Modeling Pixel Level Error in Fine-Resolution Digital Elevation Models: A Regression-Based Approach

This thesis presents and evaluates a new approach for modeling pixel level error in fine-resolution digital elevations models (DEMs), addressing shortcomings of previous studies related to: (1) the assumption of normality for pixel level errors; (2) the scale-dependence of the relationship between local topographic roughness (LTR) and error; and (3) the spatial autocorrelation of error. Ordinary Least Squares (OLS) and spatial lag (SL) regression are used to predict DEM errors for integration into geomorphic change detection analyses. The modeling approach is tested using two terrestrial LiDAR datasets. Pixel level errors prove predominantly non-normally distributed and a new measure for summarizing asymmetric error distributions is introduced. Smaller radii LTR are found to produce better fit models, but also incomplete and less conservative error grids. Lastly, although SL models account for spatial autocorrelation of errors better than OLS models, they prove overly conservative and inferior to at predicting DEM error grids.