Network analysis of sea turtle movements and connectivity: A tool for conservation prioritization

Understanding the spatial ecology of animal movements is a critical element in conserving long‐lived, highly mobile marine species. Analyzing networks developed from movements of six sea turtle species reveals marine connectivity and can help prioritize conservation efforts.

Understanding the spatial ecology of animals, particularly how movements of migratory animals link different areas, is a critical element in conserving highly mobile species Webster et al., 2002). For example, unfavorable conditions on adult foraging grounds can alter the phenology of migration and body condition, which influence reproductive frequency and success (Marra et al., 2006). Ecological connectivity between areas can shape the genetic variation that manifests in migratory behavior and, in turn, can influence population resilience (e.g., the ability to respond quickly to change; Bernhardt & Leslie, 2013;Carr et al., 2017;Olds et al., 2015;Webster et al., 2002). Describing and quantifying the spatial ecology of animal movements contributes to understanding the overall network of habitats and the specific roles that areas and connections have during various life-stages. These areas and connections (network components) can control the dynamics of area use, resource flow, and other ecological patterns (Cohen et al., 2018;Dunn et al., 2019;Treml & Halpin, 2012).
However, these techniques are limited in identifying all suitable and unsuitable habitats and, with the exception of tracking data, will not describe the "corridors" (pathways used to connect high use areas).
Tracking data can be used to reconstruct animal movements, though a common constraint occurs when small sample sizes are used to represent the corridor (Mazor et al., 2016). Additionally, information on the spatial connectivity can be difficult to apply to management measures at the appropriate scale, especially for highly migratory species utilizing areas across ocean basins (Balbar & Metaxas, 2019;Dunn et al., 2019;Shuter et al., 2011). Incorporating knowledge of connectivity in marine conservation is challenging due to (1) limited understanding of the scale and extent of connectivity necessary for population survival (Cowen et al. 2006;Foley et al., 2010;Sale et al. 2005), (2) a shortage of empirical data to conduct robust analyses (Mumby et al., 2011), and (3) non-standard ways to measure and define the ideal level of connectivity to maintain or conserve (Almany et al., 2009;Balbar & Metaxas, 2019;Keeley et al., 2021).
Networks are graphs that can display spatial connectivity within a structure composed of nodes (vertices that can represent available habitat) and links (edges that can represent landscape connectivity; Kindlmann & Burel, 2008). Functional connectivity within spatial networks can be displayed by animal movements represented as links between nodes that are high-use areas (Jacoby et al., 2012;Jacoby & Freeman, 2016;Kindlmann & Burel, 2008;Olds et al., 2015). As a first step, examining the topology or pattern within the structure of the network can be a qualitative way to visualize the role of individual nodes and links based on their position and how changes would affect functional connectivity (Bisht & Singh, 2015;Santra & Acharjya, 2013). Evaluating differences in topology could provide quick comparisons on conservation and maintenance needs across multiple networks (Bisht & Singh, 2015;Santra & Acharjya, 2013).
Graph theoretical methods can also use mathematical applications to study modeled relationships between node pairs within a network (Bunn et al., 2000;Rayfield et al., 2011;Urban & Keitt, 2001). Network metrics are used to quantitatively compare centrality (dominance or influence of an area or connection) and connectivity (the movement of organisms within paths) between and among the network node and link positions (Schick & Lindley, 2007;Treml & Halpin, 2012;Urban & Keitt, 2001). Standardized network metrics applied in a variety of other disciplines (e.g., engineering, social science, and chemistry) have been leveraged in ecology to determine flow of resources and identify sources, sinks, and isolated areas (Rayfield et al., 2011;Urban et al., 2009). These methods can also facilitate a better understanding of ecological connectivity on a global and local scale, such as evaluating the amount of resources transferred among key habitats of highly migratory species in marine environments (Jacoby & Freeman, 2016). Graph theory has a large potential to help prioritize conservation measures within or across key areas at different spatial and temporal scales.
Spatially explicit networks can distil complex movement patterns into "connections" that are more easily communicated to policymakers, helping bridge the knowledge gap in ecological connectivity for the implementation of appropriate conservation measures. The number of satellite tracking studies is increasing for marine species (Hussey et al., 2015), along with the ability to synthesize data across large spatial extents (Block et al., 2011;Davidson et al., 2020;Ropert-Coudert et al., 2020;Sequeira et al., 2019), but graph theory applications have not yet been common practice for examining movements (Jacoby et al., 2012;Jacoby & Freeman, 2016). Here, we applied graph theory to spatial networks constructed from the movements of 1235 individual sea turtles from six species to highlight marine connectivity in support of conservation prioritization.
Using this large dataset, we described and compared the structure and function among several networks. Last, we proposed recommendations on (1) how graph theory concepts can facilitate analyses on emerging connectivity patterns in networks created by highly migratory species, (2) applications for the prioritizing conservation efforts within and across regions and species, and (3) collaborative methods for gathering existing data to support the synthesis of knowledge on connectivity in the ocean.

| Sea turtle movement and connectivity data
We conducted a formal literature review on sea turtle movements and connectivity and summarized the common methods presented in publications (see Appendix S1). Relevant references from the literature review were used to determine how representative the telemetry dataset was that was collated for this study compared to the published studies. From March 2018 to April 2019, telemetry tracking data on sea turtle movements were requested from data providers for the State of the World's Sea Turtles (SWOT), the seaturtle.org tracking listserv, and the CTURTLE listserv. Most of the data collated for this study originated from online data archives (e.g., Coyne & Godley, 2005;Halpin et al., 2006;Kot et al., 2018).
No telemetry data on flatback sea turtle (Natator depressus) were directly contributed.
A total of 1235 individuals, from six sea turtle species, included in this study were tracked with Advanced Research and Global Observation Satellite (ARGOS) tags receiving ARGOS doppler or global positioning satellite (GPS) locations (Table 1; see Appendix   S2). For all ARGOS doppler and global GPS data, points on land (using Global Administrative Areas, 2018), erroneous or low-accuracy locations, and animals with only one location recorded were removed.
For locations that were not estimated using a state-space model, a speed, distance, and angle (SDA) filter was applied using the R package argosfilter (Freitas, 2012;Freitas et al., 2008;R Core Team, 2019), where all GPS data were assigned the highest location quality value (LC = 3, estimated error <250 m), following Thums et al. (2017) and Wildermann et al. (2019). We applied a conservative speed filter of 10 km/h as the maximum speed for all six species (see Eckert, 2002;Luschi et al., 1998;Walcott et al., 2012;Wirsing et al., 2008).
Finally, to better normalize data with timestamp irregularities and tags programmed to record locations on different schedules, we subsampled each track to retain just one location per day (De Solla et al., 1999). The highest quality record of the day was chosen; the first occurrence of the high-quality points was retained if multiple locations of the same highest quality level was recorded for one day (per Metcalfe et al., 2020;Witt et al., 2010). All types of sea turtle movements from one point to another were used in this study, such as those described by Dingle and Drake (2007) as short-and long-distance migrations, ranging, foraging, commuting, or seasonal movements "between habitat regions."

| Sea turtle movement and connectivity networks
We used all collated tracking data to create ten spatial networks: one global network for all data, three regionally connected networks (herein called the "Atlantic-Indian," "Mediterranean," and "Pacific" components), and six species-specific networks ( . We a priori used the centroids of marine regions that were frequently used for management purposes as network nodes (points within the graph; see Appendix S3). Marine regions included high seas (areas beyond national jurisdiction), exclusive economic zones (EEZs), and sub-regions of EEZs (Flanders Marine Institute, 2018a, 2018b. Each tracking data point was assigned to the intersecting marine region polygon that varied greatly in size (i.e., high-seas region, EEZ, or sub-region of an EEZ), which was then represented by the polygon's centroid using ESRI ArcMap 10.6.1 (ESRI, 2019). Marine region centroids were not the same as specific locations of animals, and nodes could fall outside of known geographic extents of sea turtles. Tracking data network links (arcs within the network) were created representing connectivity between two nodes when at least one animal moved from one marine region to another (see Appendix S4); links within one node (self-loops) were eliminated. Links were not representative of the specific path sea turtles travelled, and lines could be longer or shorter than distance travelled. Rather, mapping the nodes and links helped visualize their geographic arrangement within each network for qualitative comparisons on structure and function among different networks.
We also created four "neighbor networks" based on the relative geographic location of nodes, following Bange and Hoefer (1976), to TA B L E 1 Tracking data summarized by species, sex, age class, number of locations, and temporal coverage. Species: Cc serve as a baseline network of structural connectivity when assuming that neighboring areas sharing borders were connected. Neighbor networks were created using all possible nodes (n = 227 marine regions with at least one tracking data point) and links that represented connections between marine regions that shared a border (ESRI ArcMap 10.6.1 Polygon Neighbors Tool; ESRI, 2019). The structural connectivity of one global neighbor network and three regional neighbor networks (Atlantic-Indian, Mediterranean, and Pacific components) was compared directly with the functional connectivity of networks created using tracking data (herein called "tracking data networks").

| Network metrics selection and analyses
We selected centrality metrics within the current literature that 2013) from which a subset was then selected for further application to our study. For tracking data networks, node weights (based on the number of animals at each node), link weights (based on the number of routes observed), and link direction were applied whenever possible in calculating metrics. Routes were counted as the number of times animals travelled from one marine region to the other. High numbers of animals were assumed to represent high flow, and abundances were normalized within five categories using the Jenks natural breaks classification method (Jenks, 1963(Jenks, , 1967 for all data and within species. For the neighbor networks, marine regions and connections were normalized to minimize potential differences due to weight and direction by treating links as undirected and attributing the overall average number of animals, average number of routes, and median weight to individual nodes and links. To refine our set of metrics, principal component analysis (PCA; Hotelling, 1933;Pearson, 1901) was first used to determine a subset of metrics that contributed significantly to the total variance of the tracking data network using the 80% cumulative sum as a threshold (R packages factoextra and FactoMineR; Kassambara & Mundt, 2016;Lê et al., 2008). Second, uncorrelated metrics (Spearman's rho <0.80) from the subset were retained to test for significant differences among region and species-specific networks using the Kruskal-Wallis (KW)

test and post hoc Dunn's test with
Bonferroni adjustment for multiple comparisons (α = .05; R package FSA; Bonferroni, 1936;Dunn, 1964;Kendall, 1938;Kruskal & Wallis, 1952). These nonparametric tests were used because metrics were not normally distributed (Shapiro-Wilk normality test, p > .05; Shapiro & Wilk, 1965) or comparisons were made with a small sample size. Finally, the 14 node and link centrality metrics that showed the greatest differences (lowest KW test p-value) among networks were selected to describe network properties (see Appendix S5).
Global-level metrics were calculated to give an overall connectivity metric that would take into account all components, including disconnected sub-networks and isolated nodes. Node degree (number of links, in all directions, connected to the node) and the number of links (calculated using links with direction) were the global level metrics that were retained to describe global connectivity because they contributed significantly (PCA sum >80%) and were uncorrelated with other metrics (Spearman's rho <0.80). In general, both degree and the number of links related positively to overall network connectivity, where greater values corresponded to higher levels of network flux or movement (Borgatti, 2005;Rayfield et al., 2011).
However, these global metrics could not be statistically compared among the whole region or whole species networks because of low statistical power (only one network per species and region was created, and not enough data were available for a longitudinal analysis).
Local-level metrics that displayed the greatest difference among regions were node "closeness" and link "betweenness." The node closeness metric is related to "how close a node is to all the other nodes in the network beyond ones that are directly connected to" where the distance is the shortest path; a node with greater closeness that can reach all other nodes quickly is more independent than other nodes (Kim et al., 2011). Node closeness is ecologically significant to determine areas that are relatively central that may be more critical to conserve to maintain the network (Estrada & Bodin, 2008;Jacoby & Freeman, 2016;Ospina-Alvarez et al., 2020). The link betweenness metrics is related to "the number of shortest paths between pairs of vertices that run along it"; links that are highly relied upon to connect different communities or at the borders of communities can have greater betweenness (Girvan & Newman, 2002). Link betweenness is ecologically significant to determine links acting as stepping stones that may be more critical to conserve to maintain the network (Estrada & Bodin, 2008;Jacoby & Freeman, 2016;Ospina-Alvarez et al., 2020). Node closeness and link betweenness metrics were used here as proxies for measuring two network properties, respectively: connectedness (the ability to spread resources using short distances to other nodes; Muller & Peres, 2019) and vulnerability (the network's ability to continue functioning given changes to network structure, such as the removal of links; Aytaç & Öztürk, 2018). These metrics were complementary because the node closeness metric is limited to only calculating relationships and distances within connected components, while link betweenness can be calculated over networks that contain connected and disconnected components (Baranyi et al., 2011;Borgatti, 2005;Wasserman & Faust, 1994).

| Sea turtle movement and connectivity data
There was a considerable amount of information on global sea turtle movements and connectivity presented within the relevant literature (n = 969 references), and very little overlap with references attributed to the telemetry data collated for this study. About 92% of the references presenting relevant telemetry data were unique to the literature review, and 40% of references attributed to the telemetry data used in this study were found within the literature review.
Over 64% of all relevant publications from the literature review presented data from telemetry or capture-mark-recapture methods, though a diversity of methods were used (see Appendix S1). Most networks exhibited complex arrangements of nodes and links within connected components, typically described as "hybrid topologies" (Figures 3 and 4). These networks contained a combination of node and link patterns that can involve point-to-point (nodes connected along a single path), ring (nodes connected within a circular path), star (a single centere node acts as a hub when connecting to several different nodes), and mesh formations (multiple nodes are connected in multiple ways; Bisht & Singh, 2015;Santra & Acharjya, 2013). The Mediterranean component exhibited a hybrid topology closer to a mesh; the Atlantic-Indian, Pacific, most species networks included star topologies within their hybrid form with high-sea marine regions as highly connected hubs (Figure 3). Nodes identified as hubs have relatively high number of links and can act as "super-spreaders." On the other hand, the Kemp's ridley sea turtle network was the only one that displayed a single connected component displaying a point-to-point topology, the simplest structure that has a relatively low number of links (Figure 4e).

| Regionally connected networks
Comparisons among regionally connected networks showed that the tracking data network had significantly higher node closeness

| Species-specific networks
Compared to other sea turtle species, the closeness of marine region nodes was significantly higher within hawksbill, Kemp's ridley, and olive ridley sea turtle networks (KW test, p < .05, see Appendix S5). Maximum closeness was found for the Italian part of the Mediterranean eastern basin node within the Kemp's ridley sea turtle network, the network with the smallest number of individuals tracked and nodes with the lowest link betweenness, resulting from a minimal number of links along a straight path (Figure 4). On the other hand, the loggerhead sea turtle network with the largest number of individuals tracked created a network with the largest F I G U R E 1 Tracking data for six sea turtle species summarized by (a) number of locations within a hexagonal grid, and (b) number of animals per marine region and number of routes taken between marine regions within a network diagram using marine region centroids as nodes. Grid cell = 1.62 × 10 5 km hexagon; stars symbolize the centroid of individual high-sea marine regions, not animal point location; circles symbolize the centroid of individual marine regions within national jurisdictions, not animal location; links represent connections, not animal paths; two overlapping links between nodes represent connections in both directions. Data were classified using natural breaks (Jenks) within each panel; warmer colors represent higher values and cooler colors represent lower values; number of map features in parenthesis. For more information, see Appendices S1-S4 F I G U R E 2 Summary of tracking data for six sea turtle species within a hexagonal grid (left) and within a network diagram using marine region centroids and connecting links (right) for (a-b) Caretta caretta, (c-d) Chelonia mydas, (e-f) Dermochelys coriacea, (g-h) Eretmochelys imbricata, (i-j) Lepidochelys kempii, and (k-l) Lepidochelys olivacea. Grid cell = 1.62 × 10 5 km hexagon; stars symbolize the centroid of individual high-sea marine regions, not animal point location; circles symbolize the centroid of individual marine regions within national jurisdictions, not animal location; links represent connections, not animal paths; two overlapping links between nodes represent connections in both directions. Data classified using natural breaks (Jenks) within each panel; warmer colors represent higher values and cooler colors represent lower values; number of map features in parenthesis. For more information, see Appendices S1-S4 Within sea turtle species networks, regions with relatively high node closeness also had high link betweenness, but these regions differed across species (Figure 4). The loggerhead sea turtle network was the only species that showed high node closeness and link betweenness in the Pacific Ocean, specifically in the Japanese EEZs, turtles rely more on oceanic habitats than other sea turtle species (Bolten, 2013;Luschi et al., 2003). In contrast, marine regions and links within EEZs were highly centralized within the loggerhead,

| Conservation recommendations
Progress has been slow in the explicit use and application of connectivity research for managing the marine ecosystem (Balbar & Metaxas, 2019;Carr et al., 2020). However, our results provided a better understanding on connectivity to further any advances and contribute to future conservation proposals that can appropriately cover sea turtle movements throughout their life cycle, following the success of others using tracking data to influence marine policy (see Davies et al., 2021;Hays et al., 2019). As international efforts continue in the development of spatial management plans in the high seas, areas that lack the most data (Ardron et al., 2008;Dunn et  topology) may be more vulnerable to any network changes without the availability of alternative nodes or links as a backup.
However, changes such as loss of a node or link within a pointto-point network would be relatively easy to detect and manage, and given that network changes would be within a more direct path, effects could be easier to predict, and the needs to restore network function could be quickly addressed (Bisht & Singh, 2015;Forouzan & Fegan, 2007;Santra & Acharjya, 2013). Conversely, complex network topology containing many links relative to nodes may be less vulnerable to changes, given the options for alternate paths, but can require more effort to identify where resources should be applied to best conserve network function. It may be more difficult to prioritize among multiple nodes with relatively high numbers of links (hub centres) or several links identified as bottlenecks (connections without alternative paths). Examining the roles of nodes within node sets or multiple nodes within the network can also provide support for the most efficient selection of priority areas necessary within reserve networks after identifying habitats and omitting redundant nodes from conservation efforts (Pereira, 2018;Pereira et al., 2017). Many critical hubs in the high seas were identified within the loggerhead (North Pacific and Philippine Sea), green (Arabian and Indian Oceans), and leatherback (North Atlantic) sea turtle networks, emphasizing the need to include areas beyond national jurisdictions in conservation and prioritization efforts for these species.
Network node and link metrics can also facilitate conservation strategies on a more regional or local level (Beger et al., 2015;Mazor et al., 2013). Networks with higher node closeness may indicate that benefits from area-based conservation measures within marine regions could spread to other highly connected marine regions, especially when directed toward important areas.

Examples of priority marine regions included those overlapping
EEZs in the northwest Pacific for the loggerhead sea turtle network and within EEZs in the Persian Gulf/Gulf of Oman areas for the green sea turtle network. On the other hand, networks with relatively high link betweenness may benefit more from preserving seascape connectivity among marine regions (e.g., protecting migratory corridors) to increase overall stability within populations that rely on connections. Policies to conserve network structure and function may be more critical for these populations, such as leatherback sea turtles across the Atlantic.
The notion that marine geographic proximity generally relates to similar ecological habitats that drive species area-use (Spalding et al., 2007) was confirmed by the agreement in relative levels of connectivity among regions in this study. For example, the geography and sea turtle movements within Mediterranean marine regions resulted in a network with relatively high node closeness and link betweenness, contributing to high connectedness and vulnerability. In other regions, tracking data networks were less connected (lower median closeness) and more vulnerable (higher median link betweenness), indicating that dependence on sea turtle movements to inform connectivity may be more critical than relying on geography alone.
Because of these network differences, empirical data on sea turtle movements need to be directly incorporated into marine network conservation and management.

| Methods considerations
Our approach was to include as much of the available telemetry data as possible, providing a greater overview of spatial connectivity through networks on a global and regional scale, even when our dataset was not fully representative of all published studies.
Therefore, the resulting maps and networks were developed using the best available data, and many details on the individual sea turtles included in this study were unknown or unavailable. Characteristics such as sex, age-class, life-history stage, health condition, core area utilized, amount of time spent in a specific area, the availability of habitats, or movement types (e.g., foraging, inter-nesting, migrating, ranging), were likely to be highly variable within our study sample.
We also recognized that the extent and distribution of our study  (Mazor et al., 2016). Future research is needed to properly correct for sampling biases or appropriately weigh data (e.g., according to location from tagging site, distinct research methods, groups of individuals, activities, or habitat quality) to evaluate network patterns based on ecological niches, specializations, and biological traits (Block et al., 2011;Brodie et al., 2018;Hays et al., 2020;Jacoby et al., 2012;Queiroz et al., 2019). Network analysis methods developed in our study were applicable at various scales, allowing future comparisons at finer resolutions or longitudinal analysis, given the appropriate data. Lédée et al. (2015) found that network analysis added valuable information to kernel-based methods, revealing important movements within and between core habitats. However, within the current study, node closeness and link betweenness values that were scaled within global or species-specific networks were not directly related to the number of individuals or abundance of locations recorded within marine regions. Furthermore, high seas nodes represented relatively large marine regions (e.g., North Atlantic Ocean, South Atlantic Ocean, North Pacific Ocean, and South Pacific Ocean) compared to EEZs, and false connections between coastal EEZ marine regions via the high seas may be revealed or overemphasized due to higher probability of falling within a larger area. For example, olive ridleys in this analysis appeared connected across the South Atlantic, but tracking data from Pikesley et al. (2013) indicated that they spend limited time in the high seas and have fidelity to either side of the Atlantic. Therefore, we recommend supplementing results from more traditional approaches (e.g., abundance and density estimates) or higher-resolution animal movement data with network analyses to identify important high-use areas.

| Datasets for "big data" analyses
Many studies reviewed within the literature were limited to describing areas used for specific behaviors (e.g., female nesting), lacking details on how multiple marine areas were connected by individual animal movements. Among the relevant data types, data from telemetry tags were most effective in describing core area use, spatial patterns, networks, and migratory connectivity, especially when studies incorporated capture-mark-recapture, stable isotope, or genetic analyses (e.g., Godley et al., 2008;Hart et al., 2015;Hays & Hawkes, 2018;Haywood et al., 2020;McClellan et al., 2010;Rees et al., 2017;Stewart et al., 2013). While a small proportion (about 5%) of the reviewed literature combined telemetry data with another research method (e.g., capture-mark-recapture, genetic, stable isotope analyses), the trend in integrating approaches has increased in recent years. A better understanding of the geographic extent of global sea turtle movements and area use could result from future studies that either include a combination of approaches or encourage data sharing to compare with data collected by different tools.
Benefits of collating primary datasets for "big data" animal movement analyses have been widely recognized and can be a major driver for standardizing data collection, analyses, and archiving methods (Sequeira et al., , 2021Thums et al., 2018). Even with a largely inclusive approach for gathering data, the number of sea turtle tracks sharing were essential to this study for an overview of sea turtle movements and have contributed to highlighting the importance of marine connectivity among EEZs and the high seas (Davidson et al., 2020;Dunn et al., 2019). Future collective efforts to share and synthesize available information on all of the marine areas and linkages that populations depend upon will continue to be valuable for both scientists and resource managers (Halpin et al., 2009;Hampton et al., 2013;Hays et al., 2016;Kot et al., 2014).

ACK N OWLED G EM ENTS
Primary support was from the Global Ocean Biodiversity Initiative

CO N FLI C T O F I NTE R E S T
The authors declare no conflict of interest.

PEER R E V I E W
The peer review history for this article is available at https://publo ns.com/publo n/10.1111/ddi.13485.

DATA AVA I L A B I L I T Y S TAT E M E N T
The data that support the findings of this study are available in