In this paper, to address the problems of missing measurement noise and process noise covariance matrix (Q-R matrix) coupling modeling, insufficient generalization ability of nonlinear mutation data, and high computational complexity in the dynamic noise covariance estimation of GNSS in the pro-slip-slip phase of landslide disaster, we propose a trace-free Kalman filter optimization that integrates liquid neural networks (LNNs) and variational Bayesian inference (VBM) framework (LNNs-VBM-UKF). The framework utilizes LNNs to dynamically generate a priori estimates of noise covariance, constructs a joint optimization mechanism based on VBM, reduces the Q-R estimation complexity from O(n3) to O(n2), and realizes the adaptive closed-loop feedback regulation of the state variables and noise covariance. The experiments are carried out on the example of the GNSS monitoring data of Pijiashan Shunyi Slip in Huangshi City, Hubei Province, in August 2022, and the results show that this method improves the efficiency of noise suppression in all directions, reduces the estimation error, and effectively solves the problem of filter dispersion caused by the assumption of fixed value of Q-R in the case of nonlinear mutation of the surface displacement solved data. Meanwhile, this study establishes for the first time the mapping relationship between the dynamic characteristics of landslide noise and continuous state space, which provides an interpretable co-optimization framework for geohazard monitoring.