Eyes in the Sky: Safer, Faster—and Sharper—Views for Commercial Roof Inspections

By Isaac Wash

eyes-in-the-sky2.jpgTwo ladders loosely joined with rope and duct tape—that was someone’s idea of an extension ladder. And it’s what one commercial building manager provided to an ISO property inspector to access the roof of his building. Another commercial building manager had no ladder, so he offered to hoist an inspector up to the roof using a wooden pallet and a forklift. In other cases, building managers have flat out refused to grant access.

Getting eyes on commercial roofs to assess risk exposure can be quite a challenge. Unlike residential roofs, most commercial roofs are flat and can’t be seen from the ground. Large commercial buildings may have permanent fixtures such as staircases and fire escapes that allow for easy roof access. But smaller commercial buildings often have no such means of access, and this type of building is common in the United States. Property inspectors either have to carry camera poles or tall ladders—or rely on what’s provided by building managers.

All of this would be of little significance if roofs weren’t such a big part of commercial property exposure. The average cost to replace a flat roof is $5 to $10 per square foot.[1] At that rate, replacing a 20,000-square-foot flat roof could run between $100,000 and $200,000. And that’s a relatively small roof by commercial standards! Add bad weather: In some parts of the country, the sheer cost to repair and replace roofs has been exacerbated by an increasing number of damaging hailstorms. Flat membrane roofs are particularly susceptible to hail damage. If property underwriters historically made decisions without much data on roofs, many are now expressing that the gap in information is hurting their profitability, especially in Hail Alley.

The good news is that aerial imagery and artificial intelligence are making it possible for commercial underwriters to fill this gap, saving insurers from the hassles and hazards of roof surveys from the ground. Advances in the quality and resolution of aerial images have, in recent years, allowed aerial cameras to become a cost-effective way to gather detailed information about the built landscape of vast geographic areas. From a mile in the sky, plane-mounted cameras are able to resolve images down to a ground sample distance of 2 inches or better—meaning that each pixel of the image corresponds to 2 inches of distance on the ground. Drones, flying at lower altitudes, can do even better. This degree of resolution allows commercial underwriters to see the features and condition of properties with a level of detail not available from satellite imagery. And the most difficult thing to see from the ground is the easiest thing to see from an airplane: flat roofs! 

Images of flat roofs are especially valuable for small commercial buildings where access is difficult. But there’s another factor to be considered here. Insurance companies are looking for ways to streamline underwriting expenses for smaller risks, which pay lower premiums than larger properties. At times, this means forgoing traditional property inspections. But in some cases, insurers are going a step further by removing the underwriter from the process. The “underwriting” in these cases is done by computer algorithms using sets of predetermined rules for selecting risks and setting terms. In such a scenario, what good is an image if there’s no human to look at it (the aerial imagery analogue to the “If a tree falls in a forest” question)? The underwriting algorithms would somehow have to be able to see the roof in the image and extract information about it. Can computers do such things?

Luckily, there’s a way to train computers to “see” and classify commercial roofs within aerial images, and companies like Verisk are moving full speed ahead on such algorithms so that insurers don’t have to do the heavy lifting themselves. This “computer vision” is made possible by another branch of technology, which coincidentally shares the same initials as aerial imagery: artificial intelligence—more specifically, a subset of artificial intelligence called machine learning. Machine learning refers to the ability of machines to receive a set of data and “learn” from that data, adapting their algorithms as they learn more about the data they’re processing. Just as the human brain recognizes patterns in visual data, machine learning develops “neural networks” that perform a similar function. These neural networks are modeled after the human brain to mimic its ability to classify and categorize information, finding patterns amid complexity. 

What exactly can neural networks “learn” about roofs? Theoretically, any visual feature or element that’s common enough to provide sufficient sample data can be used to “teach” neural networks. In practice, that’s information like the footprint of the building, the roof cover material, the geometry and pitch of the roof, and equipment on the roof. But it can get even more sophisticated: for example, roof condition.

Flying airplanes and drones and training neural networks may sound like a long reach to overcome the challenge of commercial roofs. To be fair, they’re not the only means underwriters can use when boots-on-the-ground roof inspections are too costly or troublesome. Many underwriters have turned to publicly available satellite imagery, which provides decent-quality images of most roofs in the United States.

One difficulty with such imagery is that the resolution is relatively low; images get quite grainy as the observer zooms in. And the images aren’t always time-stamped, leaving underwriters to guess how recent they are. The advantage of satellite imagery is that it tends to have broader geographic coverage than airplane imagery. Some underwriters use municipal tax assessor and building permit data to get information on roofs. Tax assessor records may contain fields for capturing basic roof information, but those fields may not always be populated, and they contain no indication of roof age or condition.

Building permits can be a valuable source for estimating roof age based on the date the roof was last replaced or repaired. But building permits have limitations too. Not all jurisdictions in the U.S. have strong building permit enforcement. Also, the free-form text found in building permits can be challenging to mine in a systematic fashion. In the end, these sources are a supplement, but not an ideal substitute, for aerial imagery.

[1] http://www.roofcalc.org

Isaac Wash is a senior actuarial associate for ISO, a Verisk (Nasdaq:VRSK) business.