Automatic Mapping of Residential Rooftops using High-Resolution Thermal Imagery and Machine Learning

MyHEAT Inc. (industry partner) provides on-line tools/services to reduce urban waste energy. Currently, the company relies on sourced municipal, private, or publicly available GIS roof polygons which it combines with its proprietary high-resolution (H-Res) airborne thermal infrared (TIR) imagery to generate personalized rooftop heat-loss maps/metrics. Unfortunately, these GIS polygons are often incomplete, inaccurate and out-of-date. To mitigate these issues, this project proposes two main goals: (1) test and optimize two leading-edge Convolutional Neural Network (CNN) methods (SegNet and U-Net) for automatic and accurate rooftop delineation from MyHEAT’s existing TIR imagery, and (2) define the optimal TIR spatial resolution for CNN based rooftop delineation. The key benefits to MyHEAT include: (i) reduced data acquisition/processing costs as their optimal resolution TIR imagery will be the only data source required for heat-loss metrics, and (ii) speeding up their entire analytical pipe-line, as there will be no need to acquire, correct, or negotiate for sourced GIS roof data.

Faculty Supervisor:

Geoffrey Hay;John Yackel;David Goldblum

Student:

Salar Ghaffarian

Partner:

MyHEAT

Discipline:

Geography / Geology / Earth science

Sector:

Professional, scientific and technical services

University:

University of Calgary