Spotting the Roof That Changed, Not the Shadow Teaching a Model to Ignore Common Changes

This application covers a way of telling when something real has happened to a property by comparing two aerial or satellite views of it, without being fooled by shadows, seasons, parked cars or camera angle. A machine learning model is trained to turn each view into a compact representation that stays the same through those everyday differences, so that a gap between two representations flags a rare change such as damage or construction. It was filed by Cape Analytics, Inc., names Matthieu Portail, Christopher Wegg and Fabian Richter as inventors, and is a divisional of Australian application 2022409205.

The Problem

Anyone who has compared two photos of the same house knows the trouble. Between the two images the lighting differs, trees have grown leaves, a trampoline has appeared, a cloud has crossed the sun, and the aircraft was a few degrees off from last time. A naive comparison lights up on all of this. The real events an insurer or lender cares about, a house fire, a hailstorm, a new extension, are a tiny fraction of what differs between any two images.

The specification states the central difficulty directly. A comprehensive list of every common change is hard to build, because there are too many of them, and a model trained to recognise each one separately would be incomplete. It also notes that rare changes occur infrequently and that labelled examples of them are scarce or missing. That makes a conventional supervised classifier, trained on examples of damage, hard to build for every kind of rare change.

The document adds a third point. Geometric features pulled from remote imagery by appearance based methods can be highly susceptible to common changes, particularly when the two measurements come from different sources, for example aerial against satellite imagery of the same property, where georegistration errors, atmospheric effects and misalignment create apparent differences even when the images were taken at the same moment.

What This Invention Does

The independent claim 1 recites a method with four steps. Determine a first and a second measurement of a geographic region. Determine a first representation of the region from the first measurement using a representation model, where that model is trained to be agnostic to common changes. Determine a second representation using the same model. Then detect a rare change for the region based on the two representations. Claim 10 is a second independent claim with a more specific recipe: the representation model is trained by self supervised learning on a training set of unlabelled measurements of many geographic regions, a comparison metric is computed from the two representations, and a rare change is detected from that metric.

In plain terms the model is trained so that different pictures of the same place produce nearly the same vector. Dependent claim 2 says exactly this: the model outputs substantially equivalent training representations from different measurements of a training region. Dependent claims 3 and 13 then detect change from the distance between the two vectors being above a threshold. The description gives cosine similarity as one example comparison metric.

The training set does not need to be clean. Claim 12 says the share of unlabelled measurements that depict rare changes is below a threshold. The spec describes rare changes as those occurring in fewer than some threshold proportion of measurements, giving 30 per cent as an example while allowing both lower and higher cut offs.

Beyond detection, claim 7 and claim 15 add an optional classifier that labels the rare change. Claims 8 and 9 describe a route where a text encoder turns a description of a change type into a representation, and a likelihood is computed that the measurement pair shows that change, which is described as using CLIP style methods. Claim 18 names property damage and property construction as examples of rare change. Claims 19 and 20 name the parcel and the census block group as geographic regions.

Key Features

• A representation trained to be blind to common changes. The model learns that shadows, seasons, vehicles, temporary structures and camera pose should not move the output, so what remains is the signal of real change.
• Self supervised training on unlabelled imagery. Claims 4 and 10 train the encoder without labels. The description lists methods such as bootstrap your own latent (BYOL) among those that can be used.
• Comparison by distance and threshold. Two representations are compared with a metric such as cosine similarity, and a rare change is flagged when the gap passes a threshold.
• Storing representations rather than images. The description says keeping the vectors instead of the measurements reduces the storage needed for a time series of a property.
• Filtering before classifying. A large dataset of regions can first be filtered for those with a rare change, and the heavier classification step is run only on that subset.
• Uses named in the description. Listed applications include automated valuation, real estate management, loan trading and insurance underwriting, for example choosing where to send inspectors or finding properties damaged since a prior due diligence check.

Who Is Behind It

Cape Analytics is a geospatial artificial intelligence company based in Mountain View, California. TechCrunch reported that it was founded in 2014 by Suat Gedikli and Ryan Kottenstette, extracts structured property data such as roof condition and solar panels from satellite imagery, and sells it for insurance pricing and underwriting. The same report records that Moody’s announced on 13 January 2025 an agreement to acquire the company, which Moody’s own announcement describes as folding its address level analytics into the Moody’s Intelligent Risk Platform. The price was undisclosed; closing was expected in the first quarter of 2025. This application still lists Cape Analytics, Inc. as applicant, and the specification does not mention the acquisition.

On priority, the specification opens with a cross reference to two US provisional applications, number 63/290,174 filed 16 December 2021 and number 63/350,124 filed 8 June 2022, and its page headers show it is the Australian entry of PCT/US2022/051719 (WO 2023/114027). The priority country is therefore the United States. This Australian application is a divisional of 2022409205, and the title page lists Pipers Intellectual Property of Auckland as agent.

Why It Matters

Property data goes stale. A valuation model built on last year’s imagery misses the new pool, the missing roof section or the demolished shed, and an insurer pricing a policy or a lender holding collateral wants to know about those events without sending a person. Doing that across millions of addresses only works if the system rarely raises a false alarm, since each alert costs an inspection or a manual review.

The approach in this claim moves the hard work into training rather than rule writing. Instead of listing every harmless difference, the model is shown many views of the same place and learns what stays constant. That is a sensible bet where labelled damage examples are scarce, and the specification says as much. What the document does not provide is measured performance. It gives no detection rates, false alarm rates or benchmark results, so how well the representation holds up against real hail damage or seasonal change is not shown in the text.

Related Concepts

• Self-supervised learning – the training approach that lets the representation model learn from unlabelled imagery.
• Feature learning – the general field of learning compact representations from raw data, which the vectors here are an example of.
• Change detection – the GIS task of finding differences between images of the same area at different times.
• Cosine similarity – the example comparison metric between a baseline and a test vector.
• Underwriting – the insurance and lending decision that the detected changes are meant to inform.

---

AU 2026202257 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.