Sensitivity of discrete Bayesian networks to the number of intervals in fire forecasting

Wildfires are a global problem, causing environmental, material, and human losses.
Predicting the frequency of fires is fundamental for the State to prevent and intervene against these events. Bayesian networks (BNs) can be promising in this way. However, analytical approaches are easily unfeasible and simulation can become computationally expensive. In turn, exact inference is yet possible when the variables are categorical, making emerge questions about the better way to categorize quantitative variables. To overcome this challenge, this article seeks to assess the sensitivity of Discrete BNs to the number of intervals (or categories) used in the discretization. A set of monthly fire frequency time series from Cear´a was used as a case study. It was found that the networks were highly sensitive to the number of intervals used in the discretization. For some series, BNs had their performance improved by a factor of 20 depending on the number of intervals.