Computer Vision Comes of Age

www.
By : Emmett Soldati |February 13, 2017 |weather-report |0 Comment

Computer vision is all the rage in tech.  Relying on algorithms to review and make judgment calls on a collection of images presents a host of new applications in the consumer and business-to-business market. Combining imagery analytics with weather and atmospheric analytics, computer-vision will soon be reality at forward-leaning insurance companies.

Property and casualty insurance firms have an increased need and appetite for this kind of innovation. While they’ve built massive databases, full of extensive property-feature information, much of this data collection has happened over extended periods of time – through a sea change of technological advances, new demographics, and portfolio acquisitions.  All of this change stands against an evolving risk landscape of natural catastrophes, convective storms, wildfires, and other environmental perils.  The ‘known-knowns’ of typical property portfolio exposures are shifting.  But upgrading these books of business is no small task.

Future-Proofing Your Portfolio

Changes to the risk landscape – especially weather and environmental threats – present new underwriting challenges that require new data collection.  Was the underwriter aware that the region was prone to wildfire spread?  Did the coastal property business anticipate private flood insurance when they sent surveyors over a decade ago?  Understanding the present and future risks puts underwriters in the right place asking the right questions today, to future-proof their data.  However, even when underwriters are asking the right questions, and collecting the right information, carriers do not always have clean, accurate, accessible, and machine-readable data.

Enter imagery analytics – the deus ex machina for this actuarial coming-of-technological-age tale.  The predecessor for this field of analytics was Keyhole Inc., an Earth imagery company with strong ties to the U.S. Intelligence Community.  Google purchased Keyhole in 2004, renamed it “Google Earth” and began adding significant new functionality and content.  But it would be almost a decade before drivers in insurance, financial markets, energy, and agriculture would transform the consumer novelty of Google Earth’s imagery insights into commercial table-stakes.

The key was training computers to recognize objects in images and automatically detect changes in those objects, through machine learning techniques.  Bringing this pattern-recognition protocol to the visual domain relies on a burgeoning data science field known as neural networks.  Many people see the impacts of neural networks through facial recognition software, but don’t really understand the science behind the powerful ability for computers to detect seemingly complex images or faces.

The premise is fairly simple – train a model to read pixel layers, and the proximity of certain pixels to other pixels, to recognize patterns and categorize similar features. Then store these features in computer memory and search and tag similar features in new sets of images. This capability allows Snapchat to recognize where your nose is, to swap it with a cartoon dog’s nose, or helps your search engine find thousands of cat images from a Google image search. When applied to insured houses and businesses, it can transform a sea of American homes and roofs into underwriting intelligence.

Unlike faces, or cats, the availability of useable roofing and property imagery is surprisingly limited.  The sky-race to get the right equipment in the air plays out between cube-sat, drone, and aerial fly-over companies, while some other organizations opt to analyze public and open source data already collected. The test of these technologies comes down to speed of deployment, geographic coverage, and image resolution.  The three leading imagery-capture technologies – satellites, drones, and airplanes – each has its strengths and weaknesses.

As Google’s TerraBella (originally SkyBox) and other low-Earth orbiting satellite companies push for optically advanced satellites demonstrate, image resolution matters.  Most of these companies boast a pixel resolution of 30-90cm.  Being able to derive meaningful (and confident) property attributes from this level of resolution is nearly impossible.  Trying to detect roof damage on a single family home from Landsat or other satellite images would be akin to identifying a smile on the face of a blurred-out ‘Perp’ on the hit 90s show COPS. Overtime, the push for these low-earth-orbiting satellites to improve their optical vision as they travel closer to the surface will be the game-changer.

Drones offer the opposite value.  Quick and close fly-overs allow for stunningly high-resolution images of rooftops (2-5cm).  But the hardware is limited in its ability to cover ground.  Most drone imagery companies have focused their deployments in a handful of dense population centers.  Though perhaps beneficial for activities like catastrophic response and claims handling, this does little to assist with the more massive portfolio upgrades of major property carriers.

Aerial fly overs, so far, represent the goldilocks approach of the three.  Planes can get close enough to collect usable and machine-visible imagery, while the speed and frequency of trips allows for much broader coverage.

Equipped with high-resolution and widespread imagery, neural networks analytics can transform these images into actionable data at a property level.  This transformation into “roof type” or “vegetation encroachment” takes some time to train the models – but once the out-of-sample tests deliver a high confidence value, then you’ve got yourself a new portfolio database.

This data, however, is not an end in itself. Property features, absent risk, are glorified tax maps.  Fusing property features with atmospheric perils, analyzed over time, compared against real-life claims and loss information, is a recipe for the underwriter’s true dream: insight.

Like our name suggests, Weather Analytics is, at its core, a data analytics company.  We deploy data science expertise to understand and predict new risks from a climatological and atmospheric perspective – including risk scoring for under-modeled perils like hail and winter storms, to building and broadcasting the world’s most advanced machine-learning hurricane forecast.   Weather has always been our main subject.  As our company has expanded over the last several years, however, understanding the object of interest – the domain that the weather risks apply to – be them properties, infrastructure, crops, or even energy grids – has pushed the company into less chartered territory. Deriving the property characteristics for our customers is the first hurdle, but as environmental risk experts, the important question to follow is, “so what?”

Ice Dam Case Study

For some primary insurance carriers, the “so what?” is asking how they can transform their property portfolio into a hierarchy of exposure. Based on years of rigorous underwriting, one New England insurance client of Weather Analytics already has substantial machine-readable property data on hand – from year built to architectural style to heating unit.  But how to make sense of it all? In a recent work program, Weather Analytics transformed this grab bag of property features into acute underwriting logic around one of New England’s most maddening winter phenomena – ice dams.

Analysis of over 1,000 policies and 500 claims related to winter roof damage from the 2015 winter season revealed distinct correlations to property features and risk of ice dam.  Running a feature-selection model similar to Weather Analytics’ crop yield forecast, key performance indicators were narrowed down to three main factors contributing to ice dams.  First was roof type (mansard, gabled, flat, etc.), followed by house heating system, and roof material.  Using high-resolution surface and atmospheric weather data, an ice dam model is based on the specific weather events that contribute to roof damage.  Ice dams are complicated beasts that require several atmospheric conditions to fall into place over time.  Weather Analytics’ multi-variable analysis can determine how much snow pack, followed by how much warm (melting) period, and subsequent freeze period, contributed to a claim event.

Furthermore, this prediction model can determine how the severity of an ice damming event might affect the spectrum of roof types and materials most vulnerable to damage.  Using our forecast alerting app, Beacon, our customers will know what properties to focus on when the next storm hits.  And as they look to expand their business, they’ll be able to rely on Weather Analytics imagery analysis to collect and verify property features congruent with the rest of their portfolio.

The implications of computer-vision don’t stop with claims analysis.  Our fusion of weather and property data supports underwriting (e.g. property-level risk scoring, broken down by regional ice dam ‘zones’), risk mitigation and customer engagement (e.g. incentives for roof repair), and claims preparedness.
The future of computer vision will grow, as industries allow technological advances to charter their ‘known-unknowns.’   Weather Analytics can provide insights into what types of data insurers might what to start collecting – and imagery analytics will allow them to deploy data collection in a cost effective and targeted manner.