Photogrammetric point clouds are notoriously noisy, very dense, and generally lack points located underneath vegetation. Echo information is non-existent and surface roughness are not as useful for ground classification with these point clouds. Point clouds from low quality LIDAR sensors exhibit some similar characteristics. Hence, a different approach is needed
In general, it is a good first step to disregard isolated points because they are never usable and almost always not meaningful (i.e. noise). This can be accomplished using the ‘Classify isolated points’ routine which reduces noise to some degree.
Next, a potential surface needs to be identified within this still relatively noisy point cloud. One of the most used surface classification routines is the generic ‘Ground’ classification. This routine iteratively creates a triangulated surface model beginning with low points on the surface. This routine is particularly useful for traditional airborne LIDAR, but not very suitable for these noisy datasets. Its bottom-up logic will be compromised by the noisiness in the data resulting in a potential surface that is not an accurate depiction of the real world. An alternative to the traditional ‘Ground’ classification routine is the ‘Hard surface’ Routine. This is based on a planar fit algorithm. The Hard Surface routine classifies the dominant, median points of a surface which provides for better performance on horizontal paved surfaces and is most often used for mobile LIDAR datasets. This tool would work decently for horizontal or fairly flat surfaces within this dataset, but would have a difficult time classifying the median surface of natural terrain and vertical features.
Read Complete Article: Classify Surface Points Routine