In: Wagner W., Székely. B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII Part 7B 
After this, we used ground truth data for validation of the whole 
test site. The validation strategy for verification of the 
performance of double-threshold strategy includes two parts, 
i.e., double- threshold vs. single-threshold. The error matrix for 
the single threshold is shown in Table 13. The overall accuracy 
of the detection is 0.931. The producer’s accuracy is 0.874, and 
the user’s accuracy is 0.779. The error matrix for the double 
threshold strategy is shown in Table 14. The overall accuracy of 
the detection is 0.959. The producer’s accuracy is 0.937 and the 
user’s accuracy is 0.852. The accuracy shows improvement with 
the double-threshold strategy. 
To scrutinize the performance of the proposed method, two 
representative cases are discussed. These two cases explain why 
the detection failed. For the second part of the discussion, we 
look at two incorrect detections. The aerial images, LIDAR data 
and building models for case (a) and (b) are illustrated in Fig 15. 
Fig 15. Incorrect detections case (a) and (b) 
We observe that the ground truth data show no change for the 
two buildings in case (a) and (b). However, they have been 
classified as “changed”. Explanations are given as follows. In 
case (a), it is an unchanged building that has been classified to 
micro-structure changes. The reason is that some of the LIDAR 
points on the wall were not excluded. Those points cause the 
incorrect detection. As shown in Fig 16, the blue points are the 
LIDAR points within the building polygons, the green points 
are the points removed after Delaunay triangulation. The red 
points are the changed points. Notice that the detection changed 
points are almost the points on the wall that should be excluded. 
Fig 16. Changed points in case (a) 
In case (b), it is an unchanged building that has been classified 
to main-structure changes. The reason is that the building roof 
has tiny roof variations. The variations cause some of the points 
detected as change points. Those points affect the detection. As 
shown in Fig 17, the blue points are the LIDAR points within 
the building polygons, the green points are the points removed 
after Delaunay triangulation. The red points are the changed 
points. Fig 17 shows the building roof has tiny variations. 
Fig 17. Changed points in case (b) 
As shown in Fig 18, the detection results show the LIDAR 
point clouds in the new building areas by removing vegetation, 
ground and old building areas. As shown in Fig 18, these 
LIDAR points are discrete points, not regions. So, we use the 
region growing to separate the LIDAR point clouds into 
different groups. After that, we removed the wall points and 
point groups with small area of the LIDAR point groups. The 
result is shown in Fig 19. Finally, we use the boundary tracing 
to get the boundaries of new building area. The result is shown 
in Fig 20. The accuracy of new building detections is 100%. 
Nine new buildings in this test dataset are all detected by the 
proposed method. However, more test cases would be needed 
for comprehensive understanding.