-B4, 2012 
DTM 
00 University 
Road South, 
| one an idea of 
by dividing the 
ferences in later 
ince the outliers 
to identify cells 
dentified as off- 
each level. Once 
and followed by 
t assume a prior 
atasets covering 
terrain or gentle 
nma in choosing 
rindow size, the 
' point as terrain; 
relief variations 
hard to remove 
A lot of research 
ially to filtering, 
) However the 
en a barrier in 
ic filter and it is 
airborne LiDAR 
lows. Section 2 
thod. Section 3 
S WG III/3 test 
1e study. 
(MTF) method 
rain models from 
to-fine way. As 
point cloud pre- 
'M refinement. 
r scanning points 
oints of multiple 
ith regard to the 
data set. Two 
are to assign the 
following MTF 
nd noises. 
International Archives of the Photogrammetry, Remote Sensin 
g and Spatial Information Sciences, Volume XXXIX-B4, 2012 
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia 
  
Input: LiDAR Points ji 
LL 
Preprocessing 
-Non last return points removing 
-Point layering 
-Outlier and noise removal 
40 
Multi-scale Terrain Filtering 
-Rasterized pyramid level generation 
-Point identification 
-Off-terrain point interpolation 
J 
DTM Refining 
-nDSM feedback adjustment 
-Classification by refined DTM 
-IDW Interpolation 
Output: DTMs A] 
Figure 1. Flowchart of the MTF method 
  
  
  
  
  
  
  
  
  
  
  
  
Input Lidar Image; 
/[Pre-processing 
points.SelectLastReturns(); 
points.HistogramGeneration(); 
points.Layering(); 
points. NoiseElimination(); 
/[Points identification and interpolation 
level[n] = points.PyramidLevelsGeneration(); 
for(level j = n-1 to 0){ 
if(j=n) {reference} 
for(points[i] in level j){ 
if(points[i].layernum==reference) { 
points[i]-terrain point; 
else 
points[i]=off-terrain point;} } 
for(points[i] in level j){ 
if(points[i]==off-terrain point) { 
points[i].z=points[i].interpolation(); 
points[i].layernumRenew();} } } 
//DTM Refinement 
Rough DTM = RasterGeneration(points[i]); 
Refined DTM = nDSM Adjustment( 
Lidar Image.First Returns, Rough DTM); 
Terrain Points = Filtering(Refined DTM, 
Lidar Image.Last Return); 
FinalDTM = TerrainPoints.IDW Interpolation() 
, 
  
  
  
Output DTM; 
  
Figure 2. The pseudo-code of the proposed method 
The Second step is the multi-scale terrain filtering, which 
includes the rasterized pyramid level generation, iterative point 
identification and interpolation. Several rasterized pyramid 
levels are generated at first, and lowest points of every grid in 
every level are marked as representative points. The highest 
level is referred to an initial digital terrain model, from which 
the Proposed MTF is employed as a reference. Then the 
identification and interpolation is iteratively processed in every 
level from the second highest level to the lowest level in the 
pyramid. The identification is based on comparing the layer 
numbers generated in layering and a slope-based threshold. This 
is followed by the interpolation at identified off-terrain points. 
The points from a processed level then become reference points 
in the identification of next level. Iteratively, DTMs are 
recovered and densified from coarse scales to fine scales. 
In the third step, the terrain results are adjusted based on the 
normalized Digital Surface Model (nDSM). The original 
LiDAR point cloud data are filtered based on the refined DTM. 
The separated terrain points are applied to generate the final 
DTM through the IDW interpolation. As a result, this produces 
the final and complete DTM. Figure 2 presents the pseudo-code 
of the proposed method. 
3. RESULTS AND DISCUSSION 
The dataset used in this study was released from the ISPRS WG 
III/3, have been made available through the web site (www. 
commission3.isprs.org/wg3/). A total of 15 sites were selected 
to test the performance of our multi-scale terrain filtering 
algorithm and compare the results with other methods evaluated 
by ISPRS (Sithole and Vosselman, 2004). The LiDAR point 
cloud data were captured by an Optech ALTM scanner and the 
reference data were generated manually. Those data are located 
along seven study sites over the Vaihingen test field and the 
centre of Stuttgart City, Germany. The study cites have varied 
terrain characteristics and diverse feature content (e.g., open 
field, vegetation, building, road, railroad, river, bridge, 
powerline, water surface, etc.). Those sites are listed in Table 1. 
This dataset covers many different land features and filtering 
difficulties. However, it does not contain small woods and 
residence in urban area. And the reference data is only available 
for the 15 samples. The reference data for entire area is not 
available, which means will limit the algorithm testing for a 
large area. 
Sithole and Vosselman (2003b) reviewed and compared eight 
filtering algorithms, and their comparing method and data are 
frequently cited and applied in many researches of the laser 
scanning data filtering (Pfeifer and Mandlburger, 2008; Liu, 
2008; Briese, 2010; Meng et al., 2010). This paper adopts part 
of their assessment method and tests the performance of the 
MTF method. According to Sithole and Vosselman (2003), the 
cross-matrices are applied in this study to quantitatively analyze 
the Type I, Type II errors and their relationship. Type I errors 
are the errors which wrongly identified terrain points as off- 
terrain points, and Type II errors are the errors which wrongly 
identified off-terrain points as terrain points. 
The proposed method was developed on C++ by Visual Studio 
2008. The morphological filter and adaptive TIN filter 
compared in this research are included in the ALDPAT Version 
1.0, which was developed by the International Hurricane 
Research Center, Florida International University in 2007. The 
final IDW interpolation and accuracy evaluation are processed 
on ArcGIS 10. The processor of the computer is equipped with 
Intel Core2 Duo CPU T5800 @ 2.00 GHz and 4 GB RAM. 
345