Abstract: Objectives To address the difficulty of traditional Sage-Husa adaptive unscented Kalman filtering in balancing transient-response capability and the stability of weakly observable parameters, a three-layer adaptive UKF method is proposed for parameter identification of unmanned surface vehicle (USV) maneuvering dynamics models. Methods Within the standard UKF framework, the Sage-Husa mechanism is used to estimate the overall level of the process noise covariance. A strong tracking mechanism based on a sliding-window normalized innovation squared (NIS) statistic is then introduced to enhance parameter-subspace updating under transient maneuvering conditions. Furthermore, an observability-aware index is constructed from the parameter rows of the Kalman gain matrix, and the parameter process noise covariance is modulated differentially. Using the “Yukun” vessel simulation model, open-loop reconstruction, ablation comparison and generalization tests are carried out under combined excitation of variable-speed propulsion, zigzag steering and turning maneuvers. Results Compared with OA-SH-UKF, the proposed ST-OA-SH-UKF reduces the open-loop RMSEs of surge velocity, sway velocity and yaw rate by 10.42%, 53.81% and 71.35%, respectively, and reduces the RMSEs of longitudinal and transverse positions by 25.99% and 16.19%, respectively. Under a generalized condition in which the maximum rudder angle is increased from 20° to 25°, the identified model still preserves the main trajectory pattern and velocity trend. Conclusions The proposed method can balance parameter-update sensitivity and estimation stability under complex maneuvering conditions, improving the open-loop reconstruction accuracy and cross-condition adaptability of USV maneuvering models.