Triple Collocation Analysis of GNSS, MWR, and Radiosonde Water Vapor Retrievals Across Dry and Wet Seasons: A 2025–26 Analysis at Nicosia, Cyprus

Integrated Water Vapor (IWV) is a key indicator of atmospheric moisture, and its accurate quantification requires precise characterization of the errors inherent to each observing system, a task complicated by the absence of a true, error-free atmospheric reference. This study characterizes the absolute error structures of IWV retrievals from Global Navigation Satellite Systems (GNSS), Microwave Radiometers (MWR), and Radiosonde using a temporally collocated dataset of 326 trivariate samples acquired in Nicosia, Cyprus over a complete annual cycle (March 2025–March 2026). Triple Collocation Analysis (TCA) and the Three-Cornered Hat (TCH) method were applied without assuming a ground truth, with non-parametric bootstrap resampling (10,000 iterations) used to derive 95% confidence intervals and flag statistically unstable estimates. During the dry season, the recovered error ordering was radiosonde (0.38 kg m−2) below MWR (0.60 kg m−2) and GNSS (1.69 kg m−2), which is consistent with the a priori budget. The MWR error rose sharply to 5.79 kg m−2 in the wet season and to 8.91 kg m−2 in the highest moisture bin, consistent with degradation of the K-band retrieval in the presence of liquid water. The wet season value is a lower bound, since any positive error covariance between the radiometer and the radiosonde acts to increase it, whereas the GNSS error approximately doubled to 3.53 kg m−2, indicating comparatively greater resilience across weather regimes. The radiosonde error, lowest of the three sensors in dry conditions (0.38 kg m−2), could not be reliably resolved in the wet season (98.8% truncation), the lowest moisture bin (71.4%), or the highest moisture bin (94.5%), and it is reported as unstable in each. These results indicate a substantial degradation in radiometric retrieval accuracy and identify the specific regimes in which trivariate error decomposition itself becomes statistically unreliable.