Abstract: Over-segmentation is a common problem for all kinds of segmentation tasks. Automated segmentation of natural scenes is no exception. This paper proposes a solution to the over-segmentation problem, with the emphasis on satellite images of farmland. In many cases, an agricultural field can be considered as a flat region having a rather large area, a compact shape, and straight region boundaries because it is a man-made object. Our approach for dividing farmland into individual field units use region shape, as well as spectral information, when merging over-segmented regions. The results from the presented method are compared to two different methods of segmentation as well as interpreted field boundaries. The results show that task-specific knowledge adds important information to the decision step for the merging procedure of regions. About 70% of the edges are classified within one pixel away from the ground truth edges using our methods.