Locality of Topographic Ground Truth Data for Salt Marsh Lidar DEM Elevation Bias Mitigation Using Machine Learning

Light detection and ranging (lidar) digital elevation models (DEMs) are crucial for modeling coastal salt marsh systems, simulating the coastal dynamics of sea level rise (SLR), and predicting storm surge inundation depth and duration. Improvements in lidar acquisition technology and data processing over the last decade have led to increased accuracy. However, the lidar-derived DEMs for coastal salt marshes that are densely vegetated are generally unreliable without adjustment based on local ground truth elevations. In this study, Random Forest (RF) DEM adjustment models are trained for two similar Northern Gulf of Mexico salt marshes. The need for local topographic ground truth data to train the models is also investigated. Two Real-Time Kinematic (RTK) GNSS field surveys were conducted by others to acquire ground truth elevations near St. Marks, Florida (n=377) and Pascagoula, Mississippi (n=610). These elevations, along with lidar elevations and Sentinel-2A multispectral satellite imagery (MSI) reflectance values were used to train the RF salt marsh DEM adjustment models and apply them under two scenarios: local and non-local. A local adjustment relies on data collected within the adjustment domain to train the model whereas a non-local adjustment uses data collected outside the adjustment domain. The RF-local models achieved the lowest mean absolute error (MAE) values for St. Marks and Pascagoula. The predictions using non-local RF models were unsatisfactory. The evidence suggests that local ground truth data are necessary for mitigating bias in salt marsh lidar DEMs, although future work should investigate if increasing the data set size could narrow the accuracy gap. This mitigation adjustment technique can be replicated in other coastal regions with similar vegetation profiles. As the world becomes increasingly vulnerable to the effects of climate change and SLR, it is important to accurately characterize the current state of the system to model marsh restoration and migration, natural and nature based protective infrastructure, and land use planning policies, for example