FISHERIES OCEANOGRAPHY

Fish. Oceanogr. 12:4/5, 443–457, 2003

Investigating remote synchronous patterns in fisheries

P. FRE´ON,1,2,* C. MULLON1,3 AND B. VOISIN4 1

Institut de Recherche pour le De´veloppement, 213, rue La Fayette, 75480, Paris, France 2 Marine and Coastal Management, Private Bag X2, Rogge Bay, Cape Town, 8012, South Africa 3 University of Cape Town, Private Bag, 7700 Rondebosch, South Africa 4 ForeTrade Support, 9, Silwood Road, 7701 Rondebosch, South Africa

ABSTRACT The hypothesis that the population dynamics of distant fish stocks may be synchronized by climate variations is largely based on a few major pelagic fisheries or stocks (called here ‘target series’). We broadened the study of synchrony to all available fisheries in the world by making use of all available data in the FAO catch database during the last 50 years. We investigated synchronous patterns in remote fisheries using two different approaches. The first approach consisted of simulating catch time-series as random walk processes and comparing the synchronies found between these time-series to the synchronies found in FAO data. In the second approach, we applied classification algorithms to analyse synchronies in the FAO catch database. We used K-means clustering and neural networks to answer the following four questions: (1) Do the target series appear in the same cluster? (2) Do the clusters regroup series from the same region (local synchrony) or from distant ones (remote synchrony)? (3) Are there any numerically dominant patterns (clusters) in the catch series? (4) Do the species displaying the same pattern share some common bio-ecological features or ranges of abundance? Our results indicate some degree of local rather than remote synchrony. Nonetheless, it does not prove that such remote synchronies do not exist, but simply that further process studies are required before accepting them as a new paradigm. Key words: catch data, K-means, pelagic fish, neural network, regime shift

*Correspondence. e-mail: [email protected] Received 25 October 2002 Revised version accepted 13 June 2003
2003 Blackwell Publishing Ltd.

INTRODUCTION Kawasaki (1983) was the first to notice that during the 20th century several of the world’s largest populations of sardine Sardinops sagax experienced large, long-term changes. He showed that a pattern of ocean-wide synchrony emerges when one looks very broadly at the sardine landings from three regions of the Pacific Ocean (Japan, California and Peru-Chile) that have supported large sardine fisheries. The collapse of the Japanese and Humboldt sardine populations in the late 1980s and early 1990s reinforced Kawasaki’s point of view (Lluch-Belda et al., 1992a; Bakun, 1998; Schwartzlose et al., 1999). Moreover, two of the other largest coastal pelagic fish populations in the world, the Peruvian anchoveta Engraulis ringens and the South African sardine, appeared to rise and fall directly out of phase with the Pacific sardine (Lluch-Belda et al., 1989, 1992b; Crawford et al., 1991). Synchronies are also observed at the scale of oceanic basins. Many of the very large groundfish populations of the sub-Arctic North Pacific have also been varying substantially in phase with sardine populations (Bakun, 1998). Catches of several salmon species around the Pacific Rim have also indicated some coherence in the low-frequency, large-amplitude variation of the corresponding populations (Beamish and Bouillon, 1993), which, in turn, are out of phase with the anchovy populations (Beamish et al., 1999). Understanding the processes involved in these possible synchronies is commonly considered a key scientific task in terms of climate change, management, prediction or human impact. As the fish populations are certainly far too widely separated to interact through direct exchanges, a largescale climatic signal was first proposed as being the driving force of the observed synchronies. The climatic regime shift of the mid-1970s is well documented for the Northern Pacific, and is a strong candidate to account for several of the changes in salmon populations observed during the late 1970s in that region (Polovina et al., 1994; Hare and Mantua, 2000). For the small pelagic populations, the situation is unclear. Sea Surface Temperature (SST) was considered initially (LluchBelda et al., 1989, 1992b), but the SST trends have not been consistent in the various regions inhabited by populations of small pelagic fish (Bakun, 1998). Klyashtorin (2001) suggested that the Atmospheric 443

444

P. Fre´on et al.

Circulation Index (ACI) is a reliable climatic index that may be related to regular long-term changes in the major commercial fish stocks. The long-term fluctuations of the main commercial catches of Pacific salmon, Japanese, Californian, Peruvian, South African and European sardine, Alaskan pollock, Chilean jack mackerel, Atlantic and Pacific herring as well as Atlantic cod display a pronounced spectral maximum of 54–58 years and correlate with the ACI zonal or meridional component (r ¼ 0.70–0.90) (Klyashtorin, 2001). Spencer and Collie (1997) only observed shorter periodicity (around 40 years) because they used timeseries shorter than 80 years. Regional synchrony forced by environmental changes is also advocated to explain multi-centennial regimes of anomalous salmon, sardine and anchovy abundance in the northern Pacific (Beamish et al., 1999; Finney et al., 2000, 2002). Most of the previous studies focused only on a few major fisheries or stocks, and on the alternation of sardine and anchovy in the major ecosystems (regime shifts, turning points), but these synchronies are not exclusive to pelagic or semi-pelagic fisheries, as shown by recent studies (Hollowed et al., 2001; Klyashtorin, 2001). The objective of this study is to use statistical tools to investigate synchronous patterns in all available marine fisheries in the world (including crustaceans, molluscs, sharks and seaweeds). Instead of focussing the study on a limited number of major pelagic stocks for which long time-series (>70 years) are available, we broaden it by making use of all available data in the FAO catch database for the last 50 years. This dataset is first compared to simulated data (random walk process). We then use clustering methods to identify synchronies and to find out if any properties (ecological, geographical, human induced, etc.) are shared by fisheries displaying synchrony. This approach is based on a number of assumptions, implicit in previous studies: 1. commercial catches are adequate proxies for detecting low-frequency variability in stock abundance (these are the only proxies available at a global scale); 2. national reported catches by FAO areas (FishStat Plus) are the best and more consistent data available. Our aim was to answer the following four questions: (1) Do the ‘‘target series’’ of clupeoid species that supposedly vary in synchrony appear in the same cluster? (2) Do the clusters regroup series from the same region (local synchrony) or from distant ones (remote synchrony)? (3) Are there any numerically dominant patterns (clusters) in the catch series? (4) Do the series displaying the same pattern share any common properties? As the target series corresponds to large pelagic fish stocks of low trophic levels, we

focussed on large stocks (regardless of the species) and pelagic species (regardless of the stock size) in more detail, and paid attention to the trophic level distribution in the clusters. MATERIAL AND METHODS Data The main focus of our approach was to make data repeatable by other teams by using available data and as little filtering and transformation as possible in order to keep it simple. We used the FAO database (FishStat Plus), which covers 15 225 national fisheries time-series data from 1950 to 1998. We extracted 965 time-series using the following criteria: • no more than three consecutive years of missing data; • no more than four consecutive years of repeated values, except if 10 millions tons).

447

11 175 coefficients. The frequency distribution of these coefficients was compared to the frequency distribution resulting from a similar processing of an FAO data series. Furthermore, we contrasted the distribution of coefficients to the correlation coefficients between data series corresponding to clupeoid stocks identified visually as varying in synchrony by Schwartzlose et al. (1999); we will call these ‘target series’ in the rest of the paper (Tables 3 and 4; Fig. 1). Clustering algorithms Clustering techniques were applied to the complete dataset, the dataset plus its mirror image and the limited dataset of 22 data series to construct homogeneous clusters according to the shape of the catch series only. We analysed the distribution of the cluster size (n), looking for those clusters expected to be numerically dominant, and verifying if those target series described in the literature as varying in synchrony were classified in the same cluster. As all clustering methods are sensitive to the choice of parameters and settings (number of clusters, pre-clustering methods, parameters), we made several trials with 6, 12, 20, 30 and 40 clusters, using two clustering techniques: K-means and neural networks. We performed a variance analysis of the clustering to get a compromise between the number of clusters and their

Country

Common name

Scientific name

Total catches, 1950–1998 (tons)

Peru Japan Chile Peru Japan Japan Chile Norway USA Norway China Japan USA Namibia Canada Japan Norway Iceland Japan UK Japan Morocco

Anchoveta Japanese sardine Chilean jack mackerel Peruvian sardine Alaska pollock Chub mackerel Anchoveta Capelin Gulf menhaden Atlantic herring Largehead hairtail Japanese flying squid Atlantic menhaden Sardine or pilchard Atlantic cod Japanese anchovy Atlantic cod Atlantic cod Pacific saury Atlantic cod Japanese jack mackerel European sardine

Engraulis ringens Sardinops sagax Trachurus murphyi Sardinops sagax Theragra chalcogramma Scomber japonicus Engraulis ringens Mallotus villosus Brevoortia patronus Clupea harengus Trichiurus lepturus Todarodes pacificus Brevoortia tyrannus Sardinops sagax Gadus morhua Engraulis japonicus Gadus morhua Gadus morhua Cololabis saira Gadus morhua Trachurus japonicus Sardina sardineus

1.86 6.07 4.61 4.16 3.84 3.43 2.83 2.69 2.64 2.56 2.34 1.64 1.55 1.47 1.43 1.42 1.42 1.39 1.31 1.24 1.20 1.03


2003 Blackwell Publishing Ltd, Fish. Oceanogr., 12:4/5, 443–457.

· · · · · · · · · · · · · · · · · · · · · ·

108 107 107 107 107 107 107 107 107 107 107 107 107 107 107 107 107 107 107 107 107 107

448

P. Fre´on et al.

Table 3. ‘Target series’: a selection of 13 national catches of small pelagic fish whose corresponding stock is supposed to display remote synchrony with at least one of the stocks in the table (according to Schwartzlose et al., 1999). Country

FAO region

Common name

Scientific name

Tot. catches, 1950–1998 Average landing

Peru Japan Peru Chile Namibia Japan Mexico South Africa Korea (Rep.) Brazil Mexico USA USA

Pacific, Southeast Pacific, Northwest Pacific, Southeast Pacific, Southeast Atlantic, Southeast Pacific, Northwest Pacific, Southeast Atlantic, Southeast Pacific, Northwest Atlantic, Southwest Pacific, Southeast Pacific, Eastern Central Pacific, Eastern Central

Anchoveta Japanese sardine Peruvian sardine Anchoveta Southern Africa sardine Japanese anchovy Californian sardine Sardine or pilchard Japanese anchovy Brazilian sardinella Californian anchovy Californian anchovy Californian sardine

Engraulis ringens Sardinops sagax Sardinops sagax Engraulis ringens Sardinops sagax Engraulis japonicus Sardinops caeruleus Sardinops sagax Engraulis japonicus Sardinella brasiliensis Engraulis mordax Engraulis mordax Sardinops sagax

1.86 · 108 60 683 069 41 577 169 28 338 059 14 686 000 14 185 431 8 151 165 5 653 090 5 560 184 4 262 774 2 818 427 1 395 249 1 029 585

3 799 394 1 238 430 848 514 578 328 299 714 289 499 166 350 115 369 113 473 86 995 57 519 28 474 21 012

Table 4. Correlation matrix of time-series of FAO catches corresponding to stocks identified as varying in synchrony by Schwartzlose et al. (1999). Species

A.P.

A.C.

A.K.

A.J.

A.M.

A.U.

S.P.

S.J.

S.M.

S.U.

S.N. S.S.

S.B.

Anchoveta, Peru (A.P.) Anchoveta, Chili (A.C.) Anchovy, Korea (A.K.) Anchovy, Japanese (A.J.) Anchovy, Mexico (A.M.) Anchovy, USA (A.U.) Sardine, Peru (S.P.) Sardine, Japanese (S.J.) Sardine, Mexico (S.M.) Sardine, USA (S.U.) Sardine, Namibia (S.N.) Sardine, South Africa (S.S.) Sardinella, Brasil (S.B.)

1.00 0.66 0.06 0.25 –0.34 –0.01 )0.20 –0.33 )0.14 )0.28 0.58 0.18 0.11

1.00 0.51 –0.13 –0.22 –0.25 0.29 0.10 0.41 )0.19 –0.01 –0.13 0.03

1.00 –0.49 0.39 0.06 0.62 0.49 0.78 )0.31 –0.42 –0.49 0.46

1.00 –0.70 1.00 –0.09 0.22 1.00 )0.55 0.42 )0.27 1.00 )0.69 0.60 )0.16 0.88 1.00 )0.64 0.53 )0.20 0.84 0.85 1.00 0.19 )0.24 )0.24 )0.25 )0.26 )0.25 1.00 0.50 –0.36 0.26 –0.58 –0.55 –0.61 –0.19 1.00 0.40 –0.38 –0.25 –0.49 –0.51 –0.51 0.05 0.25 1.00 –0.31 0.42 0.61 0.12 0.19 0.25 )0.40 0.04 )0.44 1.00

Bold numbers are expected in-phase synchronies whereas italic numbers are expected synchronies in phase opposition. Note that Schwartzlose et al. often user longer time-series than FAO ones and that they do not always regroup the data by country.

internal variability. The criterion used to judge the strength of a given synchrony was its consistency across different settings or methods. The conventional K-means clustering algorithm iteratively assigns K centres to represent the clustering of N points (Bishop, 1995; StatSoft, Inc., 2000). The resulting clusters are of greatest possible distinction. The algorithm starts with K random clusters, and then moves objects between these clusters to minimize variability within clusters and maximize variability between clusters (analogous to ‘ANOVA in reverse’). The initial cluster centres can be objectively assigned in two different ways: first, by maximizing the between-cluster distances, and second, by taking

observations at constant intervals. We used both of these methods. The sum of squared Euclidian distances between a given cluster and all others provides an index of oddity of the cluster. The SOM (self-organizing map) neural network classifies the time-series in an unsupervised way, in an attempt to find the appropriate number of clusters to describe the major patterns (Kohonen, 1995). The SOM algorithm is a neural network that learns data features in an unsupervised way. Clusters are organized in a two-dimensional matrix (here 5 · 4 or 6 · 4). A two-stage vector computation (quantization) is used to allow smoother feature detection. The first step facilitates pre-ordering (definition of neighbourhood


2003 Blackwell Publishing Ltd, Fish. Oceanogr., 12:4/5, 443–457.

Investigating synchronous patterns in fisheries

449

Figure 1. Some examples of documented synchronies as reflected in the FAO database.

radius, which in our case was equal to the larger matrix dimension), while the second focuses on better clustering convergence (we used a value of two for the radius). Other typical settings that we used were hexagonal topology, a bubble neighbourhood, 2000 iterations and a learning rate of 0.02 during the preordering phase of quantization, versus 20 000 iterations and a learning rate of 0.05 during the final phase of quantization. We did not subset the data for training. Analysis of clusters with ancillary variables In order to quantify how the cluster composition matched the properties of the independent ancillary variables, we used a decision tree algorithm C & RT (Classification & Regression Tree) to reconstruct the clusters independently from the initial clusters. This enables the segmentation success to be used as a criterion of representativeness of the clustering (Breiman et al., 1984; StatSoft, Inc., 2000). RESULTS Simulated versus FAO data The absolute value of the correlation coefficient |r| in the 21 common identified synchronies shown in Table 4 varies from 0.12 to 0.88, with an average value of 0.59 for the 10 highest values. The frequency distribution of |r| in the correlation matrix constructed from the 965 FAO data series displays a bell shape, similar to the frequency distribution obtained from the correlation matrix constructed from the

simulated dataset (not shown). Testing the confidence limits of these coefficients according to conventional statistical tables (e.g. Zar, 1999) is debatable due to the non-independence of the observations within a series (trended and auto-correlated time-series), and most authors prefer a visual assessment. In the FAO correlation half-matrix, 82 375 (17.7%) of the 465 130 correlation coefficients exceeded the absolute value of 0.59 (which is well above the threshold of 0.46 corresponding to P < 0.001) and 13 582 (2.9%) exceeded 0.80. The corresponding values were even higher in the correlation matrix of simulated series. Indeed, we observe in both types of series high visual resemblance between many pairs of catch series, without any identifiable link in the case of FAO data. Clustering with K-means Two options used to define the initial cluster centres (maximum between-cluster distances or observations at constant intervals) provided different cluster compositions, suggesting that the classification is unstable. For instance, using the pre-clustering method that maximizes between-cluster distances, the K-means classification in 20 clusters of the 965 exponential smoothed z-score data (without adding mirror images of the time-series) provides differing numbers of observations per cluster: one with 14 observations and two dominant clusters with more than 120 observations (the third cluster having only 58 observations). The last two clusters (7 and 14) were above the 90 percentile of


2003 Blackwell Publishing Ltd, Fish. Oceanogr., 12:4/5, 443–457.

P. Fre´on et al.

450

oddity using the sum of squared Euclidian distances. We found only three clusters that regrouped two or three of the 13 target series, which is less than expected if the observations were randomly distributed in the clusters. The seven corresponding series are as follows: • USA anchovy and Brazilian sardinella with low cluster distances of 0.37 and 0.39 respectively (cluster 2, n ¼ 57; Fig. 2). • Korean anchovy, Mexican and Peruvian sardine (three increasing time-series), with low cluster distances of 0.36, 0.22 and 0.29, respectively. Interestingly, this cluster is the one with the highest oddity coefficient and the highest number of timeseries (cluster 7, n ¼ 129; Fig. 2). • Japanese anchovy and South African sardine, with medium cluster distances of 0.50 and 0.43, respectively (cluster 9, n ¼ 49). When using the constant intervals pre-clustering method on the same dataset, the numbers of observations per cluster were more homogeneous than before (frequency distribution between 13 and 101; figure not shown). Two clusters were above the 90 percentile but there were no target series in these two clusters. There were only two clusters that regrouped two of the 13 target series (which is less than expected if the observations were randomly distributed among the clusters). The four corresponding series are as follows: • as before, the USA anchovy and the Brazilian sardinella with cluster distances of 0.41 and 0.37, respectively;

Figure 2. Mean values of Z-core annual catches (year 1 ¼ 1950) of some K-means classifications in 20 clusters of the whole dataset, using the maximum distance pre-clustering option. 4 Cluster 2 Cluster 6 Cluster 7 Cluster 14

2 1

–1 –2

Year index

49

45

41

37

33

29

25

21

17

13

9

5

0 1

Z-score catches

3

• the Mexican sardine is now in the same cluster as the Japanese sardine (but no longer with the Peruvian sardine and the Korean anchovy), with low and medium cluster distances of 0.21 and 0.45, respectively. Table 5 provides a synoptic view of the differences between the two pre-clustering methods regarding the target series. After different trials of clustering the 965 timeseries into 6, 12, 20, 30 and 40 clusters, we found the most suitable value to be 20. As expected, decreasing the number of clusters increases the number of detected synchronies, but will also increase the distance of the time-series to their cluster, a distance that is already high in most of the results using 20 clusters (results not shown). On the other hand, increasing the number of clusters will decrease the probability of finding synchrony among the 13 target series: with 40 clusters and using the maximum between-cluster distance method, we found only the sardine time-series from Japan, Peru and Mexico to fall within the same cluster (n ¼ 57), with distances of 0.36, 0.21 and 0.15, respectively. Using the constant intervals method, only the sardine time-series from Japan and Mexico fell within the same cluster (n ¼ 40), with distances of 0.37, and 0.13, respectively. Repeating these analyses on the dataset combined with its mirror image allows the identification of some of the phase opposition synchronies between target species as described in the literature (Table 6). For instance, the following out-of-phase synchronies were identified by K-means when using the pre-clustering option that maximized between-cluster distances: the Peruvian anchoveta and the USA sardine on the one hand, and the anchovy time-series from Japan and Mexico on the other. When using the constant intervals method, we found the following time-series to be out of phase in the same cluster: the anchovy time-series from Japan was anti-correlated with the one from Mexico (as before) but was also anti-correlated with the Japanese sardine, the USA sardine was anti-correlated with the Brazilian sardinella, and finally the Japanese sardine was anti-correlated with the Japanese anchovy. Inconsistencies in the results were related to the two different pre-clustering methods (Table 6), and also between the initial analysis using only the original time-series and the analysis after adding the mirror image (compare Tables 5 and 6). The most consistent in-phase synchronies were between the Brazilian sardinella and the USA anchovy. When trying to identify some common features in the life histories or abundance of the time-series


2003 Blackwell Publishing Ltd, Fish. Oceanogr., 12:4/5, 443–457.

Investigating synchronous patterns in fisheries

451

Table 5. Synoptic view of the in phases synchronies described in the literature (grey blocks) and of the catch series found in the same cluster according to two clustering methods on the 965 FAO data series: K-means with 20 clusters and using two pre-clustering methods (maximizing the between-cluster distances, K1; taking observations at constant intervals, K2) and SOM neural network with a matrix of 5 · 4 clusters (S1) or 6 · 4 clusters (S2).

belonging to the same cluster, we found little structuring by the ancillary variables. For instance, in our first example of K-means classification in 20 clusters using the pre-clustering method that maximized between-cluster distances, we used ancillary variables in the decision tree algorithm to reconstruct the 20 clusters, but the segmentation success was only 36%. FAO areas had the highest contribution (Fig. 3), the species composition an intermediate one (Fig. 4), whereas the trophic level had one of the smallest contributions, in contrast to what we expected. Clustering the 109 series of pelagic species into 12 clusters did not provide clear numerically dominant patterns. For instance, using the maximum betweencluster distances, the 13 target series were located in nine different clusters, and the USA sardine series was alone in its cluster. Finally, several attempts to cluster the 22 major catch series (Table 2) into increasing numbers of clusters revealed no clear numerically dominant cluster(s) until one reached a total of 12 clusters, regardless of the pre-clustering option. Then, when using the maximum between-cluster distances, a dominant cluster of four time-series appears, which regroups three Japanese species (anchovy, squid and

jack mackerel) and the Peruvian anchoveta, i.e. two out of our 13 target series. Clustering with SOMs By clustering the whole dataset in a matrix of 5 · 4, the number of time-series per cluster varied from 15 to 79 and the map was unfolded (not shown). The SOMs map classified the time-series according to the following two characteristics (Fig. 5): on the horizontal axis, the mean cluster pattern varied from nonmonotonic (unimodal) to monotonic; on the vertical axis, it varied from a decreasing to an increasing trend. Interestingly, the two numerically dominant patterns (clusters 02 and 43) are increasing or decreasing time-series, with homogeneous compositions as indicated by the small confidence limit intervals (especially for cluster 43; Fig. 5). The following target series were found in the same cluster: • South African sardine and Japanese anchovy (cluster 01); • Peruvian, Mexican and Japanese sardine (cluster 41). For the SOMs algorithm, it was more difficult to identify clusters with bimodal patterns such as the


2003 Blackwell Publishing Ltd, Fish. Oceanogr., 12:4/5, 443–457.

452

P. Fre´on et al.

Table 6. Synoptic view of the synchronies described in the literature (in phase ¼ dark grey blocks; out of phase ¼ light grey blocks) and of the catch series found in the same cluster according to two clustering methods on the 965 FAO data series combined with mirror images: K-means with 20 clusters and using two pre-clustering methods (maximizing the between-cluster distances, K1; taking observations at constant intervals, K2).

anchovy time-series from Peru and Chile. This suggests that bimodal time-series are atypical and/or that the location of the mode varies from series to series. To obtain these bimodal clusters, one had either to increase the number of clusters up to 24 (matrix of 6 · 4) or, if using fewer clusters (such as 20 in a matrix of 5 · 4), to set the neighbourhood to zero during the final phase of computation. This last option is not a conventional one and always yielded folded maps of cluster centres, which were discarded. When using a matrix of 6 · 4, the number of time-series per cluster varied from 15 to 65 (results not shown) and the following target series were found in the same cluster: • South African sardine and Japanese anchovy (cluster 02); • Namibian sardine and Peruvian anchoveta (cluster 03); • Peruvian and Mexican sardine (cluster 52). DISCUSSION It can be argued that commercial catches poorly represent the variation in abundance of stocks because of external anthropogenic factors (technological improvements and economic health of the countries,

fishery management, wars, etc.) or poor quality of FAO data for some countries. However, we made the same assumption as authors claiming that commercial catches are adequate proxies for stock abundance when one studies low-frequency variability, because most fish populations display interdecadal variation of one or more orders of magnitude. We contend that even the increasing number of fishery regulations during recent years does not necessarily weaken the relationship between catches and abundance, as most of these regulations are aimed at getting catches proportional to abundance at the interdecadal scale. In many instances, especially in pelagic fisheries, which are the best candidates for synchrony, regulation is provided by variable quotas that are nearly always filled because of fleet overcapacity. Another serious objection could be that our method is not able to reflect these synchronies because we deal with national catches, which do not always match boundaries of stock units, populations or ecosystems. The answer to this criticism is threefold. First, authors who claim synchrony are not always consistent in their grouping of national catches into stocks, populations or ecosystems: the pelagic catches of Namibia and South Africa are sometimes considered


2003 Blackwell Publishing Ltd, Fish. Oceanogr., 12:4/5, 443–457.

Investigating synchronous patterns in fisheries

453

Figure 3. Distribution amongst FAO areas of the 20 clusters resulting from the K-means classification of the whole dataset, using the maximum distance pre-clustering option (technical realisation by Laurent Drapeau).

Herrings, sardines, anchovies Cods, hakes, haddocks Jacks, mullets, sauries Mackerels, snoeks, cutlass.

Redfishes, basses, congers Tunas, bonitos, billfishes Shrimps, prawns Others

100

Proportion (%)

80

Figure 4. Grouping of species distribution within 20 clusters resulting from the K-means classification of the whole dataset, using the maximum distance pre-clustering option.

60 40 20 0

1

2

separately, and at other times regrouped. Second, as far as synchrony is concerned, the object of study is not clearly defined scientifically: is it an exploited stock, a group of stocks belonging to the same ecosystem or a population defined on a genetic basis? In that respect, the regrouping of all the Californian sardine stocks from Canada to Mexico is questionable, especially because the national catches of this species show opposing trends (Mexico and USA). Thirdly, if remote synchronies occur in some populations, in most

3

4

5

6

7

8

9 10 11 12 13 14 15 16 17 18 19 20 Cluster

of the cases they should be reflected equally in the different components of these populations exploited by different countries. From a statistical point of view, one must bear in mind the main limitations of our approach. The K-means is a robust procedure that requires little parameterization except for the definition of initial cluster centres, but some authors do not find it to be as powerful as neural networks. The SOM procedure supposedly performs better, but is sensitive to


2003 Blackwell Publishing Ltd, Fish. Oceanogr., 12:4/5, 443–457.

454

P. Fre´on et al.

Figure 5. Map of the mean clusters in a 5 · 4 matrix of time series (year 1 ¼ 1950) resulting from the SOM algorithm used on the whole dataset.

parameterization (learning rate, neighbourhood, weighting factors). More important for the interpretation of our results is the low number of patterns that one can expect in time-series of 49 years of largely auto-correlated data (limited number of true degrees of freedom). Most of the time-series display decreasing or increasing trends, which can be interpreted either by the effect of exploitation or by non-anthropogenic long-term cycles in which the half-wavelength is close to or larger than the length of the FAO time-series. Low-frequency variations of pelagic fish stock abundance have been documented in many ecosystems (Glantz, 1992; Spencer and Collie, 1997; Francis et al., 1998; Schwartzlose et al., 1999; Bakun and Broad, 2002), and are often interpreted as the result of environmental forcing, although other processes, such as interactions between species or – during the anthropogenic period – exploitation, may also be advocated (e.g. Garcia and Grainger, 1997; Caddy and Garibaldi, 2000; Cury et al., 2000; Kornilovs et al., 2001; Link and Garrison 2002). SOM maps show a continuum of patterns and have difficulty in identifying time-series with more than two or three turning points. These turning points do not match those of the ACI proposed by Klyashtorin (1998), or those of the Pacific Decadal Oscillation index. This suggests that

only a minority of the national catch series are correlated with these indices. Our results indicate that, from nearly half a million correlations that one can compute from 965 timeseries of catches, either observed or simulated, several hundred display tighter correlations than those observed between the target series. Clearly most of the correlations occurred just by chance. Whether or not some of them have a meaning in terms of remote synchronies triggered by a common external factor is impossible to prove but remains largely debatable. If such was the case, as advocated for the major pelagic species, these stocks should be found consistently in the same cluster(s), which was not the case. Instead, we observed instability in the composition of the clusters, with different compositions according to options in the clustering algorithms. Finally, our data analysis does not support the idea that species within the same cluster should share some common features in their life histories or abundance. In that respect, the contribution of trophic levels was very low, whereas we would have expected the anticipated role of environmental forcing factors on lower trophic levels to result in a higher contribution of this ancillary variable. Only the FAO region had a reasonable contribution, mainly because one cluster (no. 14 in the K-means analysis) was dominant in the two


2003 Blackwell Publishing Ltd, Fish. Oceanogr., 12:4/5, 443–457.

Investigating synchronous patterns in fisheries

regions of south-east Asia (Fig. 3). This result suggests a certain level of local synchrony in this part of the world, although the very tight synchronies observed among Indonesian time-series are questionable. Because cluster 14 regroups decreasing time-series (Fig. 2), a generalized overexploitation of the region (Pauly et al., 1998) might explain this result. Regional synchronies might be easier to explain than remote synchronies because consistent low-frequency environmental change is more often observed at the basin scale than at the planetary one. How these basin-scale climatic changes translate into one or more local effects on the shelf is still largely unknown. However, Rothschild (1995) suggested that fishingindependent variations of abundance occurring at a large multidecadal scale are related either to multidecadal fluctuations in primary production or to changes in the trophodynamic pathways by which primary production is transformed into fish biomass. Other promising hypotheses related to hydrodynamics have been proposed recently to explain coherent regime changes observed during the last 40 years in the Pacific (Francis et al., 1998; McFarlane et al., 2000; McFarlane and Beamish, 2001; Bakun and Broad, 2002). From the 23 synchronies indicated in Table 4, we found only a certain degree of synchrony within the main Pacific time-series of sardine during certain clustering trials. Two to three sardine time-series of the following countries were sometimes found in the same cluster: Japan, Korea, Mexico and Peru, but not the USA. On one occasion these sardine time-series occurred in the same cluster as the Korean anchovy, although anchovy and sardine are expected to display out-of-phase synchrony (Kawasaki, 1983; Luch-Belda et al., 1992a; Bakun, 1998; Schwartzlose et al., 1999). The only consistent result was the association between the Brazilian sardinella and the USA anchovy, a synchrony not described in the literature. In brief, our results indicate few occurrences of synchrony between national pelagic catches, and the number of synchronies observed is less than would be expected if the time-series were randomly distributed in the different clusters. Recently, there have been some major unexplained departures from remote synchrony. One of these is the apparent time lag of about one decade in the California Current relative to the Japanese and Peruvian systems (see appendix in Schwartzlose et al., 1999). The sardine population off California has been expanding since the early 1990s, while at the same time, in the Southern Pacific, anchovy became the dominant population and sardine declined to a low level. As a consequence, the sardine populations of the

455

Humboldt and California systems are now in an opposite phase of abundance. The apparent link between the Pacific and South Atlantic sardine populations (out-of-phase situation between the Pacific and the Benguela) also seems to have vanished. Moreover, opposite patterns are occurring within the Benguela system. In the Southern Benguela, the anchovy population reached a record high level in 2000 while the sardine population has shown a positive trend for the last 15 years (Beckley and van der Lingen, 1999; MCM unpublished data). According to Luch-Belda et al. (1992a), sardine population collapse is simultaneous with the population shrinking to smaller, equatorward refuge areas, a concept that could help in the interpretation of some of our results but does not apply here. The equatorward refuge area for sardine in the Benguela should be Namibia, but the Namibian sardine stock collapsed in the middle of the 1970s and has still not recovered. Furthermore, farther south the South African stock is extremely healthy, with a persisting positive trend over the last 13 years and catches close to 200 000 tons over the last 3 years, sustained by an average biomass of 1.8 million tons (MCM unpublished data). This apparent contradiction might be solved if one considers that the Benguela differs from all other Eastern boundary upwelling ecosystems by the presence of warm water both in equatorward and polarward regions. In conclusion, although large-scale synchronies between widespread marine fish populations have occurred over more than a decade, from the mid-1970s to the late 1980s, it is becoming increasingly obvious that, for the small pelagic populations, the situation that prevailed from the mid-1970s to the late 1980s is no longer true. Our results for national catches suggest that most of the synchronies observed occurred just by chance due to the limited degrees of freedom of largely auto-correlated time-series over a few decades. Many patterns have appeared where the clupeoid species are distributed in all the corresponding clusters, but the clusters where the target series are found do not share common features in their life histories. Although the aim of this paper was not to study periodicity in time series of catches, our results support findings of previous authors (Baumgartner et al., 1992; Garcia and Grainger, 1998; Beamish et al., 1999; Finney et al., 2000) showing that some fisheries – and in many cases the corresponding fish stocks – display long-term cycles of variation. The factors responsible for these cycles are not clearly understood (ecosystem dynamics, physical forcing, anthropogenic) but they do not seem to vary in synchrony. Nonetheless, if this work demonstrates that remote synchronies may not be proven


2003 Blackwell Publishing Ltd, Fish. Oceanogr., 12:4/5, 443–457.

456

P. Fre´on et al.

with the available data and the particular methodology used, it does not prove that such synchronies do not exist. Therefore, this study suggests that further process studies are required before accepting causally linked remote synchronies as a new paradigm. Regional synchronies forced by climatic changes at a basin scale may be easier both to prove and to understand. ACKNOWLEDGEMENTS The authors are grateful to Salvador Lluch-Cota for initial discussion and analysis on this topic, to Anthony Richardson for his help in the neural network analysis, and to Laurent Drapeau for his contribution in the use of the decision tree and segmentation analysis. We also acknowledge Claude Roy and Philippe Cury for upgrading our background on reported synchronies and Jenny Huggett for improving the language. REFERENCES Bakun, A. (1998) Ocean triads and radical interdecadal stock variability: bane and boon for fisheries management. In: Reinventing Fisheries Management. T. Pitcher, P.J.B. Hart and D. Pauly (eds) London: Chapman and Hall, pp. 331– 358. Bakun, A. and Broad, K. (eds) (2002) Climate and Fisheries: Interacting Paradigms, Scales, and Policy Approaches. Palisades, New York: International Research Institute for Climate Prediction, Columbia University. Baumgartner, T.R., Soutar, A. and Ferreira-Bartrina, V. (1992) Reconstruction of the history of Pacific sardine and northern anchovy populations over the past two millennia from sediments of the Santa Barbara Basin, California. Calif. Coop. Oceanic Fish. Invest. Rep. 33:24–40. Beamish, R.J. and Bouillon, D.R. (1993) Pacific salmon production trends in relation to climate. Can. J. Fish. Aquat. Sci. 50:1002–1016. Beamish, R.J., Noakes, D.J., MacFarlane, G.A., Klyashtorin, L., Ivanov, V.V. and Kurashov, V. (1999) The regime concept and natural trends in the production of Pacific salmon. Can. J. Fish. Aquat. Sci. 56:516–526. Beckley, L.E. and van der Lingen, C.D. (1999) Biology, fishery and management of sardines (Sardinops sagax) in southern African waters. Mar. Freshwater Res. 50:955–978. Bishop, C.M. (1995) Neural Networks for Pattern Recognition. Oxford: Oxford University Press. Breiman, L., Friedman, J.H., Olshen, R.A. and Stone, C.J. (1984) Classification and Regression Trees. Belmont: Wadsworth Inc. Caddy, J.F. and Garibaldi, L. (2000) Apparent changes in the trophic composition of world marine harvests: the perspective from the FAO capture database. Ocean Coastal Mgmt 43:615–655. Crawford, R.J.M., Underhill, L.G., Shannon, L.V., Lluch-Belda, D., Siegfried, W.R. and Villacastin-Herrero, C.A. (1991) An empirical investigation of trans-oceanic linkages between areas of high abundance of sardine. In: Long-term Variability

of Pelagic Fish Populations and their Environment. T. Kawasaki, S. Tanaka, Y. Toba and A. Taniguchi (eds) Oxford: Pergamon Press, pp. 319–332. Cury, P., Bakun, A., Crawford, R.J.M. et al. (2000) Small pelagics in upwelling systems: patterns of interaction and structural changes in ‘wasp-waist’ ecosystems. ICES J. Mar. Sci. 57:603–618. Finney, B.P., Gregory-Eaves, I., Sweetman, J., Douglass, M.S.V. and Smol, J.P. (2000) Impacts of climatic change and fishing on Pacific salmon abundances over the past 300 years. Science 290:795–799. Finney, B.P., Gregory-Eaves, I., Douglas, M.S.V. and Smol, J.P. (2002) Fisheries productivity in the northeastern Pacific Ocean over the past 2,200 years. Nature 416:729–733. Francis, R.C., Hare, S.R., Hollowed, A.B. and Wooster, W.S. (1998) Effects of interdecadal climate variability on the oceanic ecosystems of the NE Pacific. Fish. Oceanogr. 7:1–21. Froese, R. and Pauly, D. (eds) (2000) FishBase 2000: Concepts, Design and Data Sources. Los Ban˜os, Laguna, Philippines: ICLARM, 344 pp. Garcia, S.M. and Grainger, R. (1997) Fisheries management and sustainability. In: Developing and Sustaining World Fisheries. D.A. Hancock, D.C. Smith, A. Grant and J.B. Beumer (eds) Collingwood, Australia: CSIRO, pp. 631–654. Glantz, M.H. (ed.) (1992) Climate Variability, Climatic Change and Fisheries. Cambridge: Cambridge University Press, 450 pp. Hare, S.R. and Mantua, N.J. (2000) Empirical evidence for North Pacific regime shifts in 1977 and 1989. Progr. Oceanogr. 47:103–145. Hollowed, A.B., Hare, S.R. and Wooster, W.S. (2001) Pacific basin climate variability and patterns of Northeast pacific marine fish production. Progr. Oceanogr. 49:257–282. Hutchings, J.A. (2000) Collapse and recovery of marine fishes. Nature 406:882–885. Kawasaki, T. (1983) Why do some pelagic fishes have wide fluctuations in their numbers? Biological basis of fluctuations from the viewpoint of evolutionary ecology. FAO Fish. Rep. 293:1065–1080. Klyashtorin, L.B. (1998) Long-term climate change and main commercial fish production in the Atlantic and Pacific. Fish. Res. 37:115–125. Klyashtorin, L.B. (2001) Climate change and long-term fluctuations of commercial catches: the possibility of forecasting. FAO Fish. Tech. Paper No. 410. 86 pp. Kohonen, T. (1995) Self-Organizing Maps. Berlin: SpringerVerlag, 362 pp. Kornilovs, G., Sidrevics, L. and Dippner, J.W. (2001) Fish and zooplankton interaction in the Central Baltic Sea. ICES J. Mar. Sci. 58:579–588. Link, J.S. and Garrison, L.P. (2002) Changes in piscivory associated with fishing induced changes to the finfish community on Georges Bank. Fish. Res. 55:71–86. Lluch-Belda, D., Crawford, R.J.M., Kawasaki, T. et al. (1989) World-wide fluctuations of sardine and anchovy stocks: the regime problem. S. Afr. J. Mar. Sci. 8:195–205. Lluch-Belda, D., Lluch-Cota, D.B., Herna´ndez-Va´zquez, S. and Salinas-Zavala, C.A. (1992a) Sardine population expansion in eastern boundary systems of the Pacific Ocean as related to sea surface temperature. S. Afr. J. Mar. Sci. 12:147–157. Lluch-Belda, D., Schwartzlose, R.A., Serra, R. et al. (1992b) Sardine and anchovy regime fluctuations of abundance in four regions of the world oceans: a workshop report. Fish. Oceanogr. 1:339–347.


2003 Blackwell Publishing Ltd, Fish. Oceanogr., 12:4/5, 443–457.

Investigating synchronous patterns in fisheries

McFarlane, G.A. and Beamish, R.J. (2001) The re-occurrence of sardines off British Columbia characterises the dynamic nature of regimes. Progr. Oceanogr. 49:151–165. McFarlane, G.A., King, J.R. and Beamish, R.J. (2000) Have there been recent changes in climate? Ask the fish. Progr. Oceanogr. 47:147–169. Pauly, D., Christensen, V., Dalsgaard, J., Froese, R. and Torres F. Jr. (1998) Fishing down marine food webs. Science 279:860– 863. Polovina, J.F., Mitchum, G.T., Graham, N.E., Craig, M.P., Demartini, E.E. and Flint, E.N. (1994) Physical and biological consequences of a climate event in the Central North Pacific. Fish. Oceanogr. 3:15–21.

457

Rothschild, B.J. (1995) Fish stock fluctuations as indicators of multidecadal fluctuations in the biological productivity of the ocean. Can. Spec. Pub. Fish. Aquat. Sci. 121:201– 209. Schwartzlose, R.A., Alheit, J., Bakun, A. et al. (1999) Worldwide large-scale fluctuations of sardine and anchovy populations. S. Afr. J. Mar. Sci. 21:289–347. Spencer, P.D. and Collie, J.S. (1997) Patterns of population variability in marine fish stocks. Fish. Oceanogr. 6:188–204. StatSoft, Inc. (2000) STATISTICA for Windows [computer program manual]. Tulsa, OK, USA. Zar, J.H. (1999) Biostatistical Analysis. Upper Saddle River: Prentice Hall. 663 pp. and appendices.


2003 Blackwell Publishing Ltd, Fish. Oceanogr., 12:4/5, 443–457.