Automated offset detection approaches: Case study in IGS Repro2 and 3

Abstract Due to their high false positive detection rates, traditional automated offset detections on GNSS station coordinate time series still cannot fully replace manual detections in practical applications and need to be improved. This work proposes preliminary and enhanced automated offset detection approaches for IGS Repro2 and Repro3 datasets. Their manually detected offset recordings in ITRF2014 and ITRF2020 are used as evaluation criteria. In the preliminary approaches, stochastic models based on covariance matrix, white noise model, and white noise plus flicker noise model of both univariate and multivariate are studied. Although a somewhat “good” performance is achieved, the preliminary automated detections still lead to many false positive detections. Thus, based on the preliminary approaches, and ancillary data, an enhanced detection approach is proposed. Enhanced detections significantly reduce 56%~80% false positive detections compared to preliminary approaches. As a result, for Repro3, the optimal overall performance is attained with true positive, false positive, and false negative rates (TP, FP, FN) of (0.57, 0.25, 0.18), along with a detection rate of 75%; for Repro2, the rates are (0.58, 0.20, 0.22), accompanied by a 73% detection rate. According to its “good” performance mentioned above, the current enhanced approach may serve as a supplementary or reference to manual detection, although still not being perfect. Furthermore, twenty manually detected unknown offsets in ITRF2020 are found to correspond to some known events (thirteen earthquakes and seven equipment changes); thirty-four automated detections that correspond to known events but are not collected in ITRF2020 are manually checked as offsets (fourteen earthquakes and twenty equipment changes).