| 27 May 2026
A Comparative Analysis of Clustering Algorithms for Characterizing Surface Ocean Variability in the Western Mediterranean
Abstract. Understanding regional dynamical structures in the sea is fundamental to characterize energy transfer and transport properties, with implications in physical and biogeochemical modeling and characterization. In this work, we study the potential of clustering techniques to identify regional patterns, persistent or recurrent configurations, out of daily snapshots of sea surface temperature and kinetic energy in a region of the western Mediterranean Sea. From the methodological perspective, we use different clustering techniques: K-means, Self-Organizing Maps and InfoMap to verify if the patterns found are coherent across methods. Our results show that K-means and Self-Organizing Maps consistently delineate four distinct clusters of sea surface temperature configurations, aligned with the seasons even after removing the annual cycle, which indicates the persistence of seasonal structures beyond a mean effect in the temperature field. The study of surface kinetic energy, characterized by higher spatial and temporal variability, reveals more complex circulation regimes. While K-means and Self-Organizing Maps provide a robust and convergent classification of the dominant large-scale energy patterns, InfoMap uncovers finer-scale features such as localized jets and eddies. InfoMap, in particular, provides a complementary perspective to the partition-based methods, validating subtle yet significant hydrodynamic structures and acting as an anomaly detector for extreme events.
This research work compares two to three clustering methods applied to SST and KE variables in the Western Mediterranean. The paper is generally well written and addresses an interesting scientific problem. However, it contains several methodological issues that prevent its publication in its present form. Therefore, I recommend a major revision before further consideration.
GENERAL COMMENTS (P = page; L = line)
I identify the following methodological issues:
** Section 3 and the subsequent sections rely on the selection of an optimal number of clusters (four in this case). This decision is based on the Silhouette Score values shown in Figure 2. However, this figure, together with the supporting analysis, lacks a minimal assessment of statistical significance. As a result, it is not possible to guarantee that selecting four clusters instead of, for instance, nine makes any meaningful difference. Since the remainder of the analysis presented in the paper depends on this particular choice, this issue is highly relevant and undermines the validity of the conclusions. A straightforward solution would be to apply a bootstrap strategy to both the K-means and SOM methodologies. This could be done by iteratively subsampling the 12996 data samples (e.g., 70%), applying the clustering techniques, and computing the corresponding Silhouette Scores. Confidence intervals could then be constructed and added to Figure 2 (and potentially Figure 3), allowing for a more robust selection of the optimal number of clusters, whether it is four or another value.
** The results presented in Section 3.2 indicate that the four identified clusters essentially reproduce the seasonal cycle, characterized by four patterns, but provide limited additional insight. If this is the intended objective, it should be stated explicitly. However, I believe the paper would offer greater scientific value if the methods were applied to data with the seasonal cycle removed. As it stands, it gives the impression that relatively complex techniques are being used to recover “only” a seasonal signal. Assuming a more elaborated approach based on deseasonalized data, the existence of the four pattern cycle suggests that a simple Day-of-Year (DOY) climatology may not be sufficient to remove the seasonal signal. In this context, a PCA retaining two or three EOFs could be an appropriate strategy. In parallel, no consideration is given to potential decadal trends in SST or their influence on the analysis, which should also be addressed.
** The inclusion of the InfoMap clustering technique should be reconsidered. The main concern is that the results are inconclusive (nearly 100% of samples are assigned to a single cluster). It is unclear whether including this method adds any value to the manuscript. Furthermore, it is applied only to one of the two variables (KE) without clear justification. Finally, the implementation details, particularly the choice of threshold (L156 and L323-325), appear arbitrary and lack proper justification.
I also consider the following point to be critical:
** Based on the current title of the manuscript, one would expect a broader range of clustering methodologies. However, only two to three methods are applied, and no discussion is provided regarding alternative approaches. This issue could be partially addressed by modifying the title. Nevertheless, at a minimum, the manuscript should include a discussion of other available clustering methods. Additionally, it may be worth reconsidering whether the term “Surface Ocean Variability” in the title is appropriate, or if a more specific formulation such as “SST and KE Variability” would be more accurate.
SPECIFIC COMMENTS (P==page; L==line)
P3L73: To what extent do the presented results depend on the selected data product? What are the limitations?
P6L145: It is not discussed why InfoMap is applied only to KE and not to SST
P7L156: See general comments regarding the threshold
P8L196-199: This is very speculative given the actual results
P10Figure4: Is this result expected, given that the seasonal cycle typically captures most of the SST variance? Also see general comments on the seasonal cycle
P10L221: Is this statement correct? Please explain how and why.
P12L250-251: How does this compare with the suggested PCA approach?
P14Section 3.3: Temporal distributions of clusters are missing in this analysis. Is there a reason?
P16L308: This statement is speculative (“significant…”)
P17Section 3.3.3: See general comments on the removal of this section
P18L347-349: “So “rare” events are “similar” enough to be classified into specific cluster(s)? Wouldn’t their nature place them outside any cluster?
P21L401: This is difficult to sustain given the limited number of techniques applied and the specific results obtained with InfoMap
P21L404-406: Specify what these implications are