Relevant material and environmental parameters are required in modelling chloride ingress into concrete. They could be determined from experimental data (concrete cores taken during inspection) but in practice data availability is limited by time-consuming and expensive tests. Consequently, the main objective of this paper is to develop an approach based on Bayesian networks (BN) to improve the parameter identification when inspection data is limited. We aim at proposing appropriate inspection configurations that reduce inspection costs and identification errors for different exposure conditions and materials. It was found that it is possible to define an optimal number of inspection points in depth for allowed identification errors defined by decision makers. The optimal number of inspection points depends on both exposure and material properties. The random variables identified with the improved BN configurations are used to assess the probability of corrosion initiation. The results indicate that the improved BN configurations are useful to identify model parameters even from scarce inspection data.