An Object-wise Classification Approach

Keeping spatial data up-to-date is a very time and cost intensive task. Every object of the database has to be checked by a human operator (for example by comparing it with an up-to-date orthophoto) to see if there has been a change in the landscape. Therefore the amount of work of updating a spatial database is nearly as high as the primary acquisition. But a large number of GIS applications rely on up-to-date data in order to solve their tasks. The higher the number of objects in the database the more difficult the problem. New satellite systems offer high resolution multispectral data in high quality with high repetition rates. This data can be used as an input for automatic change detection procedures. One approach is for example to classify each pixel of an image to one of several predefined landuse classes. Afterwards, the classification result can be compared automatically with the GIS database in order to detect updates. Whereas the classification is a very well understood and manageable problem, the matching is still a difficult task. Problems arise for example by objects which are not captured according their real shape but according ownership structures or by objects which have a very inhomogeneous appearance. Nevertheless a human operator can deal with that problems and distinguish between correct and incorrect acquisition with high certainty. The reason for this is that human image interpretation is not based on the interpretation on single pixels but on whole object structures and their relations between them. In this paper an approach is introduced that classifies not only single pixels but groups of pixels which represent already existing object geometry’s in a GIS database. This object structured classification result is than compared with the existing GIS objects and all objects are marked where the result of the classification is not the same as the object class of the object which is stored in the GIS database. The result is not only a change detection but also a classification into the most likely class.