Removing non-ground points from automated photo-based DEM and evaluation of its accuracy with LiDAR DEM

Abstract

Three sets of DEM, including LiDAR, stereo photo based DEM, and contour-based DEM, were created using different tools and sources. Because different tools produce different raster data sets, they were reprojected into the same coordinate system and converted to point clouds as vector format, then a triangulated irregular network (TIN) retaining all grid or mass points was created from each point coverage. Corresponding orthophoto involved in extracting stereo photo DEM, classified by object-oriented approach to create vector polygons representing non-ground points (building and vegetation classes) and bare-ground elevation points. Non-ground points were removed from stereo pairs DEM using classified orthophoto polygons, and filled with contour DEM data. Also a 5th order trend surface over photo-based DEM was fitted and non-ground points were removed and filled using local interpolation. It was observed that automated photo-based extraction yields high precision terrain models in a short time, reducing manual editing; their accuracy is strictly related to image quality and terrain features. Results showed that such photogrammetric extracted DEM represents better accuracy along x and y directions than LiDAR does, while LiDAR has the best vertical accuracy, compared to other DEMs. The differences between horizontal errors are large since there were no significant differences between vertical errors of LiDAR and photo-based DEM. This indicates that there is a good correlation between elevation points of DEMs, and a stereo pair-based DEM can be a good substitute, whenever LiDAR is not affordable. This study provides several important insights into the magnitudes and spatial patterns of LiDAR and photo-based DEM errors, further studies need to verify the error extent in more diverse landscape. However, automated photo-based DEM extraction is currently an efficient method for collecting data useful for rural and small study area.