A comprehensive analysis of Turkish sea level changes and future modeling using machine learning methods

Abstract

Global sea level rise is one of the increasingly evident impacts of climate change, with the potential to create irreversible changes in coastal regions and ecosystems. While the Intergovernmental Panel on Climate Change (IPCC) indicates that global sea levels will continue to rise throughout the 21st century with increasing acceleration, it also emphasizes that the rate of increase will not be uniform across the globe. The implications of the dramatic rise at a global scale, as highlighted by different climate scenarios, may vary regionally due to meteorological, oceanographic, and local factors. Therefore, understanding the regional variations in sea level and determining the coastal response to global trends are crucial for planning sustainable management practices tailored to the specific needs and risks of each area. This study was conducted to address precisely this need, motivated by the goal of producing sea level information that can be interpreted in a useful way for coastal regions and evaluated across different disciplines. Focusing on the coasts of Turkey, which were chosen as a prototype study area, sea level data were collected to analyze temporal and spatial changes. Risk areas along the coastline were identified based on the defined models and generated predictions/projections. This study, which focuses on sea level changes along Turkey's coasts, consists of four main stages: data analysis, spatial analysis, short-term prediction models, and long-term prediction/projection models. In the first half of the study, devoted to analysis, sea level oscillations along the Turkish coasts were examined both on a station basis and at a regional scale, identifying their recurring, distinct features and trends. The second half of the study, which is dedicated to prediction models, aims to understand the fundamental dynamics of sea level temporal changes and to determine probable future sea level values. The modeling studies are addressed in two dimensions, just like the analysis: point-based and regional. Short-term prediction models produce forecasts for 1, 7, and 30 days ahead at a single point. These models are supported by a rich methodological background, utilizing various prediction models and combinations of multiple algorithms to forecast data for a single station. The success of the models is evaluated using numerous performance indices. Long-term models, on the other hand, are established separately for each station along Turkey's coasts, using a single modeling approach. The station-based results obtained are interpolated and reported spatially. Map of potantial sea level increase are created, and the regional impacts of the model results are evaluated. After an overall evaluation, it is also beneficial to discuss the specific sections of the study. In the data analysis phase, temporal changes in sea level data collected from 18 different tide gauge stations along the coasts of Turkey were examined. Long-term trends were observed multidimensionally through various trend analyses. Additionally, the sea level component resulting from seasonal cycles at each station was identified. Furthermore, extreme value analysis was applied, statistically determining the highest values that could be observed in 100-year periods based on the distributions of the time series. The spatial analysis phase is dedicated to the cumulative examination of sea level data collected from all tide gauge stations along Turkey's coasts. In other words, regional analysis focuses on patterns and trends of regional sea level changes rather than the unique or location-specific characteristics of individual datasets. The general trends of all stations were analyzed using the QR pivoting-supported Empirical Mode Decomposition (EOF) method. The representability of this cumulative information with a reduced number of stations was investigated. Through this method, which could aid in optimizing sensor locations, it was shown that sea level changes across the entire region could be represented by only 4 stations with 96% accuracy. Short-term prediction models use machine learning methods that consider and learn the temporal internal dependencies of sea level data to generate forecasts or models based on these dependencies The daily sea level values for 1, 7, and 30 days ahead for the Istanbul station were obtained using various Machine Learning (ML) algorithms. The temporal internal dependencies of the series were defined using methods such as Extreme Gradient Boosting (XGB), Decision Trees (DT), Support Vector Machines (SVM), and k-Nearest Neighbors (kNN). The ML approach is frequently used and proven in the literature for modeling physical variables at both global and local scales. However, they may struggle to capture the change pattern solely in the time domain in series formed by the influence of many different frequency components. At this point, signal decomposition (SA) techniques like Wavelet Decomposition (WD), Empirical Mode Decomposition (EMD), and Singular Spectrum Analysis (SSA) were used to strengthen the prediction performance of ML methods. The results indicate a significant contribution of SA methods to the prediction success. Moreover, our methodological approach using these alternative methods allows for the comparison of the performance of different signal analysis and machine learning combinations. The study conducted at the Istanbul station presents a methodology that can be applied to sea level prediction models in various regions. Long-term models can be examined in two categories: prediction and projection. Unlike other sections, this part focuses not only on site-specific data but also on global sea level data, where long-term sea level values are produced. Sea level measurements were predicted from global sea level data produced under different climate scenarios. In long-term prediction models, which we classify as the first category, the relationship between the in-situ observation time series (tide gauge data) of each station and the sea level values presented as global climate model outputs is mathematically defined. The prediction method uses an stacking ensemble machine learning (EML) approach, evaluated for the first time in sea level prediction. Stacking in machine learning combines multiple models to improve prediction accuracy by training them on the same dataset and then using another algorithm to optimize their combined predictions. In the projection phase, the best model combinations will be used to generate local sea level projections from global ones, considering both optimistic and pessimistic carbon emission scenarios. Thus, a significant methodological contribution was made to the localization of global data. Possible sea level series until the end of the 21st century were created for each station. Maps showed how the 2100 values would affect the coastline. The results show that agricultural lands and wetlands along the coasts are the most affected regions by global and/or local sea level rise. The intrusion of saltwater into freshwater systems and soils poses a threat to drinking water sources, agriculture, and coastal ecosystems. In our study, sea level time series were primarily used. Main data sources for monitoring sea level rise include satellite observations, tide gauge records, and outputs from atmospheric/oceanographic climate models. In the sections created to identify local sea level changes, tide gauge data were used in both the data analysis and regional analysis sections. Current tide gauge data from 18 different points along the Turkish coast are provided by the TUDES monitoring service, conducted by the General Directorate of Mapping. At some points, accessible sea level data extend up to 25 years into the past. The Istanbul Station, located at the northern entrance of the Bosphorus, began operating in 2014. The sea level data recorded in Istanbul differs somewhat from other tide gauge locations in terms of both oceanographic, meteorological, and demographic characteristics. Therefore, when the need arises to select a station to apply short-term modeling procedures and test the success of prediction methods, Istanbul was determined as the most suitable point, taking into account its unique conditions. Furthermore, long-term models linked global and local data along the Turkish coast. We believe that this integrative approach, combining the advantages of different data sources, will be beneficial for achieving more comprehensive and accurate results by providing a multidimensional perspective. The conducted analyses and modeling studies have developed comprehensive models that provide both short- and long-term forecasts for sea level changes along the coasts of Turkey, and these models have been validated for accuracy and prepared for practical applications. Initially, different oscillations, trends, and distributions were observed at stations located in three different seas surrounding Turkey. However, it is possible to identify a common, underlying pattern in the regional sea level changes. Taking into account the multi-component nature of sea level when determining modeling procedures has greatly contributed to the prediction success. Signal analysis techniques have helped to resolve the complex, multi-component structures in the time series data, allowing machine learning models to better understand this data. Developing well-localized sea level projections under different climate scenarios will be crucial for making informed decisions, particularly for adapting to and mitigating the impacts of sea level changes in coastal areas. The results, especially in light of expected sea level rise, can guide decision-makers in coastal risk management, urban planning, and policy formulation. Future research can expand on this study by incorporating more complex climate interactions, exploring additional prediction methods, and developing dynamic, adaptable management strategies to enhance resilience against sea level rise. It is also worth noting that the methods developed by examining sea level changes along Turkey's coasts offer a flexible and adaptable modeling framework that can be applied to other regions. The data analysis, short- and long-term prediction models, signal analysis-supported machine learning approaches, and risk assessment methods used in this study are not exclusive to Turkey's coasts; they also contain general principles and techniques that can be adapted to sea level studies in different regions. Therefore, the approach presented in this study can be effectively used to understand and predict sea level changes in various coastal regions, accommodating their unique dynamics and data structures.