Bayesian Networks for Preprocessing Water Management Data

Environmental data often present inconveniences that make modeling tasks difficult. During the phase of data collection, two problems were found: (i) a block of five months of data was unavailable, and (ii) no information was collected from the coastal area, which made flood-risk estimation difficult. Thus, our aim is to explore and provide possible solutions to both issues. To avoid removing a variable (or those missing months), the proposed solution is a BN-based regression model using fixed probabilistic graphical structures to impute the missing variable as accurately as possible. For the second problem, the lack of information, an unsupervised classification method based on BN was developed to predict flood risk in the coastal area. Results showed that the proposed regression solution could predict the behavior of the continuous missing variable, avoiding the initial drawback of rejecting it. Moreover, the unsupervised classifier could classify all observations into a set of groups according to upstream river behavior and rainfall information, and return the probability of belonging to each group, providing appropriate predictions about the risk of flood in the coastal area.