Enhancing the modeling of dissolved oxygen concentration using machine learning, a case study in the Greek Mediterranean Sea.

In recent years, artificial intelligence (AI) tools have gained popularity as forecasting and predictive tools able to approximate with high accuracy trends and outcomes in many fields such as robotics, climatology, and hydrology, including water resources. AI models have shown remarkable performances handling big data and dealing with their nonlinearity and nonstationary features in monitoring and forecasting water quality, complementing the traditional numerical water quality models that provide precise parametrizations of near-shore and off-shore processes and their complex interactions.

In this research, we examine the accuracy of different machine learning techniques in estimating and predicting dissolved oxygen concentration (DO) in water bodies. DO is a crucial water quality variable that influences the living conditions of all aquatic organisms requiring oxygen. Low DO concentration, when persistent, can cause eutrophic conditions, thus altering the normal nutrient cycle, favoring the formation of algal blooms and furtherly reducing water quality and affecting the entire ecosystems, also causing fish mortality.

The Random Forest (RF) and the generalized regression neural network (GRNN) are explored and compared. The two models are developed using high frequency in situ data collected from Andromeda Group, a leading company in the aquaculture sector in Greece, at four different stations in the Greek Mediterranean Sea. The input variables used for the two models are temperature and currents. The performances of the models are evaluated using root mean square errors (RMSE) and mean absolute error (MAE). The RF and GRNN showed similar performances, with the best fit obtained using the GRNN model. Results are also compared with a traditional numerical model developed with the DELFT3D-WAQ modeling suite. The AI models show better performances in estimating daily changes of the DO concentration and by being less computationally expensive than the numerical model, enhance the water quality monitoring and provide aquaculture and farmers managers with a forecasting tool.

Acknowledgments:

The work has been conducted within the framework of the HiSea project, which has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No 821934 and of the Water Harmony project, an ERA-NET WATERJPI Co-fund Action. Funding received via the Dutch Research Council - NWO Project number ENWWW.2018.1