Use of random forest for assessing the effect of water quality parameters on the biological status of surface waters

Abstract

The Water Framework Directive aims to reach good status in European surface waters by 2027. Despite the efforts taken already, the ecological status of surface waters has hardly improved during the last decades. In order to find efficient measures, there is an urgent need to improve our knowledge in understanding the linkage between the anthropogenic factors and the indicators of the ecological status assessment. Due to the complexity of the ecosystems, basic statistical methods (such as linear regression) cannot help in finding relationships between the biological quality elements and the supporting water chemistry parameters. The paper demonstrates that in these cases a machine learning data-driven method can be a promising tool for supporting biological classification. With random forest, the Gini index was used for ranking physico-chemical variables based on their influence on biological elements. Variables that have the biggest Gini index were selected for predicting the biological status of phytoplankton, phytobenthos and macrophytes. Binary classification and predictions were performed on a five-class scale. Predictions tended to be fairly good (errors varied within 8–60%, median 33.3%). A comparative analysis was also made with logistic regression, however, in some cases it led to slightly worse or slightly better predictions. We concluded that due to significant errors, the biological status assessment cannot be replaced completely by model predictions, but the method is sufficient to fill in certain gaps in the data and can help in the planning of biological monitoring systems. The evaluation was performed with Hungarian river and water quality database.