Digital Elevation Model Generation using Highly Oblique Stereo Imagery via Structure from Motion in a Coastal Area

Generating a precise Digital Elevation Model (DEM) is crucial for coastal inundation predictions, as it provides essential elevation data necessary for accurate modeling of potential flooding. Coastal inundation predictions are important to anticipate and mitigate the impacts of rising sea levels, tropical storms, and other extreme weather events. This study focuses on an approach of generating DEMs using highly oblique stereo images coupled with non-metric cameras. These images are part of a time series captured at the beach located next to Horace Caldwell pier in Port Aransas, Texas. To do so, a pair of non-metric cameras, Amcrest Ultra4K, were mounted on the corners of an elevated building located at the entrance of the pier and overlooking the beach, and often overlapping with water runup on the beach. The cameras are about 8m above sand and the distance between them is ~20m, with a viewing angle of ~60 degrees from nadir. Generating highly accurate DEMs from oblique imagery is challenging since most software and equations are intended for vertical imagery. Structure from Motion (SFM) provides users with an accessible and efficient approach for generating three-dimensional models from images. However, challenges and limitations persist when utilizing stereo highly oblique images, particularly those captured with non-metric cameras. Overcoming these challenges would provide advantages such as simplified data acquisition, cost effectiveness, and flexibility in capturing diverse angles compared to conventional photogrammetry. The proposed approach enables efficient and accurate elevation assessment in a dynamic coastal environment through applying a series of calibration and image processing techniques including wavelet transformation, rectification, georeferencing the highly oblique images, and using SFM approach for creating the three-dimensional point cloud. In this study, to enhance feature matching, we employed a wavelet transformation to identify high-frequency features in vertical, horizontal, and diagonal directions. Subsequently, we incorporated these resultant images into the original image, effectively augmenting the sharpness of the initial image. In order to achieve a more precise DEM from highly oblique images, it is imperative to generate an orthoimage that undergoes a process of warping the source image. This transformation ensures consistent distance and area proportions in relation to real-world measurements. To accomplish this, the present study employed rectification within the MATLAB environment, while the georeferencing was executed using projective transformation within ArcGIS Pro software. To evaluate its vertical accuracy, the created DEM was compared with 14 RTK-GPS check point observations collected biweekly on the beach. The check points locations were measured using terrestrial surveying based on a 5´5 meter grid, and were compared with the DEM generated by structure from motion (SFM). The obtained vertical Root Mean Square Error (RMSE) using this process was approximately 0.21m.