1264 
The International Archives of the Photogrammetry. Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008 
Archives 
0^ 
Vector database 
0 
I 
alities that need to be interconnected to meet the needs of a par 
ticular user. 
About 3000 C++ classes are already available in the current ver 
sion of OTB for most of the usual operations on remote sensing 
images. 
• image access: optimized read/write access for almost any of 
remote sensing image formats, meta-data access, visualiza 
tion; 
• geometric modeling: sensor models, DEM access, carto 
graphic projections, image registration, disparity map esti 
mation; 
• filtering: blurring, denoising, enhancement; 
• feature extraction: interest points, alignments, lines; 
Figure 2: Segmentation example: from three different seeds. The 
fast marching algorithms generates three different areas. 
by a factor of four). A pan-sharpening step is necessary to obtain 
an image with four spectral bands with the highest resolution. 
Several pan-sharpening methods are available in OTB. One ex 
ample is illustrated in figure 4. 
• image segmentation: region growing, fast marching, water 
shed, level sets; 
• object extraction: road network extraction, example-based 
detection; 
• classification: K-means, SVM, Markov random fields; 
Image classification from examples is a very useful task. Sup 
port Vector Machine can produce a good classification models 
from few examples (Weston and Watkins, 1998). On figure 5, 
an example of classification by SVM is illustrated. On the mul- 
tispectral image, few regions of interest are selected to train the 
SVM. Then the entire image is classified. 
• change detection. 
As we can see, the functionalities cover the whole range of im 
age processing, from access to image format to applications like 
change detection. 
Segmentation is a basic task in image processing. On figure 2 an 
example is given for the fast marching algorithm initiated from 
three different seeds directly on the luminance image. 
On figure 3, the registration between an optical and a radar image 
of the same area is illustrated. A good registration is a compul 
sory stage before being able to exploit jointly information from 
both images (Inglada and Giros, 2004). The deformation model 
is done by a centered affine transform which is able to introduce 
translation, rotation and scaling effects. The similarity metric 
cannot be a simple correlation due to the completly different ac 
quisition process between the two sensors: mutual information is 
used instead (Maes et al., 1997). 
Most current high resolution optical sensors (Spot 1 to 5, Quick- 
bird, the coming Pleiades), have a high resolution panchromatic 
band and a multispectral band with a lower resolution (typically 
One common application of satellite images is the change detec 
tion between two images, either to detect the effects of natural 
disasters or to update vector database (Poulain et al., 2008). Fig 
ure 6 presents the application on flooding on the South of England 
using SPOT images. Many other change detectors have been im 
plemented in the toolbox using statistical similarity measures, as 
for instance the one presented in (Inglada and Mercier, 2007). 
Finally, direct objects or network extraction can also be devel- 
opped. Figure 7 presents a real-time road extraction algorithm 
(Christophe and Inglada, 2007). 
All these features are available in OTB, but not all were devel- 
opped internally. The library is based on several external libraries. 
4 USING THE (GREAT) WORK OF OTHERS 
When developping a complex library, throughful validation of the 
algorithms is always a very delicate part. To be able to provide 
well tested algorithm with limited ressources, OTB is based on 
numerous, carefully chosen, open-source libraries. For each do 
main, we select the library which has a broad base of users (the 
library is well tested) and which is compatible in terms of licence