It is necessary to obtain reliable estimates of the in situ stress state for the design of any underground engineering project in rock, but it is of paramount importance for safety-critical projects such as deep geological repositories for nuclear waste. It is widely considered that in situ stress is a function of depth below ground surface. This often leads to rock masses being partitioned into depth-based domains, but there are no universally agreed and statistically robust methods for doing so. In this paper we present a novel method that uses Bayesian linear segmented regression of Cartesian stress components to probabilistically characterize the variability and uncertainty in the depth of non-crisp stress domain boundaries, and the in situ stress state within each domain. We demonstrate the efficacy of the method using synthetically generated stress data, and then apply the method to overcoring stress measurements obtained at the Forsmark site in Sweden.