Abstract
Eutrophication episodes are commonly observed in coastal areas, causing significant damage, especially in the Mediterranean Basin which represents the major world’s tourism hotspot. For this reason, anticipate their presence is important to prevent risks. Predictive models are effective techniques for eutrophication forecasting, since ecologists are able to predict water pollution levels and take necessary precaution measures in advance. Previous studies have confirmed the superiority of Machine Learning (ML) in modeling water quality parameters. In this study, chlorophyll-a (Chl-a) is used as water quality indicator. In this thesis, the first chapter sets up a characterization, in terms of space, time and season of the physico-chemical water quality parameters in the North lagoon of Tunis, a shallow restored Mediterranean coastal ecosystem located in the north of Tunisia. The present study of seven physico-chemical parameters (Secchi depth, dissolved oxygen, pH, salinity, water temperature, total phosphorus and total nitrogen), in addition to the Chl-a has covered approximately three decades (January 1989 - April 2018), at five representative positions in the lagoon. In general, the water quality agrees fairly well with the marked improvement in the eutrophication level of the Lagoon, especially after the completion of the restoration project mentioned above. This chapter, represents the first step in predictive modeling by preprocessing the data looking for any hidden pattern (spatial, temporal or seasonal), in order to choose the most appropriate algorithm. In the second chapter. A Nonlinear Auto Regressive with eXternal inputs (NARX) neural network was fitted to predict Chl-a concentrations. The predictor variables (inputs) used in this model were those having high permutation importance ranking with Random Forest model: Secchi depth, and dissolved oxygen. Various model scenarios with different NARX architectures were tested. To verify the model performances, the trained models were applied to field monitoring data. In addition, various mathematical parameters to measure the performance of the model were calculated (R, R2, MSE). Results indicated that the developed NARX model can predict one step ahead (1 month) Chl-a concentration in the lagoon with high accuracy (R= 0.79; R2= 0.62; MSE= 0.31). In the third chapter, a Seasonal Auto-Regressive Integrated Moving Average (SARIMA) model was developed to forecast monthly Chl-a concentrations, using three decades of Chl-a historical data. Results showed SARIMA (2,0,2)(2,0,2)12 was the best fitted model. We have demonstrated that with only one input variable SARIMA model shows great applicability as an eutrophication early warning tool. Finally, the developed SARIMA model was used to forecast Chl-a concentrations for the period from May 2018 to December 2025, as an predictive approach to the ecosystem management that could serve for decision makers for the future generations. The process of eutrophication in coastal ecosystems has a complex relationship with the environmental factors. This was clearly demonstrated by the weak performance of multivariate linear regression (R2= 0.2). Based on the strong forecasting ability of SARIMA model, and the powerful expression ability of the NARX neural network in nonlinear relationships, a hybrid model which combines SARIMA and NARX is proposed in the fourth chapter, for Chl-a forecasting. The use of linear and nonlinear approaches to model the linear and nonlinear characteristics (respectively) of Chl-a in increased the efficiency and accuracy of the predictions. The NARX-SARIMA approach is aimed at highly heterogenic events, and the tests carried out proved its excellent performance. The use of hybrid model improved the prediction capability of one time step ahead of Chl-a with the best performance being achieved (R= 0.82; R2=0.68; MSE= 0.24).