Reliable operation of autonomous underwater vehicles (AUV) depends on accurate state estimation. The major challenges come from sensor noise, unpredictable underwater environments, and complex vehicle dynamics. Traditional state space models assume linear dynamics to work with classical estimation methods. Kalman Filter (KF) variants and least squares rely on this assumption. These methods struggle when the system is highly nonlinear or changes over time. This limits their estimation accuracy for underwater vehicles operating in complex environments. This work overcomes these limitations by proposing a hybrid estimation framework that extends classical Kalman filtering by combining nonlinear compensation and adaptive uncertainty modeling. It uses a confidence-gated Radial Basis Function Neural Network (RBFNN) to capture unmodeled nonlinear dynamics and applies bounded corrections to the predicted state of the Extended Kalman Filter (EKF). It adapts process noise in response to changing residual patterns. The RBFNN is retrained online in a controlled manner, using residual thresholds and scheduled updates to maintain the state estimates reliable. Tests on a public AUV dataset showed reduced position RMSE by 31\% and improved trajectory tracking over the standard EKF.