Applying Limnological Feature-Based Machine Learning Techniques to Chemical State Classification in Marine Transitional Systems

Abstract

On a global scale, marine transitional waters have been severely impacted by anthropogenic activities. Historically, developing human civilizations have often settled in coastal areas with about 2/3 of the human population inhabiting areas within 20-km range from coastal areas. Environmental management worldwide strives for sustainable development while minimizing impacts to ecosystem integrity and has resulted in several framework directives, management programs, and legislation compelling governments to monitor their coastal systems and improve environmental quality. Among the most significant anthropogenic impacts to these ecosystems are land reclamation, dredging, pollution (sediment discharges, hazardous substances, litter, oil spills, and eutrophication), unsustainable exploitation of marine resources (sand extraction, oil and gas exploitation, and fishing), unmanaged tourism activities, the introduction of non-indigenous species, and climate change. The multitude of stressors is not independent, and as such, the chemical status of marine systems has serious implications on its ecological status and needs to be addressed efficiently. Public monitoring databases provide a large amount of physico-chemical (nutrient, dissolved oxygen, and chlorophyll a concentration) and contaminant (trace metals and polycyclic aromatic hydrocarbons) data for all Portuguese transitional systems (estuaries and coastal lagoons). These data are used to classify the chemical status (eutrophication and contamination level) of these ecosystems considering pre-defined classification thresholds, which facilitates communication to government authorities and management entities. Artificial intelligence and machine learning techniques provide an automated and efficient opportunity to improve simulation accuracy and further advance our understanding of environmental problems in estuarine and coastal waters when dealing with large environmental datasets. In the present work, we applied machine learning models, namely, linear discriminant analysis, classification tree, naive Bayesian, and support vector machine, to nutrient, dissolved oxygen, chlorophyll a, trace metals, and polycyclic aromatic hydrocarbon concentrations to produce a chemical status classification of the Portuguese marine transition systems. This approach allowed us to efficiently classify in an automated way the transitional water’s chemical status within the pre-defined classification thresholds, producing numerical index values that can easily be communicated to the general public and managers alike.