Elsevier

Ocean Engineering

Volume 233, 1 August 2021, 109116
Ocean Engineering

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

https://doi.org/10.1016/j.oceaneng.2021.109116Get rights and content

Highlights

  • Wavelet-based machine learning models are developed to predict sea level up to 7-day lead-times.

  • Developed models are trained and validated with data recorded at a different date and/or at a different location.

  • The methodological innovation presented in this study provides transferring information between stations' data.

  • Adaptation of linearly distinguished model structure into machine learning regression models are highlighted.

Abstract

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.

Introduction

The prediction of the future sea level has great importance in the planning and implementation stages of marine projects which are related to different issues such as environmental pollution (Maderich and Konstantinov, 2002), coastal management (Pethick, 2001), transportation (Whittington, 2016), renewable energy (Aboobacker, 2017) and sub-marine structures (Oda et al., 2009). Additionally, sea level records can be utilized as a reference parameter in strait hydrodynamic modeling, because surface water movement is the major operational factor that drives the coastal sea dynamics (Cucco et al., 2016). Concerns about local sea level rise and the potential impact of extreme events on coastal areas have increased the necessity of the development of the level of knowledge worldwide (Pachauri and Reisinger, 2007). However, the prediction of the future sea level for a short lead time is still limited due to the lack of in-situ records (Anderson, 2013; Hil, 2020).

Generation of site-specific information depending on the analysis of the cumulative consequence of the multiple synergic effects on a marine system needs long-term monitoring and a compact measurement network (Erol, 2011). Yet, economic and technical impossibilities can limit the number of tide gauge stations and their service life (Joseph, 2014). The absence of sufficient data encourages modeling based on information transfer obtained from one dataset to another for filling data gaps in a particular study area (Huang et al., 2003). Chang and Lin (2006) designed a neural network model to predict harmonic sea level components that depend on astronomical tide generating forces and extended the simulation to obtain future tides at six neighboring points to the location where the model is trained. Zhu et al. (2020) transferred water level fluctuation memory compiled with neural network and deep learning methods from records of multiple lakes in a region to water level prediction models of each lake in the region.

The high potential of Artificial Intelligence-based mathematical approaches in perceiving complex patterns and making rational decisions has brought them among the acceptable real-world prediction methods. The machine learning (ML) approach is a type of artificial intelligence (AI) technique that designs and develops processes based on algorithms that enable computers to learn from data. The main benefit of these prediction methods is that they infer stochastic dependency knowledge between the past and the future based on the historical data without the need of being explicitly programmed (Bontempi et al., 2012). As regards the main focus of this research, several AI applications can be mentioned in specialized sea level prediction (Ali Ghorbani et al., 2010; Lai et al., 2019; Pashova and Popova, 2011). Performance comparison of data-driven approaches is carried out from different perspectives, i.e. monthly sea level rise prediction (Lai et al., 2019), forecast of sea level anomalies (Kurniawan et al., 2014), analyzing daily sea-level fluctuations in the coastal regions (Imani et al., 2018), and prediction of flow variables in curved channels (Gholami et al., 2018).

Preliminary pre-processing techniques support the analysis and interpretation capability of data-driven models by gradually extracting multilayered information embedded into time series. Harmonic analysis is a classical decomposition method used with the intent of determining the harmonic components (e.g. tidal effect) that contribute to sea-level changes in the frequency domain. However, due to concerns about the loss of information resulting from the steady-state assumption of classical methods, that obscure dynamic changes, wavelet has become to be used as a prominent preferred signal-processing tool in sea level time series (El-Diasty et al., 2018). Thanks to its capability in revealing time-frequency information embedded into the time series, wavelet supports the predictor algorithms to determine the meteorological events on coastal high-frequency sea surface motion (de Oliveira et al., 2009), the major tidal component (Lee, 2004), the effect of unstable driving forces and extreme environmental conditions (Ali Ghorbani et al., 2010) and, non-linear components of tidal current speed and direction (Kavousi-Fard, 2016).

The primary purpose of this study is to develop a predictive mathematical model to identify spatial and temporal relations in sea-level time series that are collected through the Bosphorus. Besides, a developed methodology using alternative model structures serves the second aim of the present study, which is to investigate the contribution and effects of pre-processing and their impact on the prediction accuracy of machine learning methods. A further objective of this study is to be able to predict the sea levels at any stations with developed models via one of the station data located in the Bosphorus. Past studies have been focused on lead-time prediction in a certain station with train and test components derived from a single data set of the station, which is used for learning and testing purposes, respectively. In this study, developed models are evaluated employing a non-dimensional metric Nash and Sutcliffe Efficiency (NSE) and a classic error metric root mean squared error (RMSE). The proposed modeling approach based on the transfer of existing system information to new situations is considered to be of practical benefit, especially in regions where there are limited available or missing data.

Section snippets

Study area and data

As the only waterway of the Black Sea to Marmara and thereby the Mediterranean Sea is Bosphorus, it has crucial importance not only for local but also for international oceanographic projects (Myers et al., 2003). How the vertical and horizontal movements of the water in the region influence the hydrodynamic behavior of Bosphorus and connected oceanographic systems is still a current interest and topic of ongoing researches (Altunkaynak and Kartal, 2019; Hossain and Meng, 2020). Though several

Model structure development

In complex system modeling, determining modeling strategy and designing an appropriate model structure are foundational to enhance the efficiency of machine learning. Since sea-level fluctuations are the cumulative responses to many hydrodynamic mechanisms varying in a wide frequency range, decomposition of sea-level data into its different components embedded into sea-level time series is the advised methodology in the literature to extract meaningful information (Cheng et al., 2017; Löfgren

Conclusion

This study uses the performances of the developed predictive mathematical models in predicting future sea levels in Bosphorus from available data that are tested in temporally and spatially different conditions. The success of the modeling approach introduced in this study, in transferring the learned modeling capabilities between stations located at different regions and/or recorded on different dates were evaluated by means of RMSE and NSE metrics.

The essential point in transferring

CRediT authorship contribution statement

Abdüsselam Altunkaynak: Conceptualization, Resources, Supervision, Writing – review & editing. Elif Kartal: Methodology, Formal analysis, Writing – original draft, Validation, Investigation, Visualization.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledge

We sincerely thank to the General Directorate of Marmaray Project of Ministry of Transportation, General Directorate of Port, Airports and Railways Construction of Turkey to provide us water level data.

References (43)

Cited by (9)

  • Transfer learning of long short-term memory analysis in significant wave height prediction off the coast of western Tohoku, Japan

    2022, Ocean Engineering
    Citation Excerpt :

    Altunkaynak and Kartala (2021) predicted the sea level at the exits of the Bosphorus Strait by applying transfer learning to a wavelet-based machine learning method. The lead time for accurate sea-level prediction increased to 7 days (Altunkaynak and Kartala, 2021). Furthermore, a combination of transfer learning and LSTM has been employed to predict the correlation between those affected by COVID-19 disease and death (Gautam, 2022) and soybean harvest (Wang et al., 2018b).

  • Extended lead time accurate forecasting of palmer drought severity index using hybrid wavelet-fuzzy and machine learning techniques

    2021, Journal of Hydrology
    Citation Excerpt :

    Therefore, it can generate very good model results with using a simple algorithm. Disadvantages of kNN, according to Altunkaynak and Kartal (2021), the selection of suitable parameters for the development of the kNN model is pretty challenging and may have negative effects on the processing performance. The kNN tool has no mapping function that relates input to output and also makes extrapolation.

  • Predicting Sea Level Rise Using Artificial Intelligence: A Review

    2023, Archives of Computational Methods in Engineering
View all citing articles on Scopus
View full text