Use of Spectral Clustering for Identifying Circulation Patterns of the East Korea Warm Current and Its Extension

Abstract

A graphical clustering approach was used to objectively identify prevalent surface circulation patterns in the East/Japan Sea (EJS). By applying a spectral clustering algorithm, three distinct patterns in the East Korea Warm Current (EKWC) and its extension were identified from daily maps of reanalyzed sea surface heights spanning the past 30 years. The results are consistent with previous studies that used manual classification of the EKWC’s Lagrangian trajectories, highlighting the effectiveness of spectral clustering in accurately characterizing the surface circulation states in the EJS. Notably, the recent dominance of northern paths, as opposed to routes along Japan’s coastline or those departing from Korea’s east coast further south, has prompted focused re-clustering of the northern paths according to their waviness. This re-clustering, with additional emphasis on path length, distinctly categorized two patterns: straight paths (SPs) and large meanders (LMs). Notably, SPs have become more prevalent in the most recent years, while LMs have diminished. An autoregression analysis reveals that seasonal anomalies in the cluster frequency in spring tend to persist through to the following autumn. The frequency anomalies in the SPs correlate strongly with the development of pronounced anomalies in the gradient of meridional sea surface height and negative anomalies in the surface wind stress curl in the preceding cold seasons. This relationship explains the observed correlation between a negative Arctic Oscillation during the preceding winter and the increased frequency of SPs in the subsequent spring. The rapid increase in the occurrence of SPs indicates that a reduction in LMs limits the mixing of cold, fresh, northern waters with warm, saline, southern waters, thereby reinforcing the presence of SPs due to a strengthened gradient of meridional surface height and contributing to a slowdown in the regional overturning circulation.

1. Introduction

The East/Japan Sea (EJS) exhibits a unique and complex circulation system dominated by three major surface currents: the Tsushima Warm Current (TWC), the East Korea Warm Current (EKWC), and the Liman Cold Current (LCC). These currents flow through the southern, central, and northern regions of the basin, respectively, forming a cyclonic gyre in the south and an anticyclonic gyre in the north. The EKWC, a branch of the TWC, flows northward along the eastern coast of the Korean Peninsula before turning eastward [1,2,3,4]. One branch of the EKWC follows the frontal zone separating subtropical and subarctic waters, situated between 39° and 41° N. The path of this current demonstrates strong seasonal variability, driven by monsoonal winds and associated currents [5,6,7,8,9,10,11].

Previous studies, including Lagrangian trajectory analyses [12,13,14], have sought to identify the major circulation patterns in the EJS. These studies have revealed significant variability in the EKWC’s eastward turn, influenced by regional oceanographic conditions, climate variability, and long-term climate change. However, there remains a need for further exploration beyond the meridional location of the EKWC’s branching. Manual classification methods, which rely on identifying the branching locations, have limited applicability in capturing the distinct spatial features and temporal evolution of the region’s surface circulation. To address these limitations and advance our predictive capabilities, more objective and automated approaches are required.

Variability in the circulation patterns has been attributed to multiple factors, including the EKWC’s speed, influenced by variability in the mass transport at the Korea Strait, atmospheric thermal forcing, and vertical mixing driven by changes in local stratification. These factors, however, are interdependent and collectively respond to large-scale climate variability and change. For example, the Arctic Oscillation (AO) exerts a significant influence through surface wind forcing [15], while El Niño–Southern Oscillation (ENSO) and the Pacific Decadal Oscillation (PDO) alter storm tracks, generating atmospheric perturbations [16]. ENSO additionally affects the variability in the Korea Strait’s transport [17]. Despite these insights, it remains unclear whether the long-term variability in the EJS is primarily driven by oceanic conditions, such as the Korea Strait’s transport and salinity advection, or atmospheric conditions represented by climatic indices through heat and momentum fluxes.

The EKWC plays a critical role in the EJS by transporting warm, saline water northward, where it interacts with fresh, colder waters. This interaction strongly influences deep-water formation and the regional overturning circulation. In summer, strong surface currents dominate the EJS, facilitating interactions through eddies and advection, which precondition the region for deep-water formation during the following winter. This preconditioning, driven by the northward delivery of saline surface water during warmer seasons, enhances deep-water formation through vertical mixing and cooling from the surface heat flux associated with strong northwesterly winds in winter [18,19]. Despite their clear importance, our understanding of the variability in deep circulation remains poorly connected to ambient climatic conditions, primarily due to limitations in the available data.

This study leverages a spectral clustering method applied to 30 years of daily sea surface height (SSH) data to objectively identify distinct patterns in the EKWC’s circulation. Our approach reveals three primary patterns consistent with prior Lagrangian trajectory studies, validating spectral clustering as a robust tool for characterizing the EJS’s surface circulation. Given the recent dominance of northward-flowing EKWC paths, we conduct a secondary clustering based on path waviness, distinguishing between straight paths (SPs) and large meanders (LMs). An analysis of the temporal variability shows a notable increase in SPs and a corresponding decrease in LMs in recent years. Furthermore, we explore the relationship between these patterns and large-scale climate indices, such as the Arctic Oscillation (AO), revealing sequential propagation of winter atmospheric conditions to anomalous regional current patterns in subsequent seasons. Our findings suggest that the recent shift toward straighter paths is linked to changes in the meridional SSH gradient and the wind stress curl, potentially impacting the water mass mixing and regional overturning circulation within the EJS.

This paper is structured as follows: Section 1 outlines the data and methods. Section 2 presents the results of the clustering analyses and subsequent physical interpretations. Section 3 discusses the role of distinct circulation patterns in the variability in deep-water formation. Section 4 concludes this study.

2. Materials and Methods

2.1. Spectral Clustering Analysis

The spectral clustering algorithm from the Scikit-learn library categorized the major patterns of the East Korea Warm Current (EKWC). Spectral clustering is a graph-based clustering approach that leverages the eigenvalues of a similarity matrix to group data based on their connectivity in graph space [20,21]. Unlike traditional clustering methods (e.g., k-means) that rely on the distances between data points, spectral clustering uses the eigenvalues (or spectrum) of a similarity matrix, such as a graph Laplacian, to reveal the structure of the data. It is simple and has been effectively applied to detecting the trajectory patterns in oceanic flows [22,23], making it particularly effective for identifying complex, non-linear relationships within data and capturing the intricate spatial variations and patterns in the EKWC’s path.

The latitude and longitude coordinates of the defined path of the EKWC were used as the input for clustering, initially categorizing the EKWC’s path into three distinct patterns. In this initial clustering, we employed the nearest-neighbors metric as the affinity, where affinity refers to a similarity measure that defines the connectivity between the data points. The nearest-neighbors affinity metric links each path coordinate to its closest neighboring points, effectively capturing the local spatial relationships along the EKWC’s path. As the number of clusters is user-defined, two to five clusters were tested, with three clusters ultimately being chosen, as adding more did not reveal additional distinctive patterns.

In the second stage of clustering, we focused on distinguishing one of the three clusters further based on path length. The path length (P) between each pair of coordinates was calculated as follows:

$$P=\sum _{i=1}^{n-1}\sqrt{{\left(∆x_i\right)}^2+{\left(∆y_i\right)}^2}$$

where $∆x_i$ and $∆y_i$ represent the longitudinal and latitudinal differences between consecutive points along a path, with n being the total number of points in the path. To incorporate this path length information into the clustering affinity, the path length difference (PD) is computed as follows:

$$\mathit{P}{\mathit{D}}_{\mathit{j}\mathit{k}}=\left|P_j-P_k\right|$$

Based on this path length difference (PD), we generated a similarity matrix S $\in {\mathit{R}}^{\mathit{N}\times \mathit{N}}$ using the Radial Basis Function (RBF) kernel, where N is the total number of EKWC trajectories. Each element of the matrix, s_jk, represents the similarity between the j-th and k-th paths and is symmetric (s_jk = s_kj). The similarity matrix S is constructed using the following RBF kernel formula:

$$\mathit{S}=\mathbf{e}\mathbf{x}\mathbf{p} (-\mathit{\gamma }\mathit{P}\mathit{D})$$

where γ was set to 0.001 to capture subtle variations in the patterns in the EKWC’s path. This transformation yielded an RBF kernel matrix, emphasizing shorter path lengths between similar patterns, and allowed for a refined separation of the IBC pattern into two distinct clusters based on the path length. Following the computation of this similarity matrix, the spectral clustering algorithm was used to calculate the graph Laplacian L using the following formula:

$$\mathit{L}=\mathit{D}-\mathit{S}$$

where D is the degree matrix, meaning a diagonal matrix in which each element D_ii represents the sum of the similarities connected to node i, and S is the similarity matrix. This graph Laplacian reflects the connectivity and structure of the data.

The algorithm then performed eigenvalue decomposition on L, identifying the most significant eigenvectors for projecting the data into a lower-dimensional space. Finally, in this reduced space, k-means clustering was applied to achieve the final classification of the patterns in the EKWC’s path.

2.2. Reanalysis of Datasets for the Ocean and Atmosphere

This study utilized GLORYS12 version1 (GLORYS12v1) sea surface height (SSH) data provided by the Copernicus Marine Environment Monitoring Service (CMEMS) for the period 1993–2023 to define the path of the East Korea Warm Current (EKWC) and estimate the surface’s geostrophic current. The SSH data have a horizontal resolution of 1/12° × 1/12° (approximately 8–9 km in mid-latitudes) and were assimilated with satellite altimetry and in situ observations, ensuring consistent oceanographic fields (https://doi.org/10.48670/moi-00021). Also, the monthly vertical temperature and the meridional ocean current from the monthly mean GLORYS12v1 data were included for investigation of the vertical thermal structure and the meridional overturning circulation (MOC) in the region. The temperature data, covering 50 vertical levels from the surface to a 4000 m depth, allowed for a detailed analysis of the regional thermal structure. The meridional current data were used to calculate the regional MOC, capturing the cumulative meridional flow to reveal the regional thermohaline circulation patterns influenced by the EKWC, in good agreement with a previous study on other reanalysis data [24].

The monthly mean Pacific Decadal Oscillation (PDO) index, obtained from the NOAA Physical Sciences Laboratory (https://psl.noaa.gov/pdo/ accessed on 24 October 2024), and the Arctic Oscillation (AO) index, from the NOAA Climate Prediction Center (https://www.cpc.ncep.noaa.gov accessed on 24 October 2024), for the same period were included for investigation of the influence of climate on the EKWC. The mean values of these indices during winter (DJF) were examined in terms of their impact on the subsequent seasonal conditions of the EKWC. Additionally, monthly 10 m u-wind data from the ERA5 reanalysis, provided by the European Centre for Medium-Range Weather Forecasts (ECMWF; https://doi.org/10.24381/cds.adbb2d47), were included to assess the impact of atmospheric forcing on the EKWC’s conditions during different AO phases. ERA5 data served as the surface boundary condition for the GLORYS12v1 model, ensuring consistency in the ocean–atmosphere interaction.

2.3. Definition of the EKWC’s Path

The trajectory of the East Korea Warm Current (EKWC) was defined based on the sea surface height (SSH) contours and geostrophic current velocity in the EJS. Specifically, the EKWC’s path was defined as the SSH contour aligned with the geostrophic current speed above the 99th percentile within a latitude range of 34° N to 42° N and a longitude range of 128° E to 140.5° E. This threshold was chosen to represent the areas with the strongest geostrophic currents, effectively capturing the core flow of the EKWC.

The use of SSH contours to define the EKWC’s path allows for a continuous representation of the current’s trajectory. This method ensures coherent tracking of the EKWC, as SSH contours provide smooth and uninterrupted paths, which is essential for analyzing the spatial variations in and meandering behavior of the current over time.

3. Results

3.1. Three Major Clusters for the Surface Circulation Patterns in the East/Japan Sea

Three distinct patterns of surface circulation in the EJS were identified using the spectral clustering method: the Inertial Boundary Current (IBC, Figure 1a), the Ulleung Eddy (UE, Figure 1b), and the Tsushima Warm Current (TWC, Figure 1c). Based on 30 years of daily observations, totaling 11,322 data frames, three unique types of EJS surface circulation were revealed. The three clusters are clearly separated by their meridional positions. Cluster IBC is located around 39.5° N, capturing the pathways that extend northward from the east coast of Korea. Cluster UE, centered around 37.5° N, has the highest number of observations and features wide meanders. The final cluster, the TWC, represents a current path that separates and turns toward Japan’s northwestern coast around 35.7° N before approaching Korea. These clustering results align with previous studies [12,13,14] that manually classified the Lagrangian trajectories based on specific features, such as the turning points and branching areas.

Among the three clusters, the UE, the middle cluster, prevails regardless of season, whereas the other two exhibit quite distinctive seasonal preference. The TWC, the southern cluster, prefers summer, while it is substantially rare in other seasons. The IBC, the northern cluster, contains two peaks, centered in winter and in late spring. The IBC is the only cluster that shows an upward trend over the past 30 years (Figure 1d). In contrast, the UE cluster shows a declining trend (Figure 1e), while the TWC cluster exhibits decadal variability without a significant long-term trend (Figure 1f). This suggests that the current pathways are increasingly reaching further north before turning eastward. These clusters are closely related to the local maxima of the gradient of meridional sea surface height, as seen in Figure 2, where each cluster is centered around the zonal mean SSH gradient peaks at 35.7° N, 37.5° N, and 39.5° N, directing each cluster eastward under geostrophic constraints.

Regardless of the distinctive seasonal preferences amongst the three current paths, only one seasonal cycle is found in the initial speed of the current for all three clusters. For all three clusters, the speed of the entering current maximizes around August, while it minimizes around February. Figure 3 shows that the average speed of the current at the start of the EKWC, prior to the separation of the TWC cluster, does not exhibit statistically significant differences across the clusters. This indicates that the branching of each current’s pattern is unlikely to be attributable to variability in the mass transport into the region near the Korea Strait. In each month, the average speed of the current across the western channel of the Korea Strait corresponding to each cluster differs only within the uncertainty bound of the climatological means. The frequency and strength of these clusters appear to be minimally affected by the transport through the Korea Strait.

The unusually low average current speed at the timing of the TWC during October and December are not considered significant. This is due to the very low likelihood of occurrences of the TWC during these months. Over the past 30 years, occurrences of the TWC have been recorded on only a few dozen days. Given these limited data, it is statistically challenging to make meaningful comparisons of the entrance speeds under the TWC with those of the other clusters. Therefore, it was excluded from the discussion to maintain statistical rigor.

The agreement between our analysis and previous findings highlights the effectiveness of spectral clustering in classifying the EKWC’s paths. This approach offers a more systematic classification compared to that of the traditional methods.

3.2. Straight Paths and Large Meanders in the IBC Cluster

The re-clustering of Cluster IBC (Figure 1a,d) according to the length of each path effectively highlights that the EKWC, in leaving the northern coast, tends to choose between two contrasting states: straight paths (SPs) and large meanders (LMs). While the zonal movement of both clusters is constrained around the average zonal axis of the IBC, their distinctiveness is predominantly characterized by their meridional displacement.

The LMs are widely distributed, spanning across the central EJS and exhibiting a broad spatial spread with complex undulations (Figure 4a). Compared to the SPs, the path density of the LMs along the path’s axis is significantly weaker and less organized, while the meridional extent is considerably broader. Notably, the LM pattern shows a much further northward reach into the East Korea Bay compared to that of the SPs. As shown with the strong and concentrated density values aligning with the main current axis, the SPs follow a substantially more linear trajectory with minimal deviations compared to the LMs (Figure 4b). The SPs predominantly follow a straight northeastward direction, with limited excursion.

Both clusters exhibited upward trends of a similar magnitude, collectively contributing to a significant increase in the count of occurrences of the IBC, which rose at a rate of approximately 4–5 occurrences per year. Notably, SPs have shown more rapid growth than LMs in recent years (Figure 4c,d). Meanwhile, the relative occurrence of the two clusters, which indicates the dominance of one pattern over the other, suggests that this dominance fluctuates with interannual to decadal variability, independent of the linear trend. This variability highlights that the EJS’s circulation reflects ambient variability in the climate system, rather than merely a gradual response to climate change.

The relative occurrence of the two clusters, analyzed across different seasons and years, was examined further to uncover their temporal characteristics. Climatologically, both clusters exhibit a preference for colder seasons. Additionally, LMs slightly dominate over SPs from spring to summer, while SPs surpass LMs from autumn to winter. Moreover, the seasonal anomalies in their relative occurrence reveal significant differences in their persistence. As shown in the autocorrelation heatmap in Figure 5, this asymmetry is particularly evident in the seasonal persistence of anomalous relative occurrence. Persistence refers to the influence of anomalies in one season on subsequent seasons, resulting in non-zero correlations between them. Cold-season anomalies, such as those in October–November–December (OND) and January–February–March (JFM), show no significant correlation with the anomalies in the seasons immediately after. In contrast, anomalies during warm seasons, such as April–May–June (AMJ) and July–August–September (JAS), demonstrate significant correlations with subsequent seasons. This suggests that anomalous conditions in warmer seasons can persist in subsequent seasons, whereas the reverse is unlikely.

The intensification of the zonal mean geostrophic current is strongly linked to the increased prevalence of SPs. As depicted in Figure 6, the composite-averaged daily cross-sections of the vertical temperature profiles, calculated for days marked by the occurrence of SPs, highlight distinct stages in the evolution of anomalies in the baroclinic zonal geostrophic current. These anomalies are associated with SPs through the formation of a dynamic height gradient along the cluster’s main axis.

During the winter months, the maximum gradient is observed at around 38° N, driven by surface cooling across the meridional extent of the IBC cluster. By AMJ, subtle surface warming begins to emerge in the south, while the strong meridional gradient of dynamic height persists and shifts northward. In JAS, the cooling from the previous winter diminishes, whereas southern warming intensifies, sustaining the meridional SSH gradient. Finally, in OND, the southern warming peaks, maintaining the strongest gradient near the cluster’s axis. This sequence of processes from AMJ to OND appears to be interconnected, likely facilitated by northward advection and the seasonal migration of temperature anomalies. This chain of events is further supported by the autocorrelations between the seasonal occurrence of SPs shown in Figure 5.

The investigation then explored whether winter events are truly independent of other seasons, as suggested by their apparent isolation from the sequential processes observed during the subsequent three seasons. They also show little to no connection with the preceding OND season. However, it was discovered that the atmospheric conditions during winter are significantly linked to the occurrence of SPs/LMs in the following AMJ.

As shown in Figure 7, a negative/positive AO index during the winter months (DJF) is strongly correlated with the counts of SPs/LMs in the following spring. This correlation is physically plausible due to the interaction between intense winter atmospheric forcing and a delayed dynamic ocean response.

During negative AO phases in the winter, the surface winds over the EJS exhibit westerly anomalies. These westerly winds trigger immediate cooling of the SST in the EJS through enhanced vertical mixing and turbulent surface heat loss. Additionally, they generate a positive anomaly in the wind stress curl, which spins up clockwise circulation in the southern EJS, with a lag of a few months. Figure 8a shows the linear regression coefficient of the vertical ocean temperature in DJF with the standardized negative AO index during the same period. Remarkably, this pattern closely resembles in both structure and magnitude the composite average based on the occurrence of SPs during DJF (Figure 6a).

Meanwhile, the positive anomaly in the wind stress curl associated with negative AO suppresses the thermocline in the southern EJS, resulting in a positive anomaly in the dynamic height in the region. By elevating the southern area, this anomaly creates a steeper gradient of meridional sea surface height (SSH), which enhances the zonal geostrophic current and increases the activity of SPs. However, this adjustment process unfolds over the course of a season, ultimately producing the strongest lagged correlation between the atmospheric conditions in winter and the surface current patterns in spring.

In contrast, PDO did not show a significant correlation with the SP/LM counts (Figure 7). Although it is not shown here, ENSO was also found to have no significant relationship. This emphasizes the distinct role of the AO index in influencing the SP/LM counts during subsequent seasons.

4. Discussion

The behavior of the EKWC previously studied with Lagrangian trajectory classification using branching locations [12,13,14] was revisited using a machine learning technique. This study applied spectral clustering method to 30 years of daily sea surface height data to objectively identify the EKWC’s circulation patterns, revealing three dominant patterns consistent with prior analyses. Given the increasing prevalence of northward-flowing paths of the EKWC, we further performed a secondary clustering based on the path waviness, distinguishing between SPs and LMs. Our analysis showed a recent increase in SPs and a decrease in LMs, with a strong correlation between the negative AO in winter and an increased frequency of SPs in the following spring. These results suggest that straighter paths are linked to changes in the meridional SSH gradient and the wind stress curl, potentially affecting the water mass mixing and regional overturning circulation in the East/Japan Sea.

The increase in the occurrence of SPs across multiple seasons suggests a straightforward positive feedback mechanism. A compelling argument is that the meandering of the EKWC plays a critical role in efficiently mixing the warm, saline water from the south with the cold, fresh water from the north along its path. Conversely, if the EKWC’s extension maintains a straight eastward trajectory, the chances of such mixing would be significantly reduced. When meandering dominates, the sharp meridional gradient of dynamic height along the axis of both clusters is weakened by eddy-driven mixing. In contrast, when the number of SPs increases, this mixing is limited, and the fast current reduces the heat loss to the atmosphere, maintaining a sharp meridional SSH gradient all the way to the basin exit. This process operates as a self-sustaining mechanism, requiring atmospheric forcing only to initiate and terminate the cycle.

We acknowledge that this study does not thoroughly address the role of adjacent ocean currents in determining the preferred circulation patterns. This limitation arises from our primary focus on clustering the pathways rather than analyzing the speed of the current along these paths. While our findings significantly contribute to identifying the major current patterns, incorporating the current speed could provide a more comprehensive understanding of the regional circulation. Future research should explore this aspect, potentially leveraging spectral clustering to complement our current results. Such efforts could help to bridge the gap with a previous study [25], which used numerical modeling experiments to identify the dominant role of ocean currents in the interannual variability in the circulation of the EJS.

When considering the ocean’s role in delivering subtropical water to the north of the basin, salinity emerges as the most critical factor in initiating regional deep-water formation, with the sinking depth largely influenced by the type of convection. Among the two known convection types, open convection is driven by surface winds, where wind-induced heat loss causes the surface water to sink. However, due to strong heat diffusivity, open convection cannot easily penetrate to the deepest parts of the basin. In contrast, salinity-driven deep convection, associated with sea ice formation, facilitates sinking of the surface water to the basin’s floor [19,24]. As the dominant salt supplier for deep-water formation, the EKWC plays a decisive role in determining whether convection can occur. With a fast speed and limited mixing before the exit from the basin, increased SPs would hinder deep thermohaline circulation in the region.

The meridional overturning circulation estimated in the EJS (ESMOC) supports our argument (Figure 9), revealing significant weakening when the occurrence of SPs outnumbers the occurrence of LMs. Due to the limited data span and pronounced interannual to interdecadal variability, no substantial evidence suggests long-term changes in the ESMOC. Instead, these findings imply a straightforward linear relationship, indicating that an increase in SPs would lead to slowdowns in the ESMOC. This conclusion aligns with previous observations in the East/Japan Sea [19] and other studies using ocean reanalysis products, such as [24].

5. Conclusions

The spectral clustering approach employed in this study successfully unveiled distinct spatial patterns in the East Korea Warm Current (EKWC), offering critical insights into its spatial and temporal variability. The three primary clusters identified—the Inertial Boundary Current (IBC), the Ulleung Eddy (UE), and the Tsushima Warm Current (TWC)—are well aligned with previously documented pathways but provide a more systematic classification through an advanced graph-based analysis. Its ability to capture non-linear relationships and localized spatial structures highlights its superiority over the traditional clustering methods, particularly in capturing intricate oceanographic processes. The seasonal and long-term trends observed within the clusters further underscore the dynamic nature of the EKWC and its sensitivity to changes and variability in the coupled climate system.

A deeper investigation into the IBC cluster, using path length as an additional criterion, reveals two contrasting patterns: straight paths (SPs) and large meanders (LMs). The distinct spatial distributions and temporal trends in these sub-clusters reflect the interplay between anomalous baroclinic geostrophic factors and climatological seasonality in shaping the current’s trajectory. The intensification of SPs, coupled with their increasing prevalence in recent decades, suggests a potential shift in the EKWC’s behavior under changing climatic conditions. Interestingly, the fluctuations in the dominance of SPs and LMs observed, along with their seasonally varying persistence, emphasize the influence of atmospheric conditions on the surface as external drivers, which contribute to the interannual and decadal variability beyond gradual trends.

The link between atmospheric forcing, particularly the Arctic Oscillation (AO), and the EKWC’s dynamics is particularly notable. This study demonstrates how winter AO phases induce delayed oceanic responses that manifest in the subsequent seasonal occurrence of SPs and LMs. Negative AO phases in winter are associated with enhanced zonal geostrophic currents due to intensified westerly winds, which induces surface cooling concurrently, facilitates positive wind stress curl anomalies and dynamic height adjustments in the southern East Sea in the following season. This delayed response, spanning months, highlights the importance of atmospheric–ocean interactions and positive oceanic feedback in modulating the EKWC’s pathways. Such findings underscore the EKWC’s role as a sensitive indicator of broader climate change and variability in the region.

Finally, the relationship between the activity in terms of SPs and evolving baroclinic structures, as evidenced by the composite vertical temperature profiles, reflects the intricate coupling between thermohaline processes and surface current dynamics. Seasonal anomalies in the sea surface height gradients and their associated geostrophic responses drive the spatial variability in the EKWC, particularly during warmer seasons. The apparent isolation of winter events from other seasons further points to distinct mechanistic differences in the cold-season dynamics. This study not only reinforces the value of spectral clustering in oceanographic research but also establishes a foundation for future investigations into the EKWC’s sensitivity to climate change and variability.