Scoring a House for Wildfire From the Air How an Aerial Photo Becomes a Fuel Load

This application covers a computer method that takes an image of a property, uses machine learning to pick out the house and everything around it that could burn, and turns the lot into a single figure for how much thermal energy the site could throw at the building. It was filed by Fortress Wildfire Insurance Group, names John Wall and Michael O’Dell as inventors, and is a divisional of Australian application 2022205161.

The Problem

The specification opens with a plain statement of the risk: in high-risk areas a wildfire can mean the total loss of a house. What it complains about is the tooling. Existing approaches to assessing that risk are described as lacking precision and actionable data, and as hard to scale to the size of the problem. Some are generic guidelines, such as telling an owner to remove flammable material without saying which material or how much it matters. Others need an expert to walk the property, which makes the result depend on the expert and leaves out any quantitative estimate of ignition risk from the surrounding fuel and the building’s own materials.

The specification also notes that decades of wildfire research have produced models of how fires spread, such as BeHave, FlamMap and FARSITE, but that very few models describe how a home ignites at the wildland-urban interface. Of those it says none scores a particular home in a form that an owner or insurer can act on.

What This Invention Does

The title promises assessment, mitigation and monitoring. Claim 1 covers the assessment half. It is a method in which a system receives at least one image of a property, identifies the structural features of the main building, finds the other things that could fuel a fire, and ends by calculating a fuel load in terms of thermal energy generation potential. Mitigation shows up only in a later dependent claim, and monitoring is not in the claims at all, so the article follows claim 1 rather than the title.

Claim 1 works in four steps. First, the system builds an inventory of the building’s exterior features along its perimeter, with their dimensions and material composition. Second, a machine learning feature detection algorithm reads the image to characterise the features needed to model fire susceptibility. Third, a second machine learning algorithm finds the fuel around the building: major vegetation such as a tree or shrub, secondary structures such as a shed or fence, neighbouring rooftops close enough to help a fire advance, and the building footprint, roof and local terrain, where terrain includes a slope, a road, a hydrant or an arroyo, which is a dry creek bed. Fourth, it assigns a feature-specific attribute to every structural feature and every fuel source and calculates the fuel load from those attributes.

The dependent claims carry the maths. Claim 2 lays a grid over the image with the building’s centroid at its centre and encodes each tile with fuel, structure and terrain detail. Claim 3 adds a fire risk algorithm with four threat vectors: radiant heat flux, direct flame contact, ember accumulation and ember penetration. Claim 5 tests wind from at least the four compass directions at a range of speeds. Claim 6 compiles every feature that ignites into a failure mode and effects analysis, ranks them by a risk priority number and produces a score for the whole property. Claim 8 turns the ranking into a remediation report, naming hardening the building and reducing the surrounding fuel as the fixes. Claim 11 recasts the method as a computing apparatus with a processor and memory.

The body of the specification supplies the working assumptions. Everything within 20 metres of the home’s centroid is assumed to burn. Trees are assumed to release between 0.08 and 0.15 per cent of their mass as embers. Calculations run at about 50 cm resolution in two dimensions. Scores are fitted, using a sample of more than 10,000 properties, to a scale of 1 to 100, where zero means no ignition at peak. The reference “peak burn” has all trees burning in a one minute heat pulse with 20 mile per hour winds and 30 per cent humidity. The specification gives a simple example of a fix: replace wood siding with cement board.

Key Features

• Image in, score out. The only input claim 1 requires is at least one image of the property, which claim 9 says can be oblique satellite or aerial imagery, ground photos, real estate listing databases or phone app images.
• Two machine learning passes. One algorithm characterises the building and a separate one finds the fuel around it, including trees, shrubs, sheds, fences and neighbouring roofs.
• Fuel load as thermal energy. The end product of claim 1 is a fuel load expressed as thermal energy generation potential, not a categorical high, medium or low rating.
• Four ignition pathways. Radiant heat, direct flame contact, ember accumulation and ember penetration are each assessed per tile against the material failure threshold of the building feature behind it.
• Failure catalogue that feeds back. Claim 4 records, for each failed tile, its location, the failing feature, the threat vector, the contributing fuel sources, the wind direction and the amount by which the heat exceeded the failure point, and uses those records to update the algorithm.
• Ranked mitigation. The FMEA ranking in claims 6 to 8 orders the fixes by risk priority, so an owner can see which change removes the most risk first.

Who Is Behind It

The applicant is Fortress Wildfire Insurance Group. The trading business is FortressFire, whose own footer describes it as a division of Fortress Wildfire Insurance Group. It sells a wildfire platform called AMP, which uses satellite imagery, machine learning and physics models to score a property against five wildfire threat vectors and propose mitigations, with customers that include insurance carriers, real estate professionals, property owners, community associations and utilities. Its site focuses on California and the western United States.

A trade publication, Fire and Safety Journal Americas, names Michael O’Dell as FortressFire’s Head of Machine Learning and Data Science, which fits his place as an inventor. It does not mention John Wall, and the article could not find a verified public profile for him.

On priority, the specification says it claims the benefit of United States provisional application 63/221,242, filed 12 July 2021, and is a divisional of Australian application 2022205161. The priority country is therefore the United States. The title page gives no applicant address, so the applicant country in the table below comes from the company’s public focus on the United States and is not stated in the document.

Why It Matters

Insurers in wildfire country are deciding house by house whether they will write cover at all, and a loss model that stops at “this postcode is high risk” gives them no reason to treat a house with a ceramic roof and cleared yard differently from its neighbour. This claim is a way of making the difference visible and countable, from imagery alone and without a site visit.

For homeowners, the useful part is the ranking. The FMEA step produces a ranked list of what failed and what fuel caused the failure, which turns a risk score into a to-do list. For an insurer it provides a repeatable, auditable score, which is a different thing from an expert’s opinion.

What the specification does not do is publish validation results. It states the scale and the sample size it was fitted to, but the text shows no comparison of its predictions against houses that actually burned, so how well the scores track real losses cannot be judged from this document. The model also makes simplifying assumptions that the specification lists openly, such as treating the whole house as lost if any single feature ignites.

Related Concepts

• Wildfire – the hazard the whole method is built to assess.
• Wildland-urban interface – the zone where homes meet flammable vegetation, which is where the model is meant to be used.
• Heat flux – the radiant energy measure behind one of the four threat vectors.
• Failure mode and effects analysis – the risk ranking framework used in the dependent claims.
• Defensible space – the cleared buffer around a building that the remediation step aims to create.
• Computer vision – the field that covers reading structures and vegetation from aerial images.

---

AU 2026202280 was published in the Australian Official Journal of Patents on 9 April 2026 and is open for public inspection. Patent applications represent inventions that are sought to be protected and do not necessarily reflect commercially available products.