Structure from Motion for aerial thermal imagery at city scale: Pre-processing, camera calibration, accuracy assessment

Airborne thermal cameras are a valuable source of information for energy analyses at city scale. The generation of accurate high-resolution thermal orthomosaics is a necessary but still challenging task, especially when a thermal camera is the only imaging sensor on-board, because of the peculiar characteristics of thermal imagery (i.e. low dynamic range and poor detail definition), large geometric distortions induced by the optical system and weak acquisition geometry. This paper discusses potentials and limitations of Structure from Motion approach for the automated generation of thermal orthomosaics, with the aim to define the best practices and assess the achievable accuracy. After processing with different strategies two thermal flights over a 10 km2area in Bologna city (Italy), it can be concluded that the absolute planimetric accuracy can be in the order of 3–4 pixels and the best results are obtained when computing camera calibration on a smaller subset of images, with a limited number of ground control points and an adaptive fitting algorithm. The analysis of generated point clouds (compared with reference LiDAR data) and calibration reports, in addition to check point residuals, proved to be crucial for a proper accuracy assessment.