• Physics-informed Bayesian framework for probabilistic road flood inference from sparse post-disaster observations. • Unified energy function integrating geospatial, physical, and topological priors for robust inference under data scarcity. • Probabilistic outputs enabling flexible, threshold-based risk classification for adaptive emergency decision-making. • Sequential Bayesian updating via MCMC sampling allows continuous assimilation of new data without model retraining. Extreme weather events increasingly threaten critical infrastructure, necessitating rapid and accurate assessment of transportation network damage for effective emergency response. However, such assessments are challenged by sparse and unevenly distributed post-disaster observation data. While data-driven and physics-based approaches offer possible solutions, they often face difficulties in integrating diverse domain knowledge and generalizing effectively under data-scarce conditions. This study proposes a Markov Random Field (MRF) framework to infer probabilistic flood inundation states from sparse observations. The model integrates geospatial information on elevation, physical principles governing hydrological flows, and topological characteristics of intersection vulnerability within a unified energy-based framework. Evaluation on Hurricane Harvey’s impact to Houston’s highway system indicates that the MRF model achieves consistently high predictive performance across varying levels of data sparsity. Specifically, with 40% observational data, the integration of hydrological and topological priors yielded an Intersection over Union (IoU) of 0.747, significantly surpassing the baseline elevation-only model (IoU = 0.517). The model also supports sequential updating of state estimates and parameters via Markov chain Monte Carlo sampling, allowing continuous refinement of inference results as new observational data becomes available. Furthermore, the probabilistic outputs support flexible, threshold-based risk classification, enabling adaptive decision-making for emergency resource allocation and laying the groundwork for digital twin applications. This work offers a transferable, knowledge-aware methodology for infrastructure resilience assessment under uncertainty.