We use two global datasets of sedimentary microplankton, one with dinoflagellate cysts (dinocysts; ) and the second with planktic foraminifera . We use the surface sediment samples from sites south of 65∘ N (2849 and 4017 sites for the dinocysts and foraminifera respectively) because the OFES ocean model (which is used for particle advection; see below) ends at 75∘ N, which makes the clustering results at high northern latitudes unreliable. For some statistical analyses, we only consider sites in the Southern Hemisphere (725 and 1858 sites for respectively the dinocysts and foraminifera), in order to limit the total diversity of microplankton species in the datasets. We consider the fraction (i.e. the relative abundance) of microplankton species for every surface sediment sample.

The particle tracking results from provide us with distributions of surface origin locations for a global 1∘×1∘ grid of sediment sites, for several sinking speeds. We use the results that are obtained in the eddying OFES model with a sinking speed of 6 m d-1. Results with a sinking speed of 11, 25, and 250 m d-1 can be found in the Supplement. Values of 6 and 11 m d-1 are representative of single sinking dinocysts, 25 m d-1 representative of small aggregates, and 250 m d-1 representative of large aggregates and planktic foraminifera . Sinking speeds lower than 6 m d-1 can occur (e.g. due to oxidation of organic material and the development of gas within a shell), which may have an effect on the computed clusters. However, sinking speeds lower than 6 m d-1 are not tested in this paper because the backtracking method is computationally infeasible at lower sinking speeds due to long particle travel times.

The sinking speeds and backtracking analysis from are specifically designed to be compatible with the life cycle of dinocysts : particles are released at the bottom of the ocean every 5 d and tracked back in time until they reach 10 m depth, providing a particle distribution at the ocean surface (Fig. a). Single foraminifera typically sink at higher velocities than dinocysts (≳100 m d-1), and most of their lateral transport occurs during their lifespan, when they are passively advected while they control their buoyancy and remain at their preferential depth . However, we assume in this paper that the strength, direction, and “mixing” of planktic foraminifera by ocean currents has a similar spatially varying character compared to sinking dinocysts. We test whether the clustering results match both dinocyst and foraminifera sample datasets.

Our goal is to obtain provinces of sediment sites from the backtracked surface origin locations which are oceanographically (1) disconnected (i.e. provinces between which particles are not likely to travel) and (2) isolated (i.e. provinces with sediment sites which share similar origin locations compared to the sediment sites outside of the province). We quantify these areas by disconnected and isolated clusters of sedimentary sites. Assuming that the flow from 2000 to 2005, as simulated by the OFES model, is representative of the real ocean flow in the past decades (during which the microplankton actually sedimented; ), we ideally find the disconnectedness and isolation of clusters in the surface sediment sample datasets.

We use two types of clustering techniques. First, hierarchical clustering provides boundaries where sinking particles are less likely to cross (hence it finds oceanographically disconnected areas). This technique starts with the full ocean as only cluster and splits a cluster into two clusters at every iteration (see Appendix  for more details). The clusters from this technique can be compared to areas that are known to be oceanographically (dis)connected from each other, and these clusters can be used to test if more similar microplankton species are found within each connected area compared to between connected areas. Advantages of the hierarchical clustering method are that the cluster structure is preserved as more iterations are applied, and it does not require many parameters. The only parameter that the hierarchical clustering uses is the stop criterion (i.e. the iteration number where the algorithm stops with creating new clusters).

Second, we use the Ordering Points To Identify the Clustering Structure (OPTICS) algorithm to find oceanographically isolated clusters. OPTICS provides a density-based value (the reachability) of sedimentary sites which quantifies how strongly a site is connected to other sites. Oceanographically isolated clusters can be obtained from the “dense regions” (i.e. areas with low reachability values), by setting a threshold on the slope that surrounds the dense values in the reachability plot (ξ; see Fig. b for an example). The sediment sites outside of these clusters are less isolated and referred to as “noisy”. These clusters allow us to test if sedimentary species compositions are more homogeneous inside isolated areas compared to outside of these areas (see Fig. ).

The advantage of OPTICS is that parameter values have a clear interpretation. First, the parameter smin is the minimum number of particle release locations in clusters, which represents a minimum spatial scale of clusters (in m2). The second parameter ξ determines the degree of isolation of the clustering: OPTICS generally finds fewer and smaller clusters if ξ is larger. Another advantage of OPTICS is that not all areas are clustered, such that it allows us to distinguish between noisy (not clustered) and oceanographically isolated (clustered) areas (see Appendix  for more details about OPTICS).

We also test the seasonal dependence of hierarchical clusters (see Figs. S3 and S4), by only considering particles that started sinking in a specific season (in the backtracking analysis of Fig. a). While some of the cluster boundaries changed between summer and winter, the change in the overall clustering structure was limited if only a specific season of origin locations was considered (similar to , who only found a small seasonal effect in temperature offsets due to lateral transport of foraminifera).

We apply several statistical tools to test hypotheses about the sediment sample sites and the clusters in which they are located. A partial Mantel test is used to test whether the reachability from the OPTICS algorithm correlates with the sediment sample taxonomy, independent of the spatial distance between sediment sample site locations. A partial Mantel test requires at least three types of distance matrices, which contain distances between the sediment sample sites. We calculate the Mantel correlation between taxonomic distance and a distance which is determined from the reachability of the OPTICS clustering (see Appendix ), while we remove the effect of the spatial distance between sites on this correlation. If the Mantel correlation is significantly positive between these variables while removing the effect of the spatial distance metric between the sites, then this correlation is independent of the spatial distance.

We use canonical correspondence analysis (CCA; ) to infer the relation between species in clustered sediment sites and environment parameters at the ocean surface. In this context, CCA ideally shows unique species responses to changes in environment input parameters. We use sea surface temperature (SST) and surface nitrate (NO3) as environmental parameters, as prior literature reported them to explain a major part of the species variability in the Southern Ocean . This study infers SST and NO3 for sediment sample locations from 1∘×1∘ fields of the World Ocean Atlas . Further parameters, such as phosphorus, silicate, and salt concentration, were tested (as in ), though a spurious CCA response led to their exclusion from further analysis in this paper.

We compare the CCA's explained variation in sedimentary samples (1) only drawn from only isolated clusters and (2) with samples drawn from all available locations. Comparing the explained species variation in both cases allows us to draw conclusions about the source of variation between both CCA results in order to quantify the significance of the clustering approach. We apply a one-sided randomisation test to investigate whether the increase in explained variance is significant. This implies that we randomly take subsamples of the full dataset, which are equally sized to the number of clustered sediment samples. The p value of the permutation test is the fraction of random subsamples that resulted in a higher explained variance compared to the CCA analysis with the clustered samples.

The foraminifera dataset also contains deep-dwelling species, which live near the thermocline (typically a few 100 m depth). Although it is often assumed that these deep-dwelling species relate to sea surface variables in statistical analyses, this assumption might not be valid . We applied the CCA analysis while only using the species which are known to be near-surface dwelling in the subtropical Atlantic (the red group in Fig. 7 of ; supporting Fig. S11 this paper). This leads to similar conclusions, although less significant values are obtained because the dataset size is lower.

The clusters that are obtained from the OPTICS algorithm represent areas of relative oceanographic isolation. We test whether the species distributions in sediments outside clusters are more mixed compared than samples inside clusters during their sinking journey. We use Shannon entropy to quantify taxonomic mixing, which is defined at site j as Nsj=-∑ipijln⁡(pij). Here pij denotes the relative abundance of a species i at site j. Shannon entropy is often used as a biodiversity index , being a combined signal of species richness (number of species in the sediment sample) and evenness (how evenly these species are distributed). We choose the Shannon entropy here as biodiversity index because it can be compared to the mixing of sinking particles, and Shannon entropy is often used to quantify the loss of information by mixing (e.g. in thermodynamics). We compare the average Shannon entropy of sediment sample sites within (N‾sc) and outside (N‾snc) clusters.

We interpret splits of oceanographically disconnected clusters from the hierarchical clustering method as boundaries with a low connectivity across them (Fig. ). The probability that particles cross these boundaries is larger if the iteration number is higher. We observe that these cluster edges compare well to large-scale ocean connectivity. The first iteration splits the Mediterranean Sea from the global ocean because no sinking particles travel through the Strait of Gibraltar in the simulation. Next, the Pacific is separated from the Arctic, since few particles are transported through the Bering Strait . At the subsequent iterations, the large-scale ocean basins disconnect: the Pacific, Atlantic, and Indian oceans are split from the Southern Ocean at approximately 25∘ S. We find that areas near western boundary currents may only split into clusters at relatively high iterations because the sediments in these areas have a relatively large connectivity, with particles originating from a large area (see also ).

In the North Atlantic region, we observe that Hudson Bay becomes a cluster and the subtropical Atlantic is split from the Nordic Seas at the Greenland–Scotland ridge . The Irminger Basin is still connected with the Labrador Sea, where sinking of water occurs , which makes transport of sinking particles outside of this area less likely. Only a few particles cross the connection between the Labrador Sea and Baffin Bay .

We do not find a cluster at subpolar latitudes, which isolates Antarctica along latitudinal bands (only the Weddell and the Ross Sea are a cluster). One might expect such a cluster because near-surface currents are known to isolate Antarctica . However, deep passive particles advected by three-dimensional flow are shown to move upwards along isopycnals towards Antarctica . As a result, the sinking particles can be transported towards Antarctica at depth. The location of southward particle transport is mainly determined by topographic steering of the flow, resulting in five hotspots of southward particle transport which roughly coincide with the Southern Ocean clusters in Fig. .

Some of the clusters in Fig.  are similar to connected regions based on the surface flow (see Fig. 8 from ). The North and South Atlantic are split similarly from west Africa to Venezuela. The North Pacific and South Pacific are split in a similar way from Australia to the south of Chile. A cluster around the Pacific cold tongue (East Tropical Pacific) develops . Moreover, the Benguela upwelling area (near south-west Africa) is more connected with the Southern Ocean than with the Atlantic. Near-surface currents have an important influence on the total lateral transport of sinking particles in these areas, since they are similar to the surface connectivity areas from .

We test whether sites within hierarchical clusters have a lower (Euclidean) taxonomic distance compared to sites of different clusters with analysis of similarities (ANOSIM; ). The clustering corresponds to a high statistical significance for (1) positive ANOSIM test statistic together with (2) low p values (p value < 0.001). According to the ANOSIM tests, the clustering of sediment samples is significant across all iterations (Fig. c). Hence, sediment samples within clusters are more similar than those between clusters. The test statistic increases at higher iteration numbers, for both the dinoflagellate cyst (dinocyst) and the foraminifera dataset. Although these ANOSIM results look promising, it is important to note that the ANOSIM test statistics are partly positive because the sediment sites within clusters are closer to each other (i.e. there is a distance effect independent of the clustering).

The hierarchical clustering is overall insensitive to the sinking speed used of particles (see Figs. S1 and S2 in the Supplement with sinking speeds 11 and 25 m d-1 respectively; see Appendix  for an explanation on why we did not test a sinking speed of 250 m d-1). Only some minor differences occur in the North Pacific, and some cluster separations occur at slightly different iteration numbers. The fact that similar boundaries of little cross-transport emerge at a different sinking speed proves that the clustering does not greatly depend on the sinking speed of particles.

The OPTICS clustering algorithm provides a density-based value (the reachability) of sedimentary sites, which quantifies how strongly a site is connected to other sites. Oceanographically isolated clusters can be obtained from the dense regions (i.e. areas with low reachability values), by setting a threshold on the slope that surrounds the dense values in the reachability plot (ξ). The sediment sites outside of these clusters are less isolated and referred to as noisy.

According to the OPTICS algorithm, the western boundary currents are unlikely to be part of any isolated cluster, since the points near the western boundary currents have a relatively high reachability (Fig. a). This is expected because the origin locations of sediment samples near western boundary currents comprise a large area. Dense areas are those at higher latitudes, close to Antarctica and in the Nordic Seas, and the midlatitude gyres. Sediment sample sites within these areas have more similar surface origin locations compared to sediment sites outside of dense areas. The high reachability values in the Mediterranean Sea and Red Sea are rather artificial. OPTICS searches for dense regions (low reachability) by searching for sedimentary sites with relatively many other sites having similar surface origin locations. Since the Mediterranean Sea and Red Sea are enclosed by land, these sedimentary sites have only few neighbouring sites, resulting in a relatively high reachability.

The reachability distance (Dr; see Appendix ) between sediment sample sites correlates positively with sediment sample taxonomy. Furthermore, this correlation is independent of the spatial distance between sites, according to the partial Mantel tests (Fig. c). This means that oceanographically connected sites have a similar taxonomy, independent of their spatial distance. Large values of smin (>600; i.e. the OPTICS parameter which determines the minimum surface area in square metres of OPTICS clusters) tend to have the largest correlation (also at other sinking speeds; see Appendix ). This is probably because the reachability is smoother at higher smin, which makes the reachability distance less noisy. At small spatial scales (smin≤200), the correlation between dinocyst taxonomy and Dr could be indirect because then Dr correlates more strongly with the environment (in terms of sea surface temperature) compared to taxonomy.

We compute clusters by setting a threshold on the slope (ξ) that surrounds the reachability valleys in Fig. a. For ξ=0.002 and smin=300 (Fig. ), we obtain 13 clusters, of which three regions are isolated by the Antarctic Circumpolar Current (ACC), three in the Indian ocean, one in the Pacific warm pool, the South Atlantic gyre, near the Humboldt upwelling zone, near the Caribbean Sea, the eastern North Atlantic and two clusters near the Arctic. The clusters represent locations that are oceanographically isolated, with sediment sample sites that have backtracked origin locations which are similar to the other sites in the cluster. The environmental variability within these clusters can be reasonably large (e.g. the sea surface temperatures at backtracked origin locations in the cluster west of Australia range between 10–25 ∘C; see https://planktondrift.science.uu.nl/, last access: 14 February 2022).

The comparison between the clusters and taxonomic distance of Southern Hemisphere sample sites in these clusters (Fig. c and d) becomes interesting for clusters which are spatially close (Fig. b). For instance, the red and yellow clusters in the South Atlantic Ocean are spatially close, but sediment samples in those clusters are separated by their observed dinocyst taxonomy (Fig. c). This implies that we find a signal of the oceanographic separation of these areas in the sedimentary data. If noisy sites (such as the noisy site in Fig. b) were part of clusters, sites in different clusters are likely to contain a similar microplankton composition and the taxonomic separation of clusters is unclear.

To test if sedimentary sites within clusters are better correlated with environmental conditions at the surface, we applied CCA either including or excluding the sedimentary sites outside the isolated clusters (Fig. ). We find that the amount of explained variation by the canonical axes increases significantly if noisy sediment samples are excluded for the foraminifera (∼0.92 to ∼0.95) and especially for the dinocysts (∼0.82 to ∼0.98), for the same OPTICS clusters as in Fig. . Hence, we find that the linear relationship between environmental variables and microplankton composition of the CCA explains a larger part of the sedimentary species composition if noisy sites are excluded. In that sense, the signal is “cleaner” for sediment sample sites within compared to outside clusters, which has implications for palaeoceanographic reconstructions of SST with these sedimentary data.

The robustness of the relationship between sedimentary sites and environmental variables is investigated by testing the sensitivity of the CCA results to these parameters ξ and smin (Fig. ). For the dinocyst dataset we find an increase in explained variation by the canonical axes for most tested values of smin and ξ. By increasing the reachability slope that surrounds the oceanographically isolated clusters (ξ), a higher constraint is put on the isolation of these clusters and fewer sediment sample sites are part of a cluster. If this slope ξ is chosen too high, no clusters exist or they are too small to contain any sediment sample sites at all. Moreover, a higher value of ξ means that fewer sedimentary sites are used in the CCA (i.e. the dataset size is reduced), which may lead to an insignificant result according to the randomisation test. A relatively low value of ξ on the other hand may lead to insignificant results due to the inclusion of noisy sites in clusters. Hence, there seems to be an optimal value ξ, for which this increase in variance is maximised. These results are more often insignificant for the dinocysts compared to the foraminifera because the dinocyst dataset is smaller. If smin is higher (i.e. the OPTICS algorithm finds larger clusters; in km2), the negative and insignificant values in Fig.  are partly caused by including noisy sites in clusters. These results highlight the importance of choosing an appropriate combination of ξ and smin for the CCA to show a significant increased explained species variability if only clustered sites are used.

The clustered samples are less taxonomically mixed (i.e. are less biodiverse) for most values of ξ and smin, as we can see from the comparison between the Shannon entropy (Ns; see the Method section) within and outside of clusters (Fig. ). For the high values of smin, we find that the clusters are less taxonomically mixed if ξ is increased. This result supports that measured microplankton biodiversity in sediments is relatively large in areas with strong mixing of sinking particles by ocean currents. However, the Shannon entropy is also influenced by the species distributions at the ocean surface, for which much fewer species composition data are available compared to their sedimentary remains. At smaller values of smin (smin<200 and smin<300 for the dinocysts and foraminifera respectively; Fig. ) high (surface) productivity areas are also clustered (e.g. the south-west Atlantic or the Humboldt area; see Fig. ). Hence, a relatively high sedimentary biodiversity in these clusters can be explained by the high biodiversity at the ocean surface, before these particles start sinking.

We also tested the OPTICS results for sinking velocities higher than 6 m d-1 (11, 25, and 250 m d-1; see Figs. S5–S10 and S11–S14). Similar clusters can be found with the other sinking velocities. A higher sinking speed decreases the particle travel time; thus the lateral transport and the mixing of sinking particles is overall lower . However, the spatial dependence of the lateral transport is similar: both at low and high sinking velocities, the lateral particle transport is relatively large near western boundary currents and low in the middle of midlatitude gyres . As a result, the clusters are located in similar areas for different sinking speeds. It is only the spatial scale of these clusters that might be different. The spatial scale (i.e. the size) of the clusters can again be controlled by the parameters smin and ξ. Note that high-productivity areas are less likely clustered at higher sinking speeds, which has implications for the Shannon entropy within and outside clusters (Fig. ): higher biodiversity is measured outside compared to within clusters for a larger area of smin and ξ values as the sinking speed increases.