In this article we present a hierarchical stochastic image segmentation approach. This approach is based on a framework of edge-weighted graph for minimum spanning forest hierarchy. Image regions, that are represented by trees in a forest, can be merged according to a certain rule in order to create a single tree that represents segments hierarchically. In this article, we propose to add a uniform random noise into the edge-weighted graph and then we build the hierarchy with several realizations of independent segmentations. At the end, we combine all the hierarchical segmentations into a single one. As we show in this article, adding noise into the edge weights improves the segmentation precision of larger image regions and for F-Measure of objects and parts.