Multiple-sensor photogrammetric base model alignment of large timeseries in physical scale experiments

Physical scale experiments of natural systems such as debris flows, rivers, estuaries and deltas are conducted within the geosciences to enhance our understanding of the physical processes on Earth. Methods in these experiments are under continuous development, often inspired by the advancements in remote sensing, making use of e.g. overhead fixed position cameras, movable gantry mounted cameras and line laser scanners, which are analogous to fixed-camera timelapses, drone surveys, and LiDAR, respectively. However, physical scale experiments come with additional challenges, as small uncertainties in positioning and orientation of the sensors can significantly bias the experimental results, especially because the small-scale morphology is typically in the order of a few centimeters. Moreover, photogrammetric data processing is prone to doming which can vary for each timestep, impeding change detection studies.

We will present the advancements in the data processing of our 20 by 3 m laboratory facility that is used to emulate tidal cycles to induce morphological development in coastal systems (www.uu.nl/metronome). Mounted ~4 m above our facility are 7 lower-grade overhead cameras (pixel resolution ~1.5 mm, overlap ~20%) which are simultaneously triggered at each tidal cycle. Additionally, when an experiment is paused, we conduct both DSLR surveys (pixel resolution ~0.5 mm, overlap ~80%) and 3D laser triangulation system surveys producing gridded data (planimetric resolution ~1 mm), along a movable gantry system. We have built a large dataset of >220.000 experimental tidal cycles and >2.000.000 unique images, which requires fully automated data processing that results in morphology which is consistently accurate in both the spatial and temporal domains of our timeseries.

We developed an extensive data processing workflow that incorporates a base model of our facility under idealized conditions. This base model was photogrammetrically constructed in Agisoft Metashape by aligning a total of 169 images from all cameras without downscaling. Through an automated and fast-performing python script, we are able to successfully align this base model to our complete dataset of DSLR-gantry surveys to generate orthomosaics and DEMs, as well as overhead imagery to generate timelapses of orthomosaics. This method shows a striking degree of robustness, because it has no difficulty with aligning 7 unique overhead cameras with limited overlap, as well as imagery of increasingly different morphology compared to the base model.

Finally, positions (X,Y,Z) and orientations (ω,φ,κ) of the cameras along the gantry were extracted from the base model, which were implemented in a new workflow that processes the raw laserscan data using vector algebra and transformation matrices. This results in DEMs that are geometrically aligned to the base model, without the use of a photogrammetric method. Accordingly, these DEMs have no variability in doming throughout the timeseries and therefore a higher temporal accuracy. Moreover, the implementation of variability in camera positions and orientations results in an improvement in altimetric accuracy from 7 to 2 mm (99.7% confidence), significantly reducing the bias in small-scale morphology. Our methods can be partially or fully implemented in research and industry using small to medium scale setups of both fixed and gantry-mounted camera systems.