To study urban expansion in three dimensions, digital surface models need to be extracted of the study area at regular time intervals. The Istanbul urban area is chosen as test field.
A big problem that arises, when creating 3D city model time series before 1999 is the lack of (high resolution) sensors that take images stereoscopically. Therefore there is chosen to construct across-track stereo pairs out of mono-scopic SPOT images. Imagery of the SPOT sensor was preferred because of the optimal spatial resolution, geometric robustness of the imagery and the big availability of images in the archive. Out of the extensive image archive 8 SPOT scenes were carefully selected to construct an across-track stereo pair for the periods 1986, 1990, 1994 and 1998. All images are panchromatic, level 1B with a spatial resolution of 10m.

One of the main challenges of WP3 is to cope with the specific problems inherent to across-track stereo pairs. Because the images are taken from a different orbital pass of the satellite at a different date, across-track pairs have some major restrictions. Different illumination conditions, ground surface changes and different image geometry leads to dissimilarities between the images and to mismatching during the photogrammetric processing.
First of all, stereo are were carefully selected from the archive. Area coverage, stereo constellation, cloud-freeness and image acquisition time interval are chosen as main criteria for the selection. To reduce radiometric dissimilarities between the images, a radiometric enhancement is performed. To enhance the contrast for each image individually and to equalize the radiometric differences between the imagery, a Wallis filter is applied. The Wallis filter performs a non-linear, locally adaptive contrast enhancement, resulting in good local contrast while reducing the overall contrast between bright and dark areas.
Next to radiometric enhancement, a method for geometric normalization was devised. As the images of each across-track stereo pair are taken from different orbits, the images are displaced to each other and the internal geometry will be slightly different because of the different scan direction. A first-order polynomial transformation is performed to geometrically align the multi-temporal imagery. A first-order polynomial transformation corrects for rotation, translation, scaling and shearing.
Based on a dataset of 40 ground control points homogeneously distributed over the study area, photogrammetric products are derived from the four across-track pairs with the photogrammetric software package SAT-PP.                                

The surface models and ortho-images are processed at a grid size of 1 pixel or 10 meters. All models have a geometric accuracy in sub-pixel range for x and y and between 1 and 2 pixels for Z residuals. Outliners and noise due to clouds, ground surface changes and other errors are detected and removed from the models. The remaining gaps are interpolated with the kriging method. Future work will focus on the comparison of the multi temporal surface models with each other to detect changes and possible urban growth.