The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008 
727 
to polygon conversion. In the second part, a supervised 
classification technique is used to classify the AVIRIS image 
after removing building pixels from the image. A number of 
image processing techniques including thinning, edge detection, 
and line fitting are performed on all pixels classified as road 
and water to generate two vector layers for the road network 
and the shoreline. 
5. BUILDING VECTOR LAYER 
In order to find buildings cues in the LIDAR-based DEM, the 
digital surface model (DSM), i.e. representing bare ground, 
need to be extracted. Minimum filters are used to perform this 
task (Masaharu and Ohtsubo, 2002; Wack and Wimmer, 2002). 
The main objective of the filtering process is to detect and 
consequently remove points above the ground surface in order 
to recognize height DSM points in the data set. The minimum 
filter size should be large enough to include data points that are 
not part of the noise. However, iterative approaches could be 
used to avoid the effect of noise. In this research, the size of the 
filter is 3x3. The filtering is repeated iteratively until the DSM 
is extracted. If the difference between the DSM and the DEM 
for any pixel is greater than a given threshold, the point is 
treated as a building pixel. The value of the threshold is 
determinate using previous knowledge about the area. 
The next step is to eliminate extraneous features such as trees 
and any other object above the ground that are not buildings. A 
local statistic analysis (Maas, 1999) is used for this purpose. 
The process is implemented as follows. Each group of pixels 
that lies within a small square window is fitted to a plane. Then, 
a least squares adjustment algorithm is used to obtain the plane 
parameters. After the adjustment procedure, the RMS is 
computed for each window. A high RMS indicates an irregular 
surface that can be interpreted as a characteristic of a tree or a 
rough surface, since most buildings have smooth roof surfaces. 
For each pixel, the algorithm is implemented with different 
window orientations and positions and the minimum RMS is 
reported. 
A split and merge image segmentation process is then used to 
segment the LIDAR based DEM to regions after removing the 
extraneous objects from the raw data. The split and merge 
image segmentation technique, Horowitz and Pavlidis (1974), is 
implemented as follows. First a quad-tree representation is 
constructed for the image splitting as necessary when in 
homogeneities exist. Then adjacent regions are merged to form 
larger regions based on a similarity criterion. In the last step, 
small regions are either eliminated or merged with larger 
regions and holes are removed. Border points for each region 
are extracted and used to fill a 2D Hough transformation 
parameter space, (Hough, 1962). The Hough transformation 
parameter space is then searched and analyzed to find all 
borderlines. For each cell in the parameter space a non-linear 
lest squares estimation model is employed to refine the 
borderline parameters using all border points contributing to the 
cell. 
The next step is to convert the extracted lines to polygons using 
a rule-based system. The rules are designed as complex as 
possible to cover a wide range of polygons. The mechanism that 
is developed here works in three steps. The first step is to find 
all possible intersections between all borderlines. The next step 
is to generate all feasible polygons from all recorded 
intersections. Each combination of three to six intersection 
points is considered to be a polygon hypothesis. Some 
hypotheses are ignored if the difference in the area between the 
region and the hypothesized polygon is more than 50%. The 
third step is to find the optimal polygon that represents the 
region borders. This polygon is chosen from the remaining 
polygons using a template matching technique. The template is 
chosen to be the original region, while it is matched across all 
polygon hypotheses. The hypothesis with the largest correlation 
and minimum number of vertices is chosen to be the best fitting 
polygon. Figure 2, shows the extracted building polygons 
overlaid on the one-meter LIDAR-based DEM. 
Figure 2. Extracted building polygons overlaid on the one- 
meter LIDAR-based DEM, (rotated) 
6. ROAD NETWORK AND SHORELINE VECTOR 
LAYERS 
In this section the process of generating vector layers for the 
road network and the shoreline is presented. The results of 
classifying the raw AVIRIS image showed an overall accuracy 
of about 68.7%. This is due to the diversity of the building roof 
materials and the small number of buildings. Several buildings 
were classified as roads. This motivated the elimination of the 
building pixels before the classification process. Hence, the 
following algorithm was applied. Building polygons were first 
used to define building pixels in the AVIRIS hyperspectral 
image using a point-in-polygon process. All pixels defined as 
building points are identified as background pixels and not used 
in the classification process. In the next step, a supervised 
classification technique is performed on the AVIRIS 
hyperspectral image using the extraction and classification of 
homogeneous objects (ECHO) classifier, (Biehl and Landgrebe, 
2002). 
Six classes are defined in the image; road, water, sand beach, 
grass, bare soil, others. The last class includes all objects that 
cannot be identified. Training pixels are then identified and 
used to calculate each class statistics. A number of test samples 
are used to evaluate the classification results of each class. 
Table 1 shows the size of the training and test samples for each 
class and the classification results for the training and test 
samples. The total accuracy of all training sites is 92.7%, while 
the total accuracy of all test sites is 87.6%. However, the 
average accuracy of the road and water test sites, that are used 
to generate the road network and shoreline layers, is about 98%.