|
The emphasis of this paper is on an improvement of the classification procedure for change detection purposes using remotely sensed imagery as well as various data sources combined in a GIS. Our contributions will be on an advanced knowledge-based methodology for the training phase within the change detection process, as well as on the evaluation of the applied software package (Erdas Imagine Expert Classifier) and of the used image data source, namely Landsat TM 7 data. Firstly, we will demonstrate an improved and flexible methodology for defining and describing training areas in the course of a change detection process using knowledge-based image analysis techniques. Using the GIS-database which comprises several data sources at point of time t0 the outlines of the desired object classes will be determined and rated according to their accuracy respectively reliability. Combining this information with the image data of the first phase (t1), we are entering the first training stage. Here, not only a single standard object signature like a multispectral reflectance but a large amount of additional parameters are checked. For each parameter the inference mechanism automatically checks the separability for different object classes and evaluates the suitability of each signature for further use within the first classification stage.
The reliability of an object is derived from all information stored in the initial database. As an output of the classification stage -- again applying knowledge-based rules -- we obtain probability vectors which decide in favor of a confirmation, a modification or an elimination of the given outlines for the specific class. The new outlines together with the imagery of the next acquisition phase and up-dated ancillary data are put into the next training phase. Due to possibly changed image properties it is meaningful to test all signatures for the given outlines again, and proceed as described above. The implementation of this knowledge-based approach is performed by means of the "Expert Classifier" from Erdas Imagine (version 8.6). In conclusion it can be stated that the proposed knowledge-based method and its implementation has been proved to be a very valuable and reliable method for environmental change detection purposes.
|
