taken from the air or from space, together with claimants' own photographs, could help pay people within hours.

After the floods of 26 October 2024 in the Valencia region of Spain, the Consorcio de Compensación de Seguros (CCS), the state-backed body that covers natural-catastrophe losses, registered 250,663 claims by 24 November 2025 and had paid out over €4 billion against an estimated final cost of about €4.8 billion (operator-reported, November 2025)². Its operations director calls damage assessment "the bottleneck in any natural catastrophe", and nearly 2,300 additional loss adjusters were mobilised for this one event (operator-reported, 2025)². More than €2.6 billion had been paid about four months after the flood (operator-reported, March 2025)².

Cars were the largest single category. The insurers' association UNESPA counts 144,000 vehicle claims out of about 250,000 (industry-reported, 2025)³, and CCS reports more than 140,000 vehicle claims costing close to €1.2 billion, an average of €8,600–€8,700 each (operator-reported, 2025)². For vehicles that were total losses, CCS dropped the on-site inspection and relied on representative georeferenced photographs and on photographs or video supplied by policyholders². A year after the flood, government data showed 210,106 claims paid, 246,398 handled (98.3%) and 4,276 still pending (official data, 25 November 2025)⁴. None of the sources read gives an average time to payment.

The problem is wider than one event. In the United States, the Government Accountability Office (GAO) estimated that 86% of properties it classed as high flood risk had no National Flood Insurance Program (NFIP) cover as of April 2026, and reported that flooding in 2024 alone produced about $9 billion in NFIP claims, about $7 billion of it from Hurricane Helene (GAO estimate and official data, October 2026)⁵. In the United Kingdom, the Association of British Insurers (ABI) reported that domestic flood claims rose 38% to £312 million in 2025 and that the average flood payout rose 60% to £30,000 (ABI figures reported by trade press, 2025)⁶. Speed helps only people who hold a policy, and the US figure shows how many do not.

What already works

Seeing a flood from above. Two peer-reviewed studies show that software can outline floodwater in images. A 2026 study fine-tuned a general image-segmentation model, software that outlines regions of an image, on 290 drone and helicopter flood photographs from Germany, India, Malaysia and Bangladesh. Its best version reached an accuracy of 0.96 and an intersection over union of 0.90, a score where 1.0 means the outlined area matches the labelled area exactly (measured against hand-labelled masks, 2026, abstract)⁷. A 2025 study from India's space-research community used radar-satellite data, which sees through cloud, and a digital elevation model, a map of ground height, to estimate water depth. It reported flood-extent accuracy above 90% against optical imagery and a root-mean-square error, a typical size of error, of 0.805 for its depth method against 5.23 for a comparison method; the abstract gives no units (measured, 2025, abstract)⁸.

In plain terms, the research shows that software can outline where water lies and estimate depth across flooded land. It does not show that an estimate is accurate enough to say how much water entered one particular house.

Vendor products. ICEYE, a radar-satellite company, publishes that insurers can see flood extent and depth around insured buildings within 24 hours of the flood's peak, and that payments "can be made without the need for an adjuster visiting the impacted property first, or even at all in some circumstances" (vendor-reported, June 2023). The article gives no accuracy, speed or claim-share figures and names no insurer using the approach to pay claims⁹.

Reading damage from photographs. The anchor use case shows the principle for vehicles, and a laboratory model that classified damage severity from user-uploaded images reached 73.3% overall accuracy (measured on a test set, 2025, abstract)¹⁰. That figure is modest, and it describes a laboratory model, not an insurer's system.

Paying without a visit. Valencia shows that a large insurer can set aside on-site inspection for one class of asset, total-loss vehicles, when the rules allow it².

What does not exist. This research found no published comparison of an imagery-based flood estimate against the amount finally paid. It also found no published case of property flood claims paid on imagery alone.

Seeing a flood from above and paying a claim on what is seen are separate steps. The first has research support. For the second, this research found no published evidence.

The possibility

One way this could work is as a staged payment rather than a full settlement. If a policyholder reported a flood loss, the insurer could match the insured address to a flood map built from radar-satellite or aerial images produced within a day or two of the peak. If the address lay inside the mapped flood and the estimated depth exceeded a threshold agreed in advance, a first payment of a fixed share of the likely loss could be released the same day, without a visit. Policyholders could then upload photographs of contents and vehicles, which software of the kind described in the anchor paper could triage. An adjuster could be sent only where the map and the photographs disagree, where the depth is borderline, or where the loss is large. The first payment would then be reconciled against the final assessment, so "hours" would describe the first payment, not the closing of the claim.

The role of AI. AI would be needed in two or three places: software that outlines water in aerial or satellite images⁷⁸, software that reads policyholder photographs to triage damage¹⁰, and, if used, a model that flags claims that look inconsistent with the map. Much else would be ordinary. Depth estimation combines the flood outline with a digital elevation model, which is geometry rather than machine learning⁸. Matching addresses to a flood outline, setting payment thresholds, holding the cash to pay, writing policy wording and obtaining regulatory permission are ordinary software, money and policy. If the idea were to work, AI would be a small enabling part, and the hardest steps would be contractual and evidential.

What would have to be true

1. The depth estimate would have to be reliable for a single building. The studies report accuracy for flood outlines and for water levels over flooded land⁷⁸; the abstracts read report no error figure at the level of an individual building.
2. Payment would have to tolerate error. A first payment reconciled later would address what supervisors call basis risk, "where the payout may not accurately reflect the policyholder's actual losses, potentially resulting in either overpayment or underpayment"¹¹.
3. The rules would have to allow it. FEMA's NFIP manual describes an adjuster who contacts the insured within 24 to 48 hours and visits the property, and says the insurer must hold the proof of loss and supporting documents, due within 60 days, before paying; no route to payment without an inspection was found in that document (April 2020, more than three years old)¹². Whether other countries' or private insurers' rules differ was not established, though Valencia shows a procedure that did².
4. Policyholders' photographs would have to be good enough and checkable. The accuracy of photo-based estimates in the field has no independent measurement¹¹⁰.
5. Fraud and manipulation would have to be controlled. ICEYE says claims can be flagged when they fall outside the mapped flood or look excessive for the mapped depth (vendor-reported, 2023)⁹. The same map that triggers a payment would also become a target for false claims.
6. Policyholders would have to understand and be able to contest a payment. The FSI and IAIS paper on parametric insurance notes "educational challenges for both consumers and insurers" (December 2024)¹¹.

What could stop it

What the evidence shows

• Valencia was not simple even with relaxed inspection. CCS still needed nearly 2,300 additional loss adjusters². UNESPA states: "There were problems, and many decisions had to be taken in a hurry, but the system worked"³. A year later 4,276 claims, 1.7% of those presented, were still pending⁴.
• Regulators have restated the right to a visit. The New Hampshire Insurance Department told insurers on 14 April 2025 that photographs and videos do not always show the full extent of damage, and that they must provide an in-person inspection within a reasonable time if the consumer asks, even after the consumer first chose a virtual process¹³.
• The technical evidence is thin where it matters. The imagery studies measure outlines and water levels over land, not losses inside buildings, and ICEYE's article gives no figures⁷⁸⁹.
• Most flood losses are uninsured. An estimated 86% of high-risk US properties had no NFIP cover in April 2026⁵.
• Surge handling in the UK relies on people. A July 2026 trade report on UK loss adjusting describes insurers combining field and desk adjusting and adding back-office capacity during surges, and mentions no imagery tools⁶.

Reasoning

The challenge question is what would make this less likely, less useful or more harmful than it first appears. Three answers stand out. First, the speed-up in the best-documented case may owe little to imagery. The sources on Valencia attribute the faster handling to relaxed inspection rules, a purpose-built claims platform and many more adjusters, and none of them mentions satellite or AI imagery²³. If that holds more widely, a simpler arrangement could capture most of the benefit: a flat emergency payment to every policyholder inside an officially mapped flood zone, adjusted later, which needs no machine learning.

Second, errors would not fall evenly. An overpayment is a cost to the insurer, which can absorb it. An underpayment, accepted because it arrived quickly, would fall on a policyholder who cannot see how the figure was set. If an imagery-based estimate were used to settle rather than to advance, the insurer's incentive to pay less could meet a customer with little means to contest it.

Third, the people most in need of fast payment may be those least able to benefit: those without cover⁵, and those whose homes sit in places or conditions, such as shallow or fast-moving water, that the studies do not test. No public price for imagery-based flood data was found, so who would pay, and whether the cost is below the saving on adjuster visits, is unknown.

Signs to watch

• A published evaluation, by an insurer, regulator or university, that compares imagery-based flood estimates with final paid amounts on the same claims.
• A post-event review of the Valencia handling, by CCS, UNESPA or the Spanish supervisor, that reports days to payment and says whether imagery played any role; the CCS journal Consorseguros Digital has been the channel so far²³.
• A revision of the NFIP claims manual that allows payment before or without an adjuster inspection¹².
• Further US state insurance bulletins following New Hampshire's April 2025 position on in-person inspection¹³.
• An insurer naming a flood event and the number of claims paid on imagery, which the vendor article does not do⁹.
• Supervisory guidance requiring insurers to disclose how an imagery-based estimate or trigger was calculated¹¹.