Common-mode variability and the limits of rainfall-only prediction in short-term PPP-GNSS vertical coordinates at four CORS stations in Western Java

Abstract Short-term vertical coordinates from daily static Precise Point Positioning (PPP) may contain environmental signals, but they are also sensitive to processing choices, station effects, and network-wide errors. This study re-evaluates whether local antecedent rainfall provides out-of-sample predictive information for PPP-derived vertical-coordinate variability at four Continuously Operating Reference Stations (BAKO, CJKT, CLBG, and CSBG) in western Java during 1 May – 31 July 2025. Of 352 unique station-days, 310 were classified as PRIMARY_PASS, 13 as SENSITIVITY_ONLY, and 29 as EXCLUDE. The audited RTKLIB 2.4.3 workflow used forward GPS-only dual-frequency static PPP, an ionosphere-free combination, a 15° elevation mask, estimated zenith wet delay and horizontal gradients, IGS final orbit and clock products, no ambiguity resolution, and disabled tidal corrections. Unsmoothed daily solutions and formal vertical SDU were inspected directly. RINEX headers provided approximate station locations and equipment metadata. Five retained BAKO sessions with inconsistent initial XYZ entries were reprocessed using the dominant header coordinate while preserving the original navigation, orbit, clock, and configuration inputs; their heights, SDU values, and epoch-status counts were identical at the reported output precision. Rainfall missing values were preserved as unmeasured, station-specific anomalies were referenced to May–June training means, and July was retained as an independent holdout. Rainfall-only ordinary least-squares (OLS) models using rainfall at t−1 and antecedent 3- and 7-day features did not outperform station-mean baselines; MSE skill ranged from −83.0 % to −2.7 %. A leave-one-station-out common mode fitted only in May–June reduced July RMSE from 21.50 to 14.38 mm and bias from 14.97 to 1.78 mm. After common-mode variability and station-specific residual offsets were controlled, pure incremental rainfall skill was +0.07 % for t−1 rainfall and negative for all longer antecedent features. The results do not support rainfall-induced deformation or rainfall-only predictability. Within the tested three-month design, shared network variability was the dominant predictable component, while physical attribution requires longer records, official station logs, complete geodetic corrections, and independent hydrological or loading observations.