Water quality in Minas Gerais, Brazil: Evaluating the past 25 years using ensemble decision trees and robust trend analysis

Danieli Mara Ferreira¹Rosane Barbosa Lopes Cavalcante²Gabriel Negreiros Salomão²Paulo Rógenes Monteiro Pontes^2*

• ¹Sistema de Tecnologia e Monitoramento Ambiental do Paraná, Curitiba, PR, Brazil
• ²Instituto Tecnologico Vale Desenvolvimento Sustentavel, Belém, Brazil

Water quality monitoring provides essential insights into the health and safety of water resources in a watershed. This study presents a comprehensive analysis of water quality spatial and temporal trends in the rivers of Minas Gerais, Brazil, from 1997 to 2022. For this aim, we use 258,233 samples from 675 water quality stations monitored by the Minas Gerais Institute for Water Management (IGAM). The study evaluates the risk of exceeding the established limits for class 2, as defined by a national guideline (CONAMA 357/2005). The analysis includes water quality parameters representing organic matter, nutrients, and metals related to agriculture runoff, urban and mining activities, and vegetation cover. The spatial-temporal changes in water quality are evaluated using exploratory data analysis techniques the machine learning Extra Tree regressor method, and the Theil-Sen non-parametric trend estimator. As an example, the Extra Trees regressor provided a reliable adjustment for total arsenic, yielding a mean absolute error of 0.002 mg/L. The results indicate that, while median concentrations have declined over the 25-year period, exceedance frequencies remain substantial for Mn, Fe, and TP. The results also indicate a higher risk of limit transgressions during the rainy season, underlining the importance of controlling diffuse sources and understanding hydrological processes. Using surrogate monthly mean flow, the Extra-Trees regressor ranked flow as the most important predictor among the tested variables, followed by urban infrastructure and areas with high metal content. The role of forest cover in reducing the risk of transgressions is also emphasized. In this sense, the study provides valuable insights to support decision-making for pollution control and remediation efforts to guarantee water quality safety. This study uniquely combines robust, non-linear statistical modeling with a 25-year water quality dataset in Minas Gerais, offering new insights into long-term environmental changes in a socially and economically important region.