Developing bare-earth digital elevation models from structure-from-motion data on barrier islands

Unoccupied aerial systems can collect aerial imagery that can be used to develop structure-from-motion products with a temporal resolution well-suited to monitoring dynamic barrier island environments. However, topographic data created using photogrammetric techniques such as structure-from-motion represent the surface elevation including the vegetation canopy. Additional processing is required for estimating bare-earth elevation, which is critical for understanding the underlying geomorphology of these islands. In this study, we used a vegetation and elevation survey to produce bare-earth digital elevation models from structure-from-motion-derived elevation products for two sites on Dauphin Island, Alabama (USA). One site was exposed to high wave energy and included a mix of beach, dune, and barrier flat habitats that were dominated by supratidal/upland herbaceous vegetation. The second site was exposed to low wave energy and was dominated by intertidal marsh. Aerial imagery was collected in late fall of 2018 and spring of 2019. We tested several machine learning algorithms for predicting and removing elevation bias for vegetated areas using predictors that included spectral indices from unoccupied aerial systems-based multispectral imagery and landscape position information (e.g., relative topography and distance from shore). Models were developed for each site and season. We also explored how well the model from one season generalized to data from a different season for the same site. For developing initial digital surface models, we found that utilizing a minimum bin algorithm, as opposed to interpolation, led to lower elevation bias. For bias removal, Gaussian process regression performed the best and led to a root mean square error for the bare-earth digital elevation models of around 0.10 m for the high energy site and 0.15 m for the low energy site. Compared to the digital surface models, the root mean square error for the bare-earth digital elevation models was reduced by at least 29 percent for the high energy site and 69 percent for the low energy site. For all models, common predictors included surface elevation, vegetation greenness, and distance from the shoreline. The models produced comparable results when trained using data from a different season. The error estimates for all analyses were within published elevation standards for lidar data for vegetated areas. With calibration, this approach could be portable to other areas or data, such as aerial lidar (conventional or unoccupied), to provide an efficient and repeatable framework for monitoring geomorphology or provide baseline elevations for predicting changes to these environments under future conditions. More here.