Machine‑learning fusion of hourly E-GVAP near‑real‑time GNSS ZTD: ERA5-referenced evaluation and uncertainty estimation

Hourly near-real-time (NRT) GNSS zenith total delay (ZTD) observations provide continuous information on tropospheric variability and are increasingly used for tropospheric monitoring. Within E‑GVAP, many analysis centres (ACs) deliver hourly NRT ZTD estimates over Europe. While this multi‑centre setup provides redundancy, analysis-to-analysis differences in processing strategies and varying data availability/latency introduce time and site-dependent inconsistencies that complicate downstream use.

We present a machine‑learning (ML) fusion framework that combines hourly NRT ZTD from E-GVAP AC streams into a single, quality-controlled consensus ZTD with an associated uncertainty estimate. The ML component is formulated as a lightweight supervised “ensemble/meta‑learner”, where each AC is treated as an expert and the model learns adaptive, station, and time-dependent weights from features derived only from the NRT streams and station metadata. Predictors include Inter AC consistency metrics (spread/robust dispersion), recent ZTD tendencies, station coordinates, and completeness (latency indicators). The ML fusion is benchmarked against robust non‑ML baselines (mean, median, and best single‑AC selection).

To avoid dependency on post‑processed tropospheric final products (e.g., IGS/CODE final ZTD), performance is assessed against ERA5 reanalysis by deriving station‑specific hourly tropospheric delays at each GNSS site, accounting for model and station height differences. Station surface pressure is used to compute the hydrostatic delay and isolate the wet delay component, enabling targeted evaluation of humidity‑driven variability. We quantify bias, dispersion, and temporal variability for individual AC solutions and for the fused product, and examine how learned weights and uncertainty respond to changing meteorological regimes and data availability. The resulting hourly and uncertainty/QC information support more reliable NRT tropospheric products for monitoring and assimilation‑oriented workflows.