Interrelationships between eutrophication predictors in Brazilian semiarid reservoirs: how to measure? A decision tree machine learning application

ABSTRACT

An emerging issue for water security is the consequences of eutrophication on water quality. Conventional regression methodologies have not been sufficient to satisfactorily explain the complexity of the relationship between the hydrological and limnological variables of this process. In this sense, this research aimed to identify predictors for eutrophication variables (cyanobacteria, chlorophyll-a, nitrogen, phosphorus, and Secchi disk measurement), through their relationships with each other and between 17 physiographic and climatic variables of the watersheds of 155 reservoirs in the Brazilian semiarid region. A machine learning method was applied with the classification and regression trees algorithm for decision trees. The results revealed that the eutrophication indicators are intrinsically related to each other, especially the concentrations of chlorophyll-a with the others. The variability of the inflow resulted in an increase in the concentration of cyanobacteria; the reduction in the volume of stored water generated an increase in the concentration of nitrogen and phosphorus; and, the drainage density generated an increase in the concentration of nitrogen. Nitrogen concentrations greater than 5 mg.L⁻¹ had significant consequences on chlorophyll-a, which was strongly associated with cyanobacteria. The volume of stored water, precipitation and the inflow to the reservoirs were also predictors of water transparency. Although the model’s performance indexes indicate wide margins of error for datasets with high coefficients of variation, decision trees can help in understanding the processes that have taken place and in planning strategic actions for water governance.

Keywords:
nitrogen; phosphorus; cyanobacteria; chlorophyll-a; classification and regression trees algorithm