In segmentation, an image is divided into homogeneous regions in which the observations are similar to each other. Ideally, a segment in an image corresponds to a forest stand homogeneous with respect to forest characteristics. When individual pixels yield poor estimates of stem volume, the mean of observations within a segment may provide more accurate estimates for small areas. We have experimented with segmentation methods incorporating initial classification of pixels. A path-connected set of pixels classified into the same class is then regarded as a segment. Our segmentations were based on six SAR images, that is, a six-dimensional observation vector was associated with each pixel. A logarithmic transformation of observations was necessary. To use National Forest Inventory (NFI) sample plots as a training set, we grouped the plots into classes using the development class and the dominant tree species of each plot. Each class was characterized by the mean of observations in the SAR images. The pixels of a test area were classified into the classes by the ICM algorithm. It yields satisfactory segmentations, but the results can be improved by more time-demanding methods. After initialization by ICM, the segmentation was continued by adaptive clustering. In this method, the class means depend on the location of a pixel, and the means are updated iteratively. At each pixel, a class mean is estimated over a small window by the average of observations currently classified into the class. The estimation and classification alternate until convergence. Finally, it was worth eliminating the smallest segments by a merging algorithm. A randomly chosen segment was merged with such a neighboring segment that the difference between the means of observations in the segments was as small as possible. If the difference was statistically significant, the two segments remained separate.