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

Abstract

Lidar-derived digital elevation models (DEMs) are crucial for modeling salt marsh evolution, forecasting inundation depth, frequency, and duration, and simulating sea level rise. Advances in lidar acquisition and data processing techniques have led to increased accuracy; however, in densely vegetated coastal salt marsh areas, lidar-derived DEMs are generally unreliable without adjustment. In this article, we investigate the need for local topographic ground truth data to train random forest (RF) DEM adjustment models for two similar northern Gulf of Mexico salt marshes. Two RTK-GNSS field surveys were conducted to acquire ground truth topographic elevations near St. Marks, Florida, USA (n = 377) and Pascagoula, MS, USA (n = 610). These elevations, along with Sentinel-2A MSI reflectance values and lidar DEM elevations, were used to validate and train local and combined RF salt marsh DEM adjustment models. The local RF models achieved mean absolute error values of 0.054 m and 0.045 m in the leave-one-out cross-validation for St. Marks and Pascagoula, respectively. Elevation bias predictions using remote RF models were far worse and those using the combined RF model were marginally worse. Using the local RF predictions to mitigate the bias in the lidar DEMs improved their accuracy by 69.1% for St. Marks and 90.9% for Pascagoula. The DEM elevation was identified as the most important predictor. This evidence suggests that local ground truth data are necessary for mitigating bias in salt marsh lidar DEMs although it remains to be seen if increasing the data set size and incorporating additional hydrologic predictor variables could narrow the accuracy gap.

Keywords

• Coastal
• lidar
• multispectral
• remote sensing
• Sentinel