Flood risk quantification in large-scale watersheds poses a persistent challenge due to the nonlinear interactions among geological, hydrometeorological, and anthropogenic variables. Conventional deterministic models often lack the flexibility to incorporate uncertainty and interdependencies inherent in such systems. This study introduces a Bayesian probabilistic modeling framework to evaluate flood occurrence probability by systematically integrating multi-source data and expert knowledge. A Bayesian network (BN) was constructed to capture the conditional dependencies among key flood-driving variables, including antecedent soil permeability, rainfall intensity-duration-frequency (IDF), land use/land cover, slope and drainage density. The structure of the BN was informed by hydrological process understanding and refined using a combination of mutual information and structure learning algorithms. The model was applied to the Ottawa River watershed in eastern Canada, a region frequently impacted by spring floods due to snowmelt, rainfall interactions. Historical flood events, streamflow records, and spatially distributed physiographic data were used to calibrate and validate the model. Sensitivity analysis demonstrated that antecedent permeability and extreme precipitation indices exhibited the highest influence on flood occurrence. The probabilistic output of the BN allowed for scenario-based flood likelihood assessments, highlighting critical thresholds in contributing variables. Compared to conventional empirical models, the proposed Bayesian approach reduced dimensionality, explicitly quantified uncertainty, and provided probabilistic predictions aligned with observed patterns. The framework is particularly well-suited to data-sparse regions and supports risk-informed floodplain management under a changing climate.