Multi-class random forest model to classify wastewater treatment imbalanced data

The odor emissions generated by treatment plants imply complex environmental and economic issues. The modern instrumental odor monitoring systems based on an array of several sensors, continuously record the gaseous compounds. However they are characterized by poor selectivity, compromising the possibility to discriminate and identify the emission sources. In this paper, the ability of odor sensors to distinguish between the treatment plant sections generating the gaseous compounds is evaluated on the basis of the random forest classifier, and is also compared to the discriminant analysis performance. Taking into account that a multi-parametric system of sensors can be affected by the presence of a small sample size with imbalanced classes, several strategies for data balancing are proposed and analyzed. The findings show that the random forest classifier is characterized by a better capacity to distinguish the emissions sources with respect to the classical multiple discriminant analysis, in terms of all evaluation metrics. This is also confirmed for different resampling techniques, especially in the over-sampling case. The data concerning measurements from 10 sensors of a multi-parametric system of odor monitoring collected from a company specialized in environmental assistance are considered for this analysis.