Bayesian Probabilistic Modeling applied to some environmental data using PyMC

Machine learning models can be employed to antecipate environmental problems and even overseesystems and processes. The Bayesian multivariate probabilistic model (BMP) is able to predict the concentrationof Dissolved Oxygen (DO) in water. With the use of free software which, based on historicaldata from water analysis, we are able to achieve results similar to those produced by artificial neuralnetworks. In this study, we applied a BMP model developed in Python using the PyMC3 library to two differentnonlinear environmental data sets. In the first one, a multivariate calibration of DO was performedin raw water from the Piracicaba River (São Paulo - Brazil), and the BMP model obtained a Root MeanSquare Error of prediction (RMSEp) of 0.835 mg L-1 for 20 samples tested. In the second set, the samemodeling process was used to obtain an RMSEp of 0.839 mg L-1 for 20 samples tested. Both resultscomparable to those obtained with a back propagation neural network. Another non-linear environmentaldata set was tested with excellent results. A BMP model for the concentration of benzene in air, asused for monitoring air pollution, obtained a RMSEp of 0.584 μg m-3.