Abstract: Bayesian method infers the posterior distribution of slope parameters by combining prior distribution with filed time-series monitoring data. This process requires extensive computational resources due to repeated calls to time-consuming numerical models. While surrogate models can replace numerical models to improve efficiency, current Bayesian inversion methods still exhibit limitations. On the one hand, conventional surrogate models inadequately capture the spatiotemporal evolutionary characteristics of slope output responses, requiring separate model constructions for distinct temporal and spatial points. On the other hand, integrating time-series monitoring data requires multiple Bayesian inversions. During this process, the prior distribution progressively transitions into the posterior distributions, leading to the phenomenon of distribution shift. Employing surrogate models constructed based on a fixed prior distribution results in poor computational accuracy during parameter inversion. To address these issues, this study proposes a Bayesian inversion method that combines subset simulation with a fine-tuned deep operator network (DeepONet). Specifically, the DeepONet model is employed to construct a spatio-temporally evolutionary surrogate model. Subsequently, additional training samples are selected in each subset simulation layer to fine-tune the DeepONet model, ensuring the accuracy of posterior distribution inference. The proposed method is validated using a case study of a slope in Hong Kong. The results demonstrate that the proposed method enhances the computational efficiency of Bayesian inverse analysis while ensuring the accuracy of posterior parameter estimation. This study provides an effective tool for addressing the problem of posterior distribution inference of slope parameters based on time-series monitoring data.