A Scene Labeling Strategy for Terrain Feature Extraction Using Multisource Data

Limits of stereo processing in photogrammetry using image matching techniques are often due to very low capabilities of those procedures regarding aspects of interpretation. For example, if matching aims at recovering terrain surface it would be very helpful to know whether the processed area belongs to an urban area, agricultural land, forest region or other thematic object classes. This is one of our motivations to do research in the area of scene labeling. On the other hand the number of imaging sensors increases which record stereo and multispectral chanels simultaneously. This naturally demands for integration of multispectral image analysis and stereo image processing techniques. Drawbacks of classical pixel-by-pixel classiication for terrain feature extraction using multispectral data are well known. There is salt-and-pepper-noise caused by mixed pixels or great variations of single pixels within each object class. Further, objects having similar spectral characteristics can not be discriminated very well. Most of those experiences result from processing satellite images, e.g. with 10 m or 30 m pixel resolution. The addressed problems become more dramatic if high resolution images with pixel sizes, for example, of 0:5 m are used. To overcome some of these problems we develop a strategic concept for scene labeling based on the idea of exploiting multispectral data together with stereo and textural information. Because the work in this eld is highly diverse we will review in detail classiication aspects relating to the data and the methods which utilize diierent data sources. The conceptual work is accompanied by experimental investigations using multispectral data in conjunction with height information. First results demonstrate the achieved improvements with respect to classical multispectral classiication.