Optimizing the Processing of UAV-Based Thermal Imagery

The current standard procedure for aligning thermal imagery with structure-from-motion (SfM) software uses GPS logger data for the initial image location. As input data, all thermal images of the flight are rescaled to cover the same dynamic scale range, but they are not corrected for changes in meteorological conditions during the flight. This standard procedure can give poor results, particularly in datasets with very low contrast between and within images or when mapping very complex 3D structures. To overcome this, three alignment procedures were introduced and tested: camera pre-calibration, correction of thermal imagery for small changes in air temperature, and improved estimation of the initial image position by making use of the alignment of RGB (visual) images. These improvements were tested and evaluated in an agricultural (low temperature contrast data) and an afforestation (complex 3D structure) dataset. In both datasets, the standard alignment procedure failed to align the images properly, either by resulting in point clouds with several gaps (images that were not aligned) or with unrealistic 3D structure. Using initial thermal camera positions derived from RGB image alignment significantly improved thermal image alignment in all datasets. Air temperature correction had a small yet positive impact on image alignment in the low-contrast agricultural dataset, but a minor effect in the afforestation area. The effect of camera calibration on the alignment was limited in both datasets. Still, in both datasets, the combination of all three procedures significantly improved the alignment, in terms of number of aligned images and of alignment quality.