When quantifying the dynamics of coupled natural-human systems it is vital to consider the ways in which human activity occurs and where it is focused (Levin et al., 2013). For example, the spatial distribution of agriculture, urban expansion, and maritime shipping have respectively affected fire regimes in African savannas (Archibald et al., 2012), whole terrestrial food-webs (Faeth et al., 2005), and large-scale patterns of phytoplankton diversity (Hallegraeff, 1998). So too in fisheries, where the spatial distribution of fishing effort is both fundamental to the organization of marine food-webs (Essington et al., 2006) and to our calculations of sustainable fisheries production (Murawski et al., 2005). Indeed, almost all fisheries management actions are designed to influence the (spatial) behavior of fishermen (Branch et al., 2006; Hilborn, 2007), yet the issue of where they decide to fish and why, remains a key obstacle to achieving sustainable fisheries (Fulton et al., 2011; Hobday et al., 2011). This uncertainty is not for want of effort. Many disciplines have examined fishing location choice, from economics (Gordon, 1954; Holland and Sutinen, 2000) to anthropology (Orbach, 1977) to ecology (Hilborn and Ledbetter, 1979). From this work it is clear that fishermen behavior is at least partially explained by a desire to maximize individual well-being (via profits), but that this consideration alone is not the whole story. In the absence of a full understanding of the factors driving fishermen's spatial behavior, their actions will continue to surprise fisheries managers, potentially undermining management systems and jeopardizing the sustainability of fisheries (Fulton et al., 2011).

The choice of where to fish is in part determined by gradients in ocean properties such as temperature, salinity, and pollution, that persist over timescales equivalent to or greater than those of a fishing trip (Lehodey et al., 1997). These often sharp gradients are created by fronts of various kinds, including buoyancy currents, tidal mixing, shelf-slope/shelf-break flow, upwelling and boundary currents, and marginal ice zone fronts. Most fronts are characterized by a surface convergence, which contributes to the aggregation of nutrients and increases in primary production (Kahru et al., 2012). As a consequence, ocean fronts collectively define a dynamic mosaic of transient micro-habitats that numerous organisms use (MacKenzie et al., 2004), and are known to be “hot spots" of marine life (Bost et al., 2009; Godø et al., 2012; Scales et al., 2014; Woodson and Litvin, 2015). Fronts are observable in satellite derived maps of sea surface temperature, primary production and ocean color. However, cloud cover, flow discontinuities and noise, and low spatial resolution of satellite data present challenges for identifying the location of fronts (Ullman and Cornillon, 2000). To overcome these challenges, recent work has focused on the spatiotemporal extent of fronts identified by Lagrangian Coherent Structures (LCSs; frontal areas of convergence in the surface ocean, see Figure 1 and Supplementary Movie; Harrison and Glatzmaier, 2012). These are areas of attraction and repulsion, typically calculated for the surface ocean, which identify frontal structures in advected tracers (d'Ovidio et al., 2004; Lehahn et al., 2007). They have also been shown to correspond to biological activity. For example, it has been shown that the spatial distribution of planktonic organisms (Harrison et al., 2013), foraging frigate birds (Olascoaga et al., 2008), elephant seals (Della Penna et al., 2015), and baleen whales (Kai et al., 2009), correspond with the location of LCSs. However, while there exists evidence that fishing effort is colocated with LCSs (Prants et al., 2012), it remains unknown whether this is a systematic relationship across different marine systems and across fisheries, and also whether any relationship between LCSs and fishing influences catch rates and the revenue generated by fishing.

To understand whether fishermen track LCSs and whether doing so has economic impacts, we compiled data on the location of over 1,000 fishing vessels every hour in the U.S. California Current Large Marine Ecosystem (CCLME) for the period 2009–2013 produced from a Vessel Monitoring System (VMS), and collated corresponding fisheries catch and price data. We focused on the spatial dynamics of vessels operating in three commercially important fisheries: the albacore tuna (Thunnus alalunga) troll, chinook and coho salmon (Oncorhynchus tshawytscha and Oncorhynchus kisutch respectively) troll and pink shrimp (Pandalus jordani) trawl fisheries, which account for roughly 35, 45, and 50 million US dollars in annual revenue respectively (values from www.oceaneconomics.org/). Complementing this, we gathered oceanographic information for the same time period, from a 4-dimensional variational (4D-Var) data-assimilation Regional Ocean Modeling System (ROMS) solution to the US west coast (Moore et al., 2011a). This ROMS model operates at a higher spatial and temporal resolution than satellite altimetry data, which is commonly used to answer similar questions, and assimilates as much data as possible to make the best possible reconstruction of past oceanic conditions for the U.S. West Coast. We used the ROMS velocity fields to identify LCSs, and with these data, we assessed whether fishing events from the tuna, salmon and shrimp fisheries are associated with attracting LCSs, and whether this impacts the revenue generated by fishing.

To assess the relationship between the spatial distribution of fishing effort and the location of LCSs, we synthesized fisheries landings data. Using these data we then defined fisheries for the CCLME. Then, using spatially explicit vessel monitoring data for the tuna, salmon and pink shrimp fisheries, we identified fishing locations. In parallel to these analyses of fisheries data, oceanographic data produced from a regional ocean model were used to identify LCSs in the CCLME for the same time-period as covered by the fisheries data. Combining the two, we assessed whether vessels, operating in different fisheries, appear to search for and catch fish preferentially on LCSs. The following subsections expand on each of these steps.

The VMS-FTLE and random-FTLE distributions for the tuna, salmon and shrimp fisheries are shown in Figure 2. Values range from past −0.1 day⁻¹ to 0.5 day⁻¹, with median values for each fishery all around 0.1 day⁻¹. From these violin plots it is possible to identify systematic qualitative differences between the VMS-FTLE and random-FTLE values. For example, for the tuna fishery the VMS-FTLE median, 25th and 75th percentiles are all greater than the corresponding random-FTLE values. The Kolmogorov-Smirnov tests identified that the tuna and salmon VMS-FTLE distributions are significantly different from random, but that there is no significant difference between the shrimp VMS-FTLE and random-FTLE distributions (Figure 2; Kolmogorov-Smirnov p-values are given below the fishery names). In addition, we found that for the tuna, salmon and shrimp fisheries respectively, 57, 47.5, and 55% of fishing events happen on FTLE values ≥ 0.1day⁻¹, relative to 52, 50.5, and 54% of random events (see Figure 3A). The G-test p-values reflect those from the Kolmogorov-Smirnov tests, that is the fraction of fishing events occurring on FTLE values ≥ 0.1day⁻¹ in the tuna and salmon fisheries are significantly different from random, whereas in contrast, there is no significant difference for shrimp fishing. The differences between the VMS- and random FTLE values, for the tuna fishery, are 2–3 times less than those observed for foraging baleen whales (Cotté et al., 2011) (the fraction of baleen whales found on FTLE values greater than 0.1day⁻1 was roughly 10–15% greater than would be expected at random). This suggests that tuna fishermen are not tracking these surface frontal features as closely as baleen whales.

To highlight the way in which these distributions differ from random, we calculated the difference in VMS- and random-FTLE survival functions, Δsurv, bootstrapping values with an 83% sample size to create confidence intervals that reflect the behavioral segmentation precision (in the case for shrimp, for which we had fishing observations, we used 75% precision: see Table 1). Survival functions identify the fraction of points with a certain FTLE value or greater (formally, they are 1 minus a cumulative density function), and the difference between the VMS- and random-survival functions identifies the magnitude and type of non-random association fishermen have with LCSs. For example, from the Δsurv curves shown in Figure 3B, it is evident that tuna fishing events always occur on FTLE values greater than would be expected from random, that is the red curve is always positive over the entire range of FTLE values. This strongly suggest that tuna fishermen preferentially search for and catch tuna on LCSs, as quantified by FTLEs.

In contrast to tuna fishing, salmon fishing always occurs on FTLE values less than would be expected from random, that is Δsurv values are negative. This can be interpreted in three ways. First, salmon fishermen may search for and catch fish in specific areas of the coast that have reduced frontal activity relative to nearby waters. Second, the spatial resolution of the ROMS model, from which the FTLE data were calculated, is not sufficient to resolve the physical features that salmon fishermen use to locate salmon. This is an important caveat to this work, and is expanded upon in the discussion. Third, salmon fishing may occur in places where LCSs were recently. We do not account for the possible time-lag between fishing and the presence of an LCS, but we can hypothesize that a time-lag would only happen if salmon track some trailing effect of LCSs. This is clearly possible, as the ecosystem effects of fluid convergence, caused by LCSs, will be integrated over space and time. In contrast to both tuna and salmon fishing, there is less difference between the shrimp VMS- and random-survival functions. In general, the shrimp Δsurv curve hugs the zero-line (the point at which there is no significant difference to random). Hence, when combined with non-significant p-values from the Kolmogorov-Smirnov and G-tests, this information confirms our expectation that shrimp fishing has no spatial correspondence with LCSs and surface frontal features.

The results of the whole-domain random test contrast strongly with those produced from the fishery-specific random test (compare Figure 3 and Figure S8). Here, all fisheries—tuna, salmon and shrimp—show negative Δsurv curves. This identifies that each of these fisheries operates in areas of the California Current with low FTLE values relative to the rest of the domain. It is an interesting oceanographic question to ask why this happens, but because our focus is on the spatial relationship between individual fishing events and specific FTLE features, and not with LME-scale biophysical patterns, we do not answer it in this paper. One point to note is that the qualitative differences between the Δsurv curves of each fishery are consistent across the fishery-specific and whole-domain tests. That is, the tuna curve is the least negative, the shrimp curve is an intermediate case and the salmon curve is the most negative. This qualitative consistency reflects the results of the fishery specific tests (i.e., that tuna fishermen track FTLE values that are relatively high when compared to shrimp and salmon fishermen).

These FTLE statistics reveal that tuna fishermen preferentially search for and catch fish in areas of high oceanic filamentation, in other words areas of the ocean with strong fronts. In contrast salmon fishing occurs in areas of low filamentation, and shrimp fishing occurs independently of the FTLE context. Do these behaviors translate into economic terms for fishermen? In order to answer this question we assessed the relationship between total revenue gained per trip (in log₁₀$) and the maximum FTLE value experienced whilst fishing, for every trip for each fishery (Figure 4). Our data consisted of 753 tuna, 670 salmon and 1907 shrimp trips (made by 94, 69, and 56 unique vessels respectively), and we found that for tuna fishing, there is a significant positive linear relationship between logarithmic trip revenue and the maximum FTLE value experienced while fishing. The slope of this linear relationship is 1.44 log$ per unit FTLE (with a standard error of 0.12 log$ per unit FTLE, calculated from bootstrapping the regression and subsampling the data by 83% to account for the segmentation precision). This univariate analysis implies that a tuna vessel that searches for and catches fish on a front with FTLE value 0.1 d⁻¹ can be expected to make roughly $3,500 on a single trip, while a vessel that catches fish on a front with FTLE value 0.4 d⁻¹ is expected to make $10,000 on a single trip, roughly three times the revenue. Flat and non-significant relationships were observed for salmon and shrimp trips indicating that trip revenue in these fisheries are not affected by the FTLE conditions experienced while fishing. These results are mirrored when trip revenue is substituted by trip landings (in units of mass: Figure S9), but they are not found when the expected FTLE experienced whilst fishing is used in place of the maximum (Figure S10; this highlights that the maximum FTLE experienced while fishing on a given trip is a better measure of catch success).

Vessel size plays a role in determining trip revenue (Figure 4: marker size is proportional to vessel length), because larger vessels have more storage space and hence a greater capacity for bringing fish to harbor. Using generalized linear mixed models (bootstrapping results using subsamples similar to the univariate analysis) we indeed found that vessel size and tuna revenues are strongly and positively correlated (slope = 1.805, p-value < 0.001). However, we still also found a positive and significant (albeit secondary) effect of FTLE values on trip revenue (slope = 0.582, p-value < 0.001). In contrast, vessel size and FTLE values were not important determinants of salmon (slopes = 0.588, 0.344; all p-values > 0.05) or shrimp (slopes = −0.001, 0.369; all p-values > 0.05) trip revenues.

Our results have identified a connection between fine-scale physical ocean features and the spatial patterns of human activity, and ultimately the economic welfare of fishermen. Specifically, our results infer that fishermen targeting albacore tuna track LCSs, as identified by large backward Finite-time Lyapunov exponents. This is reflected in the location of tuna fishing events, which are consistently found on LCSs, more so than would be expected at random, and also in the revenue gained by tuna fishermen. That there is significant positive correspondence between the spatial distribution of fine-scale (and ephemeral) physical features of the ocean and the income of fishermen is surprising, for there are many possible reasons why one fisherman might make more money than another (Fulton et al., 2011). For example, the institutional context, skill and experience, gear and weather all play a role in how much money a fishermen might make on any given trip (Hilborn, 1985; Bjarnason and Thorlindsson, 1993; Allison and Ellis, 2001). Thus, the clear signal of the physical ocean environment—the seascape—in the revenue garnered by tuna fishermen highlights the potential for strong natural-human coupling in the California Current system.

In contrast to tuna fishing events, we observed a negative relationship between salmon fishing and FTLE values, and no relationship for shrimp fishing. The absence of any relationship for shrimp fishing was expected, as shrimp fishing targets the benthos (Eales and Wilen, 1986), and as a consequence the surface physical structures described by FTLE values would not be expected to have an impact. However, the negative relationship seen in salmon fishing raises an important caveat of this work. It is impossible to tell whether this observation arises from an actual avoidance behavior of salmon fishermen, or whether this simply arises due to the inability for the ROMS model to resolve ocean processes close to shore, which is where salmon fishing typically occurs (see Figure 1). Although at the time of this work, the ROMS model data has a finer spatial resolution than satellite data, which is typically used to assess the impact of LCSs on predator foraging (e.g., Kai et al., 2009; Cotté et al., 2011; Della Penna et al., 2015), it is still too coarse to resolve coastal zone processes, which likely include even finer-scale physical ocean features than the LCS assessed here. For example, the FTLE value of 0.1d⁻¹ that we used has been adopted in previous works studying the relationship between marine top-predators and LCSs (e.g., Cotté et al., 2011). But these studies all focused on ocean regions far from shore (at least several tens of kilometers). Most likely, a different FTLE value will identify dominant frontal features in coastal zones. Hence, while 0.1d⁻¹ is appropriate for tuna fishing which occurs relatively far from shore, for studying the spatial behavior of vessels operating closer to shore, a different value should be identified using new higher resolution oceanographic information.

Indeed, the spatial scale at which fisherman behavior is assessed is a critical choice on the part of the investigator (Levin, 1992). Our choice to use the ROMS modeled data (produced at a spatial resolution of roughly 10km) allowed us to assess the impact of physical ocean features at this same spatial scale, that persist over timescales of roughly less than a month. If we were to have used higher spatial resolution data, we would have assessed the impact of even finer spatial scale and potentially more ephemeral physical ocean features on fishing. Importantly, at some fine scale there will cease to be any relationship. This is the spatial and temporal scale at which fishermen no longer are able to sense and react to the seascape. It is an interesting question to ask at what scale this is, and it likely relates to the technology that fishermen employ to search for fish. Because it is related to technology, this minimum spatiotemporal scale is likely decreasing, as vessels improve in speed and sensory ability.

In contrast to examining finer scale behaviors, one could coarse-grain the (ROMS) data, and assess the relationship between fishing and aggregate measures of the seascape. These coarse graining experiments have been used to examine the relationship between pelagic predators and relatively large areas of the ocean, characterized by different biophysical properties (Scales et al., 2017). Coarse graining experiments like these move the focus away from the behavior of individual fishermen and fishing vessels, and instead ask questions about where groups of vessels (fleets) operate. This is a vital consideration because while management typically focuses on directing the behavior of individual humans, their performance is often measured at the fleet level (Hilborn, 2007), and fleets themselves vary in their composition and in where their aggregate effort is directed (e.g., Figure S7;Boonstra and Hentati-Sundberg, 2016).

An important aspect of fisheries management that this study relates to is the impact of LCS-focused foraging on the organization and dynamics of the underlying marine food-web (Woodson and Litvin, 2015). Spatially focused foraging and consumption has been shown to impact fish stocks through a phenomenon known as “local depletion” (Bertrand et al., 2012), and in ecology these non-linearities in predator-prey interactions change the long-run mortality rate of prey, and the population growth rate of predators (Fryxell et al., 2007). This is important because the ecosystem models that we currently employ to quantify and predict the dynamics of marine ecosystems (e.g., Watson et al., 2015), possibly in the context of evaluating the potential response of management changes, use very simple linear mathematical functions to represent predator-prey encounter rates (e.g., the numerator in a Type II feeding function). These linear functions assume that prey are uniformly distributed over a given area, and predators move randomly over this area (Barbier and Watson, 2016). We know that in reality that this is not so, and our analysis confirms this. It is fair to say that all models are abstractions of reality; however, it is important to know the effect of these simplifying assumptions on the predictive skill of our ecosystem models. Our analysis suggests that more sophisticated predator-prey encounter functions that account for spatially focused foraging, and other important factors such as cooperation and advanced fishing technologies, are likely to provide new insights into fisheries dynamics (Fulton et al., 2011; Barbier and Watson, 2016).

One major policy tool that will benefit from improved understanding of the fine-scale spatial behavior of fishermen is marine spatial planning. Marine spatial planning takes many forms, for example Marine Protected Areas (MPAs) which have been widely adopted around the world (Edgar et al., 2014), and define areas of the ocean that are off-limits to fishing. Obviously, because vessels operating in different fisheries search for and catch fish in different places, MPAs will have a varying impact on people's income, depending on which fisheries they work in. Knowing in which oceanic regions and at what times of the year different fisheries operate is key to designing MPAs efficiently (Le Pape et al., 2014). Further, there are more sophisticated approaches to marine spatial planning, for example temporary closures (Brown et al., 2015), and policies directed at specific spatial behaviors, for example bycatch move-on rules (Dunn et al., 2014). Like MPAs, improved understanding of the spatial behavior of fishermen, at fine spatial and temporal resolution, will be critical to advancing these arguably more sophisticated forms of marine spatial planning.

Marine spatial planning extends to other ocean-industries, not just fishing. Indeed, much applied research is currently focused on “cross-sector” marine spatial planning, which aims to develop policies that optimize the (sustainable) use of the oceans by all sectors, for example fishing, aquaculture, shipping, tourism and oil and mineral extraction (Lester et al., 2013; Klinger et al., 2017). Like in the case of marine spatial planning geared solely for fisheries (e.g., MPAs), cross sector spatial policies will depend on a detailed quantification of spatial behavior. Importantly, this advanced understanding will not only improve our ability to design cross-sector policies for optimal/sustainable use of multiple ocean resources, but it will help us anticipate how people (like fishermen) will respond to these new policies (Klein et al., 2017). This is key to developing cross-sector policies that will provide opportunities for (economic) growth in the long-run, as the social and ecological organization of marine systems change (Klinger et al., 2018).

A final consideration for integrating detailed understanding of fishing spatial behavior into marine management is that fishermen typically work in numerous fisheries over a given year (Kasperski and Holland, 2013; Fuller et al., 2017). As a consequence, fishermen will move across, exploit and make money from different parts of a seascape at different parts of the year. Hence, they will have many relationships with different parts of the seascape. For example, a fisherman on the U.S. west coast that works in the albacore tuna fishery may also work in the dungeness crab fishery. As a consequence, high FTLE ocean features will be important to this fisherman during the tuna season, but not when they operate in the crab fishery. This is similar to agriculturalists that grow different crops in different parts of a landscape at different parts of the year. Indeed, just like in agricultural management (Iverson et al., 2014), spatial marine policies will be improved if they acknowledge that fishermen's diverse harvest portfolios connect different parts of seascapes (Hobday et al., 2011).

In summary, coastal management and marine spatial planning that recognizes and accounts for feedbacks between fishermen behavior and the dynamic and structured nature of seascapes is likely to meet with greater compliance and provide for stronger resilience in the face of cumulative anthropogenic impacts. Indeed, flexibility and resilience—the ability to cope with change (Folke, 2006)—are key properties of coastal human communities that managers are looking to bolster, especially in the context of impending shifts in species ranges expected with climate change (Pinsky et al., 2013). Furthermore, incorporating fine spatial-scale information into management policies is essential if we are to continue to improve the efficiency of fisheries (Hobday et al., 2011). Increased fisheries efficiency would lead to fishermen spending less time at sea, and as a consequence reduced risk of harm. Of course, in the interest of long-term sustainability this increase in efficiency should be matched by changes in governance institutions that prevent over harvest. Ultimately, given the coupled nature of social-ecological systems, integrating high resolution data on physical, chemical, ecological and social factors will improve both how and to what extent we use living natural resources (Wilen, 2004), leading to improved chances of prosperity now and in the future.

JW, EF, FC, and JS all contributed data, designed the scientific approach, analyzed the data and wrote the manuscript.