The detection of non-stationarities (trends) in the hydrological behaviour of watersheds affected by environmental change has traditionally been achieved through the comparison of "control" (reference) and "modified" watersheds. These comparisons are probably the most efficient solution for trend detection, and are extensively documented in the hydrological literature. Outside experimental watersheds however, control watersheds are seldom available, and another approach is needed to assess the evolution of watershed behaviour. In this paper, we present a methodology using a parsimonious 4-parameter rainfall-runoff model (GR4J) to detect non-stationarities. The parsimony of the model makes it relatively easy to identify stable representative parameter sets over short time periods, and to quantify the calibration uncertainty for these parameters. Using this uncertainty knowledge, we generate equi-probable parameter quadruplets for successive periods of time, from which we derive through simulation a distribution of a hydrological variable (e.g. total runoff), representative of the watershed behaviour during this period. We then propose a non-parametric statistical test to identify non-stationarities from the distributions, and we validate this test on a deforested experimental watershed.