Presenter: Jenna Brown (USGS)

Description:
Topographic maps produced from aerial imagery and Structure-from-Motion (SfM) photogrammetric processing are used to study changes in coastal environments at high spatial and temporal resolution. One disadvantage of these elevation products is that they represent the surface and not the bare earth. When anything covers the ground, including vegetation, structures, or even people and their gear at popular recreational beaches, the resulting obstructions make it difficult to accurately measure bare earth surfaces, volumetric changes, cross-shore profiles, and alongshore metrics such as shoreline position. This was the case for a 12-month dataset of SfM-derived topographic products at Madeira Beach, Florida; “beach gear” (chairs, umbrellas, etc.) occupy the beach year-round, some measuring 2m tall with 2m² of surface area, collectively resulting in a non-trivial surface volume that should not be included in coastal change studies. Described here is a semi-automated method for detecting and removing non-bare-earth classes of interest in imagery datasets and using corresponding digital surface models (DSMs) to approximate bare earth digital elevation models (DEMs). First, orthophotographs are tiled and image segmentation is performed using the human-in-the-loop Machine Learning (ML) program “Doodler” to label areas at the pixel level for each user-defined class (e.g., “beach gear”). For larger datasets, a subset of the imagery can be labelled with Doodler, then used to train a ML model to predict labels for the remaining imagery. The points of each class of interest are identified in the corresponding SfM-derived DSM and removed, and the bare earth DEM is estimated using spatial interpolation. Limitations of this method, such as size and underlying topography of classes being interpolated over, are discussed. Overall, the method was effective for removing anthropogenic “beach gear” from the dataset and improved the volumetric calculations and beach profile extractions.