An operational wind-speed-dependent quadratic bias correction for Voluntary Observing Ship wind measurements evaluated against moored buoys and ERA5 reanalysis

Abstract

Voluntary Observing Ship (VOS) wind reports constitute one of the densest marine in situ datasets, yet they carry a persistent, wind-speed-dependent negative bias that intensifies at low wind speeds. Using 18,957 collocated ship–buoy–ERA5 records over the nearshore waters of China, we diagnose this bias against two complementary references, a moored-buoy network and ERA5 reanalysis, and construct a correction that is deployable in practice: a three-parameter quadratic in the observed VOS wind speed itself, requiring no reference information at application time. On held-out data the correction reduces RMSE from 0.65 to 0.41 m s −1 against buoys (37%) and by 41% against ERA5, removes nearly all of the −0.42 m s −1 mean bias, and restores the ship-to-buoy variance ratio from 1.25 to 1.02. The diagnosed bias is robust to the collocation criteria: it persists, and in fact strengthens, when the matching radius is tightened from 100 to 30 km, and is insensitive to temporal separation within 30 minutes. Because consecutive ship reports are strongly autocorrelated, generalization is assessed with day-blocked and month-blocked cross-validation rather than random splitting; the improvement survives both (36% and 37–44% outside September, respectively). A triple-collocation check attributes comparable random errors to the ship and buoy epochs, while the bias curves obtained from the two references agree to within 0.1 m s −1 over 3–12 m s −1 , consistent with a platform-origin systematic error. The correction is compact enough for operational VOS quality control.