ISPRS Commission III, Vol.34, Part 3A »Photogrammetric Computer Vision“, Graz, 2002 
one is making use of “first minus last” return analysis and the 
second step is utilizing the local statistical interpretation. The 
two techniques are illustrated below in more detail. 
2.1 First minus last return height analysis 
The LIDAR system, an Optech ALTM 1210 operated by 
Woolpert Consultants, has the ability to capture two returns 
(first and last) per each height point. This is due to the fact that 
the laser pulse is not a single ray but an extended solid angle. 
It has an angular beamwidth and its footprint will a take 
circular shape when it reaches the ground. Based on the laser 
characteristics and the scene characteristics, the laser beam 
could penetrate some objects. Therefore some of its energy will 
be reflected back from the object top surface and other portions 
might penetrate to different depths before they are reflected as 
shown in figure 1. Generally, this produces an extended return 
signal. Therefore, the computed height based on the first 
received return will be called first return height. Those heights 
contain more noise since they reflect every object on the 
ground such as trees, cars, and buildings as shown in figure 2. 
On the other hand, the computed height based on the last 
received return will be called last return height. Those heights 
represent only the non-penetrable objects such as the ground, 
and buildings as shown in figure 3. So each derived height 
point will have two recorded the heights, the first return height 
and last return height. As a result of that, two different heights 
of one point give an indication of the presence of a penetrable 
object such as tree. In contrast, if a data point has the same 
height for first and last returns, then this point belongs to a 
non-penetrable object. This step is just to locate the tree 
regions in the scene by examining the difference between first 
and last return. 
First return 
A^ 
| 
Last return 
  
Figure 1. first and last return 
The level of the discrepancy between first and last return 
heights is shown in figure 4a. The discrepancy was larger than 
zero in the tree regions as expected. However, building 
boundaries also show a large response. After analyzing the raw 
data, we found the following explanations for that. When the 
laser beam hits the exposed surface it will have a footprint 
with a size in the range of 15-30 cm or more. So, if the laser 
beam hits the edge of a building then part of the beam footprint 
will be reflected from the top roof of the building and the other 
part might reach the ground. In another case, the laser beam 
might hit the side of a building which results in multiple 
returns. The high gradient response on building edges was 
utilized to filter out the two returns using equation (1) and the 
procedure is described more in figure 4. 
if gradient > threshold (1) 
then (first — last ) = 0.0 
  
  
  
  
  
  
   
akon] 
    
& 
  
  
  
Figure 3. color coded map of the first return heights 
After filtering the discrepancy map, we now conclude that the 
remaining responses occur only from objects which are not part 
of a building. So, the first step was to filter the data based on 
filtered discrepancy responses. The aim here is not to filter the 
first return height data but to use the discrepancy map to locate 
the penetrable objects in the last return height data. 
Consequently, those detected regions will be used to filter the 
last return data using the local minimum filter. The result of 
this step was significant since a considerable amount of the 
noise was removed as shown in figure 5. However, some noise 
was not cleaned since it represents the center of a dense tree 
region or its boundaries have a large gradient response. 
Therefore, a second filtering approach was introduced. 
Fi