Segmentation is meaningful when it is carried out with a certain intention. Aiming at the knowledge based image analysis, three methods for region-oriented segmentation of multispectral images are presented in this paper: hierarchical aggregation clustering, region growing, and quadtree. Their satisfactory performances are illustrated by several examples, where the quality of the results are visually judges by their superpositions with their original input images. All three methods can achieve good segmentations, while the hierarchical aggregation generally tends to be more suitable for images with clear object boundaries. In comparison with two other ones, the hierarchical aggregation runs very slowly but it has an advantage that an initial segmentation can always be integrated into its process. In the following knowledge based processing, segmented regions are described with attributes and modeled with knowledge base. There the possibility should exist to be able to detect and correct segmentation errors, which is usually called backtracking for resegmentation. The quadtree method may provide the possibility to do it, owing to its inherent data structure. However, the concrete implementation should be studied in the future.