SALT MARSH ELEVATION UNCERTAINTY CORRECTION USING FULL-WAVEFORM LIDAR AND NONPARAMETRIC PREDICTIVE MODELING

Lidar is an important source of elevation data for studying, monitoring and managing salt marshes. However, previous studies have shown that lidar data tend to have greater vertical uncertainty in salt marshes than in other environments, which hinders the ability to analyze small, ecologically significant, elevation differences. Previous attempts at improving salt marsh lidar data have ranged from interpolation/filtering methods, to subtracting off the global elevation bias, to computing vegetation-specific, habitat map-based constant correction factors. It is hypothesized that correcting salt marsh lidar data with nonparametric regression using location-specific, point-by-point corrections, computed from lidar waveform-derived features, tidal-datum elevations, distance from shoreline and other variables, will produce better results. RTK-GNSS measurements of ground elevation were collected on tidal and marsh surfaces for three marshes in Cape Cod, Massachusetts, to be used as learn/test samples for model development/evaluation. Five different nonparametric regression algorithms were evaluated with the same dataset. The TreeNet algorithm consistently produced the best results and using all predictor variables it produced an R² value of 0.98 and slopes within 4% of a 1:1 correlation with RTK GNSS ground elevations. Uncorrected lidar in vegetated areas exhibited a positive bias of 0.24 m with a 0.23 m standard deviation, while the correction essentially eliminated the overall elevation bias (µ = 0.00 m). An even more significant result is that, when examining the error statistics for the entire data set, the point-by-point elevation correction also reduced the standard deviation of elevation residuals about the mean down to 0.07 m. The model was then scored on a full marsh dataset to create corrected DEMs and vegetation classification maps. The developed methods are very promising for correction of salt marsh lidar data without the need for a priori vegetation data.