In: Wagner W., Szekely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B 
628 
DEM GEMERATION FROM AIRBORNE LIDAR DATA BY AN ADAPTIVE DUAL- 
DIRECTIONAL SLOPE FILTER 
C. K. Wang 3 , Y. H. Tseng 3 ’* 
Dept, of Geomatics, National Cheng Kung University, No. 1, University Road, Tainan, 701, Taiwan - (p6896102, 
tseng)@mail.ncku.edu.tw 
KEY WORDS: LIDAR, DEM, Generation, Classification, Aerial 
ABSTRACT: 
An airborne LiDAR system is capable of collecting three-dimensional information over a large area effectively. Because of the rapid 
data collection, DEM generation using airborne LiDAR data has become a standard process since last decade. Filtering out non 
ground points from point clouds to obtain terrain relief is the key process for DEM generation from airborne LiDAR data. Many 
filtering methods have been proposed for this process. Basically they can be categorized into three main approaches: linear 
regression methods, slope-based methods, and morphology-based methods. Filters apply a certain assumption of smooth terrain, 
which cause an over-filtering problem in some terracing fields and cliff areas. This paper proposes an adaptive dual-directional filter 
based on a slope filter to deal with this problem. While the original slope filter is performed according to its whole adjacent covered 
window, the dual-directional adaptive filter is designed along an alternative direction in one dimension. The main difference 
between them is the designed filter shapes. The adaptive filters of different directions are complementary to each other, so that over- 
filtering situation can be avoided. Comparing with original slope-based filter and the commercial software TerraScan, our method 
shows better results in handling data of abrupt surfaces. The data used for comparison is the ISPRS test data. The variance, omission 
errors and commission errors are shown for the comparison. Our method has better performance in avoiding over-filtering situation 
and can keep as good accuracy as the compared methods. 
1. INTRODUCTION 
The rapid development of Airborne LiDAR system makes the 
acquisition of three-dimensional surface information more 
conveniently and directly. Comparing with photogrammetry, 
LiDAR system has two advantages which are more cost- 
effective to obtain the vertical information over a large area and 
fewer pre-processing of data(Meng et al., 2009). Besides, the 
less limitation of weather and time of a day to enable a LiDAR 
measurement assignment also makes LiDAR more and more 
popular in obtaining 3D information(Shan and Sampath, 2005). 
Many applications have been applied so far, e.g. mapping of 
corridors, mapping of transmission lines, measurement of 
coastal areas, rapid mapping and damage assessment after 
natural disasters, ground surface modelling, object 
classificationAxelsson, 1999; Wehr and Lohr, 1999) and so on. 
Among them, DEM generation is the most important 
application and has become a standard process. 
Since the raw data of LiDAR encodes the 3D coordinates 
already, DEM generation using Airborne LiDAR data can be 
simplified only by filtering non-ground points out from point 
clouds. Many methods about filtering have been proposed. 
Sithole and Vosselman (2004) and Zhang and Whitman (2005) 
have compared some of these methods. Generally, these 
methods can be categorized into three main approaches: linear 
regression methods, slope-based methods, and morphology- 
based methods(Silvan-Cardenas and Wang, 2006). For linear 
regression, Kraus and Pfeifer (2001) present two models which 
are the stochastic model and the functional model to estimate 
the ground surface. The stochastic model defines a weight 
* Corresponding author. 
function and the functional model determines the interpolated 
ground surface, In their research, an approximate ground 
surface would be estimated and then the residuals between point 
clouds and the estimated surface can be calculated. The residual 
of a point would give the point a new weight through stochastic 
model and then a new approximated surface would be estimated 
again until a regression stop condition is satisfied. For example, 
the difference between later and previous estimated ground 
surface is slight. For slope-based filter, Shan and Sampath 
(2005) use slope and elevation conditions to determine a point 
is a ground point or non-ground point. Since the slope value 
usually would be significantly large between non-ground areas 
and ground areas, and the elevation of ground surface usually 
higher than non-ground surface, using the above two conditions, 
they design a mathematical model to describe the ground 
surface. For morphology-based filter, it is known that the 
opening operation would smooth tall objects. For filtering out 
non-ground objects which are taller than grounds, opening 
operation is therefore suitable as a DEM filter. However one 
major problem in using morphology-based filter is the window 
size of the filter. If a large size of window is chosen, the non 
ground objects can be smoothed effectively but an over 
smoothed situation will occur in the abrupt surface. For this 
reason, this paper develops a dual-directional filter to overcome 
this problem.