An experiment on the elevation accuracy improvement of photogrammetrically derived DEM

This paper focuses on a topic barely considered in the literature: how to improve the accuracy of a given Digital Elevation Model (DEM) irrespective of its lineage by identifying its most suspicious values (also denoted here as outliers). Methods tailored to a specific procedure and source (contour maps, remote sensing image, etc.) exist but they are not valid for other cases. This is a problem for both the producer and end user. The results of a comparison of two methods using six DEMs intended to be representative of different landscapes are reported here. Both methods have been applied to each DEM, producing a number of height candidates to be analysed. Assuming that all candidates are wrong, their elevations have been blindly replaced by interpolated heights, simulating the behaviour of the inexperienced user. The improved (or degraded) DEMs are then compared against the ground truth, and updated accuracy figures are calculated. The RMSE can diminish by 2 to 8% of the original value by changing less than 1% of the elevations in the dataset.