Drones, New Intelligence Platforms Offer Carriers Highly-Accurate Alternative to Manual Roof Measurements
Written by Dan Ciprari, CEO and co-founder, Pointivo Inc.; Justin Kestner, president and CEO, Haag Engineering Co.; David Shearer, vice president, Marketing, Kespry
The use of drones has grown tremendously in just a few years. While figures vary, sales grew 60-100 percent between 2015 and 2016, and represent as much as $4.5 billion in revenue, according to a recent Recode article.
The insurance industry has jumped on the drone bandwagon, using drones as a means to document roof conditions and calculate roof areas for determining premiums and evaluating claims. When roof measurement reports based on aerial imagery first appeared approximately 10 years ago, it was unclear as to the precision and reliability of aerial-based measurements. While unmanned aircraft system (UAS)-generated measurements can be much faster and eliminate the potential for injury during manual measurement, the debate about accuracy continues.
An independent accuracy study was recently completed to validate the precision of roofing intelligence algorithms, which automatically extract roof geometry and measurements from drone imagery, and compare these to manual measurements.
Experienced field surveyors independently measured 13 roofs using traditional survey methods, while an autonomous UAV flew over these roofs to capture images and generate 3D models. Roof pitches ranged from flat to 12:12 and individual roof areas spanned approximately 10 to 62 squares. The test included 17 buildings, totaling approximately 535 squares (one roofing square equals 100 square feet). All sloped roofs were asphalt composition shingles, the most popular type of sloped roofing in the U.S. Flat roofs were modified bitumen. All properties were located in the Dallas-Ft. Worth Metroplex.
Drone software delivered these models to a cloud-based intelligence platform via its public application programming interface (API), which allowed complex calculations to run in a specialized computing environment. On the platform, proprietary computer vision and machine learning algorithms automatically extracted the measurements. Manual measurements and automatic measurements for each property were compared, specifically comparing individual areas of roof planes and edge lengths.
When comparing roof area, the study found that for the 13 roofs that were measured both manually and through automated means, variations ranged from +1.2 percent to -2.7 percent per individual roof. The average difference of 0.6 percent was within the industry goals of +/- 2 percent. When comparing differences in absolute values, the average variation was 1.1 percent, still comfortably within the +/- 2 percent range. Automated measurements were highly accurate on edge lengths, as compared to manual measurements, thus providing the highly accurate area results.
The Testing Process
The drone included a proprietary UAS platform, which includes autonomous drone flight, high-resolution image capture and 3D processing in the cloud. The data was then transferred to the intelligence platform that included computer vision and machine learning algorithms, which detected the roof structure, identified specific roof planes and extracted accurate geometry and measurements for the entire roof to generate a detailed CAD model. These automated measurements were then sent for comparison with the manually collected measurements, which included lengths for each roof edge and area and slope for each roof plane.
A separate team measured the roofs independently and without knowledge of the automated measurements or calculations. Automated measurements were rounded to the nearest millimeter and manual measurements rounded to the nearly ¼ inch, even though measurements to the nearest inch is a typical industry practice.
The manual and automated measurements were then compared. Areas of these roofs totaled approximately 412 squares and included 106 individual slopes. The automated measurements for calculated roof area totaled 0.6 percent less than the manual measurements. The greatest roof area difference was 95 square feet (2.7 percent of the roof area) while the smallest was just 3 square feet (0.2 percent). The roof with the greatest difference, #6, was covered by overhanging tree branches along its front edge. Roof #9 contained a flat roof section that measured approximately 17 squares and was partially overhung by the adjacent sloped roof. The automated computed area for this flat roof section was within 1 percent of the manually-calculated area. In terms of absolute value, the average difference between automated and manual measurements was 1.1 percent of the roof area – well within the +/- 2 percent unofficial industry benchmark.
The final report summed up the results well. “The automated solution proved reliable for the 13 roofs sampled, as the total area computed was only 0.6 percent different than that of the manual measurements.” These results proved the UAS can be a viable option to capture accurate roof measurements from the safety of the ground.
The areas could be measured and calculated by the UAS much faster than by manual means, and much more safely. In fact, the original intent of the study was to measure 17 roofs, but 4 of the 17 were too slick and/or steep to reliably measure manually, without better weather and/or a rope and harness.