Transfer sea level learning in the Bosphorus Strait by wavelet based machine learning methods

Accurate prediction of sea level fluctuations is fundamental to coastal engineering. However, limited monitoring service on a regional scale is the most important constraint in analyzing, identifying, or predicting sea-level fluctuations driven with several nonlinearly integrated deterministic processes. The present study investigates the prediction performance of machine learning (ML) models based on available sea level time series information recorded in different conditions. Discrete Wavelet Transform (DWT) combined with Support Vector Machine (SVM), k-Nearest Neighbor (KNN), and Decision Tree (DT) methods were used to transfer sea level information between 3 stations located in different regions of the Bosphorus Strait. The developed models are tested in predicting sea level lead-time up to 7 days based on the Root Mean Square Errors (RMSE) and Nash-Sutcliffe Efficiency (NSE) indicators. Modeling strategy is determined by taking the sensitivity of a classical regression technique, Multi-linear Regression (MLR) into account, to additional decomposition or standardization processes. The developed models are found to be more successful in the information transfer between spatially close stations than periodically close stations. Considering the relative success of ML methods in defining the sea level fluctuations, SVM and KNN models provide relatively close results while DT model results are far behind the others.

Automated summary

Accurate prediction of sea level fluctuations is crucial for coastal engineering, and machine learning models show potential in sea level time series information under different conditions. The developed models perform better in transferring information between spatially close stations, with SVM and KNN models providing relatively close results while DT model results lag behind significantly.