Many segmentation methods used for the homogenization of climate time series from station data use a reference series against which the station data is compared. The main advantage of this approach is to remove the common climate signal and thus improve the power of the change-point detection method. One drawback is that it is difficult to decide whether the detected change-point is due to the main station or to the reference. This paper describes a statistical method to help in this decision. It works by combining the data from the main station with the data from at least one nearby station, where the data from each station is actually composed of two series: a target series and a reference series. In our application, the target series is from daily GNSS IntegratedWater Vapour (IWV) measurements and the reference series from the Fifth ECMWF reanalysis (ERA5). Six series of differences are formed from these four base series and a statistical test is used to detect, in each of the six series, if the change in mean before and after the tested change-point is significant. Finally, a predictive rule is used to determine which of the four base series is (are) affected by change-point(s). The statistical test is based on a generalized linear regression approach, taking both heteroscedasticity and autocorrelation into account. The predictive rule is constructed on a dataset built fromthe test results obtained on the real data using a resampling strategy. Four popular machine learning methods have been considered and evaluated using cross-validation. The proposed methodwas applied to a real data set and the results looked very consistent and plausible with respect to GNSS metadataand our knowledge of the data. Results conclude that 41% of the change-points are due to GNSS data, 15% to the ERA5 data, and 25% are due to coincident detections.