A clustering analysis of eddies' spatial distribution in the South China Sea

Jiawei Yi, Yan Du, Xinyu Wang, Zhonghua He, Chenghu ZhouOcean 2013 · DOI

Abstract

Spatial variation is critical for understanding mesoscale eddies in the South China Sea (SCS). This study employs a K-means clustering approach to investigate these variations objectively. Analysis of the Hopkins Statistic revealed that while the clustering tendency of anticyclonic eddies (AEs) and cyclonic eddies (CEs) is weak ($H \approx 0.56$), it is statistically non-random. Finer clustering identified 10 regions of dense AE population and 6 for CEs. These data-driven partitions align with known physical generation mechanisms, such as monsoon forcing and topography. Comparisons show AEs are generally more aggregated than CEs, with distinct spatial separation observed southwest of Luzon Island and dipole structures identified off the coast of Vietnam.

Introduction

Mesoscale eddies are the "weather" of the ocean, serving as primary drivers for the transport of heat, salt, and nutrients. In the South China Sea (SCS), these swirling bodies of water are ubiquitous but exhibit high spatial variability. Traditionally, researchers have divided the SCS into four broad geographic zones to study these phenomena. However, these arbitrary boundaries often fail to align with the underlying physical mechanisms. This study seeks a more objective, data-driven method to identify eddy "birthplaces" and characterize their spatial distribution using unsupervised learning.

Methods and Data

The study utilized output from the 1/32° resolution global Naval Research Laboratory’s Layered Ocean Model (NLOM), spanning 2005 to 2011. This model, validated against AVISO satellite data and cruise measurements, provided a dataset of 371 tracked eddies, consisting of 159 AEs and 212 CEs.

Spatial distribution of cyclonic (CE) and anticyclonic (AE) eddies in the South China Sea study area, showing the raw data points used for clustering.

To analyze clustering, the researchers first calculated the Hopkins Statistic ($H$) to determine if the eddy birth locations were truly clustered or merely random. Subsequently, K-means clustering was applied. The optimal number of clusters ($K$) was determined by analyzing the Sum of Squared Error (SSE) and the Silhouette Coefficient (SC), balancing spatial detail with model simplicity.

Statistical Clustering Tendency

A key finding was the statistical validation of clustering. As shown in Figure 6, the Hopkins Statistic for both AEs and CEs consistently remained above the 0.5 threshold associated with random distributions.

Hopkins Statistic (H) analysis for (a) Anticyclonic, (b) Cyclonic, and (c) combined Eddies, demonstrating a weak but statistically significant clustering tendency.

While the clustering is considered "weak" ($H \approx 0.56$), it confirms that eddies are not randomly scattered. Interestingly, the analysis showed that AEs are generally more aggregated ("clumpier") than CEs. Furthermore, coarse clustering results ($K=4$) largely validated the traditional four-zone geographic divisions used in previous literature, providing a mathematical basis for those intuitive models.

Regional Insights and Hotspots

The optimized clustering revealed 10 AE hotspots and 6 CE hotspots, providing a much finer resolution than traditional models. Key regional findings include:

  • Luzon Island Dynamics: Southwest of Luzon, AEs and CEs form in distinct zones. AEs typically generate north of 14°N and closer to the coast, whereas CEs form further south and offshore.
  • The Vietnam Dipole: The analysis successfully identified the characteristic "dipole" structures off Vietnam—pairs of counter-rotating eddies driven by summer circulation patterns.
  • Topographic and Seasonal Influence: The Nansha Trough and central SCS were found to be dominated by AEs, highlighting the role of bottom topography. In the northeastern SCS, patterns remained stable across seasons due to the consistent influence of the Kuroshio intrusion.

Conclusion

By moving beyond arbitrary geographic boundaries, this study provides a mathematically validated map of eddy generation in the South China Sea. The results confirm that eddy distribution is a non-random process driven by complex interactions between monsoons, topography, and the Kuroshio intrusion. Future work will expand this analysis to include the spatiotemporal evolution of these clusters, tracking how these "ocean weather" systems move and dissipate over time.

Summary updated

Abstract

Spatial variation is important for studying the mesoscale eddies in the South China Sea (SCS). To investigate such spatial variations, this study made a clustering analysis on eddies' distribution using the K-means approach. Results showed that clustering tendency of anticyclonic eddies (AEs) and cyclonic eddies (CEs) were weak but not random, and the number of clusters were proved greater than four. Finer clustering results showed 10 regions where AEs densely populated and 6 regions for CEs in the SCS. Previous studies confirmed these partitions and possible generation mechanisms were related. Comparisons between AEs and CEs revealed that patterns of AE are relatively more aggregated than those of CE, and specific distinctions were summarized: (1) to the southwest of Luzon Island, AEs and CEs are generated spatially apart; AEs are likely located north of 14° N and closer to shore, while CEs are to the south and further offshore. (2) The central SCS and Nansha Trough are mostly dominated by AEs. (3) Along 112° E, clusters of AEs and CEs are located sequentially apart, and the pairs off Vietnam represent the dipole structures. (4) To the southwest of the Dongsha Islands, AEs are concentrated to the east of CEs. Overlaps of AEs and CEs in the northeastern and southern SCS were further examined considering seasonal variations. The northeastern overlap represented near-concentric distributions while the southern one was a mixed effect of seasonal variations, complex circulations and topography influences.
BibTeX
@inproceedings{yi2013clustering,
  author = {Jiawei Yi and Yan Du and Xinyu Wang and Zhonghua He and Chenghu Zhou},
  title = {{A clustering analysis of eddies' spatial distribution in the South China Sea}},
  booktitle = {Ocean},
  year = {2013},
  doi = {10.5194/os-9-171-2013},
}
Figure