Akademik Veri Yönetim Sistemi
Regional Studies in Marine Science, cilt.90, 2025 (SCI-Expanded)
In this study, we present a novel methodology to predict water levels (WLs) from its components (meteorological, hydrological, tidal, seawater, and air parameters) by using a machine learning (ML) approach. The novelty of the approach is 1) random selection of training data, 2) to determine significant predictors and eliminate the others by applying the t-test statistic, and 3) to check linear (via Variance Inflation Factor (VIF)) and nonlinear (via Distance Correlation (DisCor)) dependencies between the significant predictors to avoid using similar parameters. Various linear and non-linear ML approaches are employed to obtain the optimal regression model. We tested the prediction capacities of the model in four coastal seas: the enclosed Black Sea, the inland Sea of Marmara, the semi-enclosed Aegean Sea, and the open (marginal) Levantine Sea. We achieved the most significant correlations for the Ensemble of Trees (ET) approach in the Black Sea and Levantine Sea WL predictions, with correlation coefficients of R≅ 0.93–0.95 in the former and R≅ 0.89–0.93 in the latter. On the other hand, the Gaussian Process Regression (GPR) yielded the most significant correlations in the Sea of Marmara (R≅0.93) and the Aegean Sea (R≅0.82). Water level (WL) variability was primarily driven by Danube River discharge and air pressure in enclosed and inland areas, the air pressure and tidal forces in semi-enclosed seas, and predominantly by tidal influences in open sea regions. Two parameters (sea surface temperature (SST) and 2-m air temperature (2m-T)) were relevant in all locations considered after applying the VIF test, and the lower-correlated predictor was excluded from predictions. Overall, the proposed method notably increased the predicted WLs' accuracy compared to previous ML studies.