Reducing vertical bias and error in tidal marsh digital elevation models with machine learning and LiDAR derivatives

Abstract

The vertical bias and uncertainty of LiDAR-derived elevation data in tidal marshes is a major challenge for researchers in these critical ecosystems. Small positive bias errors in land surface elevation can lead to large underestimates of coastal inundation. Previous studies have shown that the bias and uncertainty of elevation data can be reduced using a variety of statistical techniques. However, many studies cover a relatively small extent, cover a large extent at coarse resolution, or rely on local data products that may not be widely available or are of unknown quality. This study aimed to develop and compare bias reduction approaches for digital elevation models (DEMs) using LiDAR derivatives in tidal marshes along Delaware Bay, Delaware, USA, a region that experiences high rates of relative sea-level rise and frequent coastal flooding. In this study, several statistical and machine-learning approaches were evaluated based on their ability to reduce vertical DEM error using field GPS survey data and LiDAR and terrain derivatives as inputs. The evaluated approaches for reducing DEM error included ensembles of multiple linear regressions, kernel-K Nearest Neighbors, gradient boosted regression trees, random forests, and deep neural networks (DNNs), which were compared based on statistical performance metrics. An ensemble of deep neural networks performed best at removing vertical DEM bias and reducing DEM uncertainty, though other approaches also performed well. GPS survey data indicated a mean absolute error, mean bias, and root mean square error in GPS surveys of 11.9 cm, 10.9, and 14.8 cm, respectively, in the original DEM. These error metrics were reduced to 3.5 cm, −0.2 cm, and 5.1 cm in the DNN ensemble-corrected DEM. The DNN ensemble was then applied to all tidal marshes throughout the study area. This corrected DEM clearly showed how interpretations of marsh platform inundation based on uncorrected DEM surfaces could be misguided. The approach used in this study only requires LiDAR-derived products and field surveys, so it may be applied readily in other regions.