Automatic Mapping of Residential Rooftops with High-Resolution Thermal Imagery

This study demonstrates the application of the commercially available ENVI Deep Learning module for automated rooftop extraction and evaluation for an industry partner using high-resolution (25 cm) night-time thermal infrared (TIR) airborne imagery. Specifically, we describe (i) generating accurate (93%+) GIS-ready rooftop polygons directly from TIR data, (ii) identifying the operationally optimal spatial resolution (40 cm) for deep learning-based TIR rooftop delineation, and (iii) comparing the performance of multi-resolution models (25 cm to 1.0 m in 5 cm increments) developed from a single TIR image versus a derived three-channel image. We then (iv) introduce five well known pixel-based accuracy assessment methods but apply them within a novel object-based assessment, and (v) we introduce Converging Factor (CF) as a new boundary-based accuracy method to evaluate the visual fit of the segmented rooftops.