Adaptive Dynamic Bayesian Fusion for Multisource River Surface Velocity Estimation

Operational river monitoring increasingly combines video and radar, yet conventional fusion methods seldom accommodate source-specific scaling, correlated errors, outliers, and changing sensor availability. We developed an adaptive dynamic Bayesian fusion model (A-DRBF) combining affine calibration, correlated Student-t errors, dynamic state estimation, and a sensor-inclusion mask. Three video sources (June-December 2023) were evaluated by leave-one-month-out validation; radar was tested in bidirectional July–October comparisons. Across all predictions, A-DRBF yielded the lowest RMSE (0.014 m/s), an MAE of 0.011 m/s, and an 89.1% empirical coverage of a training-calibrated 90% prediction interval. Relative to the three strongest comparators, A-DRBF produced RMSE differences of 0.1–2.5%, with substantially overlapping bootstrap intervals. Radar increased July RMSE from 0.013934 to 0.014315 m/s but reduced October RMSE from 0.015035 to 0.014286 m/s, indicating that its incremental value depends on the hydrodynamic and error regime. Removing affine calibration increased RMSE to 0.165 m/s, while equal-precision weighting increased it to 0.044 m/s. In simulations, A-DRBF reduced RMSE by 45.0% under random outliers and 23.5% under compound disturbance relative to Gaussian dynamics. The model therefore provides accurate estimates and calibrated uncertainty under changing sensor configurations, with its main advantage emerging under contaminated or incomplete observations.