Progress in Oceanography 59 (2003) 275–300 www.elsevier.com/locate/pocean

Estimating environmental preferences of South African pelagic fish species using catch size- and remote sensing data J.J. Agenbag a,∗, A.J. Richardson b 1, H. Demarcq a,c, P. Fre´on a,c, S. Weeks b, F.A. Shillington b a

Department of Environment Affairs and Tourism, Marine and Coastal Management, Private Bag X2, 8012 Rogge Bay, Cape Town, South Africa b Remote Sensing Unit, Oceanography Department, University of Cape Town, Cape Town, South Africa c Institut de Recherche pour le De´veloppement, 213 rue La Fayette, Paris, France Revised 13 April 2003; accepted 28 July 2003

Abstract We have studied the relationship between the variations in density of South African anchovy (Engraulisi capensis), sardine (Sardinops sagax) and round herring (Etrumeus whiteheadi) from commercial catch records (1987–1997) and a suite of variables describing the environment. The indicator of density (local fish abundance) used was Catch-perset, obtained from the more than 130 000 catches made during the 11-year study period. The set of environmental parameters included: temporal (Year, Month, time of day or Hour), spatial (Latitude, Longitude and water Depth), lunar (Moon phase and Moon elevation), and thermal conditions of the environment (sea surface temperature together with indices of thermal frontal intensity and temporal change). Boat length was used to account for fishing gear effects. Previous investigations of this nature have tended to use simple bivariate correlation approaches, which suffer from the problem of covariance between the predictive variable and other environmental- or fisheries related variables not included in the analysis. We have, therefore, adopted a multivariate modelling approach, which identifies relationships between Catch-per-set and each environmental variable, accounting for covariation amongst predictors. Model building consisted of first constructing generalised additive models (GAM) as an exploratory tool to identify the shapes of the relationships, followed by parameterising these relationships using general linear models (GLM) to provide a robust predictive tool. Using this modelling approach, the suite of environmental variables explained 19.6% of the variation in Catch-per-set of anchovy, 33.9% of sardine, and 54.3% of round herring in the final GLM models. Temporal variables (Year, Month, Hour), accounted for the major part of the variability in Catch-per-set but variables derived from SST and the lunar cycle provide insight into the effects of environmental factors on fish behaviour. For instance, it appears that schooling behaviour of anchovy and round herring is affected by the level of solar and lunar illumination but that sardine is not affected. Model results further indicate that anchovies prefer water cooler than about 15 °C, demonstrates a weak tendency to concentrate near thermal fronts and avoids recently upwelled water. The sardine’s preferences seem to be more or less the opposite to anchovy, i.e. it occur further from the coast and tends to concentrate in upwelled ∗

1

Corresponding author. Tel.: +27-21-402-3309; fax: +27-21-425-6976. E-mail address: [email protected] (J.J. Agenbag). Present address: The Laboratory, Citadel Hill, Plymouth, PL1 2PB, UK.

0079-6611/$ - see front matter  2003 Elsevier Ltd. All rights reserved. doi:10.1016/j.pocean.2003.07.004

276

J.J. Agenbag et al. / Progress in Oceanography 59 (2003) 275–300

water which have been warmed to14–19 °C; it shows no discernible tendency for aggregation near fronts. Round herring Catch-per-set strongly increases with water depth and reach a maximum near the shelf edge; it also demonstrates a notable tendency to concentrate near thermal fronts. SST seems to have no influence on round herring which is most often caught in the 15–18 °C range, typical of SSTs found in the vicinity of the oceanic front.  2003 Elsevier Ltd. All rights reserved.

Contents 1.

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 276

2. Material and methods . . . . . . . 2.1. Catch data . . . . . . . . . . . . 2.2. Predictors . . . . . . . . . . . . . 2.2.1. Temporal . . . . . . . . . . . 2.2.2. Spatial . . . . . . . . . . . . . 2.2.3. Lunar . . . . . . . . . . . . . 2.2.4. SST . . . . . . . . . . . . . . 2.2.5. Fishing boat characteristics 2.2.6. Data validation . . . . . . . . 2.3. Modelling approach . . . . . . . 3.

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

279 279 280 281 281 281 282 282 282 283

Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285

4. Discussion . . . . . . . . . . . . . 4.1. Anchovy . . . . . . . . . . . . 4.2. Sardine . . . . . . . . . . . . . 4.3. Round herring . . . . . . . . . 4.4. Summary of predictor effects 5.

. . . . . . . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

288 290 293 294 296

Conclusion and perspectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 297

1. Introduction Approximately one-third of the annual global fish catch comprises of small pelagic species, such as sardine, anchovy and related species. These fish are mainly confined to the very variable environments of coastal upwelling regions. They are adapted to a variable environment and adopt strategies such as feeding on organisms from different trophic levels. In our case, the feeding strategies of anchovy and sardine have been discussed by James (1987), Van der Lingen (1994) and Louw, Van der Lingen and Gibbons (1998). These populations nevertheless appear to be extremely vulnerable to subtle shifts in the balance of physical and biological processes, particularly during the recruitment phase of their life cycles (Bakun, 1996). Understanding the effects of environmental forcing on the different development stages may be crucial to effective exploitation management of these fish resources which are subject to such large population fluctuations on interannual scales. Environmental conditions, however, also influence fishing operations on shorter temporal scales, by affecting the distribution and local abundance of fish within the fishing grounds. A better understanding of such influences is therefore of considerable financial value to the fishing industry because such knowledge will assist them to reduce the time and fuel expended by the boats in search of fish concentrations. From a resource management point of view, the ensuing greater efficiency of the commercial

J.J. Agenbag et al. / Progress in Oceanography 59 (2003) 275–300

277

operations is important as it would, at least in principle, help to make operations more profitable and thus increase the viability of smaller allocation in particular. Gaining an understanding of the relationship between the environment and fish is, however, no easy task. Part of the problem is the lack of comprehensive data on marine environmental variables and fish distributions. The variable that has been used the most frequently to predict the distribution of fish catches is sea surface temperature (SST), because it is both a biologically important variable and one which is available with good spatial- and temporal resolution from two decades of satellite generated data. There are many examples in the literature where catches and the distribution of large pelagic fish such as tuna and swordfish have been related to SST and SST fronts. Williams (1977) reported on SST maps derived from airborne radiometry as an aid to tuna fisheries off Australia. The author describes the preferred temperature range for the southern bluefin tuna (Thunnus maccoyii) and shows a tendency for the fish to concentrate near strong gradients. Similar reports followed for yellow fin- and skipjack tuna in the Gulf of Guinea (Stretta, 1977), albacore off California (Laurs, Fiedler, & Montgomery, 1984) and swordfish in the western North Atlantic (Podesta´ , Browder, & Hoey, 1993). These results have found practical application through developments in satellite technology, which made available in near real time, SST maps for locating the best fishing areas for tuna and related species (eg., Laurs et al., 1984; Fiedler & Bernard, 1987). In the case of small pelagic species, the relationship with SST is not as clear as for large pelagic fish. Anchovy spawning, for example, is thought to be related to certain temperature ranges (Lasker, Pela´ ez, & Laurs, 1981; Fiedler, 1983; Richardson et al., 1998; Van der Lingen, Hutchings, Merkle, Van der Westhuizen, & Nelson, 2001) and in the North Sea greater densities of herring are associated with gradients of SST and salinity (Maravelias & Reid, 1995). A similar association of anchovy, sardine and jack mackerel distributions with thermal fronts was demonstrated in northern Chile (Castillo, Barbieri, & Gonzalez, 1996). But there are also a number of studies that have failed to find a relationship between small pelagic fish and SST. For example, the spawning distribution of the Californian anchovy was related to temperature in some studies (Fiedler, 1983), but poorly in others (Lasker et al., 1981). In South Africa, Kerstan (1993) could only explain 2–10% of the variance in sardine population density from several environmental variables (including SST). Although there are no strong and scientifically proven relationships between SST and pelagic fish distributions, satellite thermal images are nonetheless quite commonly used by pelagic fishing fleets in an attempt to locate the best fishing locations. Such is also the practice in South Africa. Fairly strong evidence of lunar effects on the behaviour of surface schooling fish has been observed. Luecke and Wurtzbaug (1993) used hydro-acoustics to study the diel vertical migration pattern of pelagic fish in a freshwater lake and found that at night fish biomass in the pelagic zone was 45% lower during full moon than during new moon. During full moon, the fish were deeper and indistinguishable from the bottom. In another hydro-acoustic study on a freshwater species, Gaudreau and Boisclair (2000) showed that the number of fish migrating from the littoral zone to the pelagic zone was smaller by a factor of eight during full moon compared with the number during new moon. This effect was only observed during the night. Fish abundance at midday was not affected by the moon phase. In the Bay of Bengal, Pati (1981) could find no lunar effect on the gill net fishery but Di Natale and Mangano (1986) demonstrated a significant correlation between moon phase and CPUE of swordfish (Xiphias gladius) in the Italian driftnet fishery. Off Namibia, an analysis of purse-seine catches of sardine (Sardinops ocellatus), anchovy (Engraulis capensis) and juvenile horse mackerel (Trachurus trachurus capensis) demonstrated that all three species form larger schools during full moon than new moon (Thomas & Schu¨ lein, 1988). There is also the perception amongst skippers of the South African purse seine fleet that more round herring are caught at full moon. The South African pelagic fishery is dominated by three species, anchovy (Engraulis capensis), sardine (Sardinops sagax) and round herring (Etrumeus whiteheadi). These species are found along most of the South African coastline, with no clear geographical separation between the populations (Fig. 1). Juveniles shoal together in the cool, nearshore upwelled waters of the West Coast during autumn and winter, but

278

J.J. Agenbag et al. / Progress in Oceanography 59 (2003) 275–300

Fig. 1. Location of the study area with an example of the SST weekly composites maps used (a), and the distributions of mean annual catches of anchovy (b), sardine (c) and round herring (d) after kriging the data from the initial 10⬘ × 10⬘ fishing grid.

J.J. Agenbag et al. / Progress in Oceanography 59 (2003) 275–300

279

thereafter develop greater behavioural differences as they grow older (Crawford, 1981; Armstrong, 1986; Roel & Armstrong, 1991). Anchovy tend to remain more inshore and thus in cooler waters than either of the other two species (Hampton, 1987). Mature sardine inhabit most of the shelf width, often aggregating near the shelf edge (Anders, 1975; Armstrong, Berruti, & Colclough, 1987). Adult round herring also tend to occur in deeper, warmer water near the shelf edge. There are also marked dietary differences among the three species, with sardine being the most dependent on phytoplankton (Roel & Armstrong, 1991; James, 1987, 1988), anchovy using both phyto and zooplankton, and round herring being entirely zooplanktivorous (Wallace-Fincham, 1987). Sardine and round herring are very similar in appearance, both growing to a mature size of about 24 cm. Mature anchovy, however, only reach a length of about 14 cm and because of this size differentiation, adult sardine and round herring commonly shoal together but mainly mix with anchovies during their juvenile stages. With the current study we aim, firstly, to gain a better understanding of the environmental preferences of the three species (anchovy, sardine and round herring) and, secondly, to express these preferences as mathematical functions which might be used in conjunction with SST images to predict the locations where the higher densities of fish are to be expected. For this purpose, we use commercial catch data to quantify fish density, together with a suite of temporal, spatial, lunar and SST variables in a combination of generalised additive models (GAM) and general linear models (GLM), procedures which are now used with increasing frequency within the fisheries context (Swartzman, Huang, & Kaluzny, 1992; Borchers, Buckland, Priede, & Ahmadi, 1997; Maravelias & Reid, 1995; Fox, O’Brien, Dickey-Collas, & Nash, 2000). Our study is by no means the first to examine environmental effects on South African pelagic fish, but previous studies (e.g. Cruickshank & Boyd, 1985; Thomas, 1986; Armstrong et al., 1987; Shannon, Crawford, Brundrit, & Underhill, 1988; Armstrong, Chapman, Dudley, Hampton, & Malan, 1991; Roel & Armstrong, 1991; Kerstan, 1993) used a correlative-type approach with only one predictor (environmental variable), ignoring covariation with other potential predictors. The GAM and GLM procedures, by contrast, are multivariate model-building procedures which adjust for covariance amongst a suite of pertinent predictors. We employed GAMs as the principal tool for studying functional relationships between the fish and predictors because the procedure accommodates a very wide range of functional forms and requires no assumptions as to the shape of the functions. GLM is a statistically more robust procedure than GAM and also produces less complex mathematical expressions, more suited for predictive purposes, but is requires preliminary exploration of the functional relationships. In our two-step approach, this requirement was met through the use of the GAM.

2. Material and methods 2.1. Catch data Data on pelagic fish density are only available from two sources, i.e. from scientific hydro-acoustic surveys and from records of commercial catches. There are biases associated with both sources, with hydroacoustic surveys having good spatial but poor temporal coverage, whereas commercial catches tend to have good temporal but poorer spatial coverage. For this study, we chose to use the much larger database of catch records available at Marine and Coastal Management (M&CM, South African Department of Environment Affairs and Tourism). The investigation was limited to the period 1987–1997 for which satellite data were available and for this period 133 382 catch records were available, containing information on the time, location, catch weight and species composition of every catch made by the 60–100 purse-seine vessels in the pelagic fleet. Because the exact geographical co-ordinates are often not known for catch positions, they are recorded in terms of a grid system which divides the fishing grounds into 10⬘ latitude × 10⬘ longitude squares (Fig. 1). Fishing boats commonly make several sets (catches) before returning to the

280

J.J. Agenbag et al. / Progress in Oceanography 59 (2003) 275–300

factories for offloading and at this point the weight of all fish in the hold is accurately determined. For our purposes, we need the weights of the individual catches, hence we have used estimates of the individual catch weights reported by the skippers. These estimates are surprisingly accurate, but provide no indication of the species composition of the catch. This is determined by sampling of the hold content during the offloading process. If more than one catch is made before offloading then all catches are assigned this ‘average’ species composition. If we assume that fish’s preference for environmental conditions is reflected by geographical variations in the population density, then the next question is how best to define variations in density using catch records? The most obvious and frequently used index of abundance is catch per unit of effort (CPUE). However, not only do the data at our disposal not contain information on search time, making it difficult to estimate effort, but the concept of effort is problematic in the pelagic fishery as there is considerable communication among skippers. We performed simple correlations between SST variables (SST, SST gradient and SST temporal change) and three other potential density indices, viz. catch-per-unit-area (ton per unit area), number-of-catches-per-unit-area and catch-per-set, to determine the best. While the tests demonstrated that, on average, similar relationships were obtained between the three indices and SST variables (not shown), the test results also demonstrated much variation in the relationships over shorter time- and spatial scales, thereby emphasising the need for a multivariate approach. The first two indices involves spatial integration which, if computed from the catch data, implies also temporal integration. They are therefore less suited for a multivariate analysis than catch-per-set which is based in the properties of individual catches. Also, it should be borne in mind that fishing operations radiate from a relatively small number of fishing ports which tends to impart a strong spatial bias to these two indices. Consequently, Catch-perset (the tonnage of fish caught in a single haul) was selected as an index of fish density. This selection is supported by the observation that increases in fish populations generally lead to the formation of bigger schools and increases in catch-per-set in purse-seine fisheries (Fre´ on, 1986, 1991; Wada & Matsumiya, 1990). Assuming that a similar relationship also exist at intermediate spatial scales we extrapolate this result to reach the conclusion that Catch-per-set may be used as an indicator of fish density. Finally, a caveat is necessary. Pure catches of a single species are quite common in the data set and raise the question of whether the corresponding zero- or null catches for other species are indicative of low density or not. As shall be discussed in greater detail later, the South African pelagic fishing fleet target specific species to a very great extent and not only are the skippers skilled in identifying shoals of the target species but also avoids fishing in areas where recent catches mainly consisted of non-target species. The fishing industry also functions in an organised and self-regulatory manner to perform test catches and to close parts of the fishing grounds, particularly to avoid fishing on mixed schools of anchovy and juvenile sardine. Under such circumstances it seem that the zero catches would primarily arise from targeted fishing and are likely to introduce artefacts in the models. We expect relationships involving temporal predictors to be particularly prone to distortion because, for any one of the species, the number of zero catches will be inversely proportional to the ratio of its own abundance to the combined abundance of the other two species. Zero catches were consequently not used in the construction of our statistical models. 2.2. Predictors The suite of predictors included those related to the time and location of the catches, those related to fishing gear performance and environmental variables. The latter group consisted of lunar- and SST-related variables, variables known for the time and location of the catches and linked to variations in catch size through biologically meaningful processes. Although a plethora of other variables such as oxygen concentration, food abundance, and reproductive activity could also play a role in fish distribution, there were no

J.J. Agenbag et al. / Progress in Oceanography 59 (2003) 275–300

281

in situ data available on appropriate time and space scales. After removal of redundant variables (for example, Longitude redundant with Depth), a suite of 11 variables remained (Table 1). 2.2.1. Temporal Three temporal factors were used: Year, Month and Hour (time of day). By considering the Year effect, we adjust the model for interannual variations of school size (Catch-per-set) arising from variations in the overall population size. Moreover, as both catches and SST vary seasonally, by including the Month effect we account for possible seasonal variations in behavioural or migration patterns that are not attributable to temperature. Hour was included as a variable, since many fish species perform a diel vertical migration which, coupled with a different shoaling behaviour, makes them more available to the fishery at certain times of the day. 2.2.2. Spatial Two spatial indices were used: water column depth (at the centre of the grid square) and Latitude. Depth was used because survey results and catch distribution (Fig. 1) show a tendency for the three species to inhabit different depths. Moreover, because of the substantial covariation between SST and Depth (SST generally increases as Depth increases), a more accurate representation of the relationship between Catchper-set and SST variables could be obtained by including Depth in the model. Because most of the catches used in the analysis were from the west coast of South Africa, Longitude was highly correlated with Depth and therefore not included. 2.2.3. Lunar Lunar data were obtained from the Solar System Dynamics Group, Horizons On-Line Ephemeris System, 4800 Oak Grove Drive, Jet Propulsion Laboratory, Pasadena, CA 91109 USA. http://ssd.jpl.nasa.gov/cgibin/eph. The ephemeris was generated for sea level at co-ordinates: 17°30⬘00.0⬙E, 32°00⬘00.0⬙S, at twohourly intervals and linearly interpolated to the time of fish catches. To identify possible lunar effects we

Table 1 A summary of the parameters used in the GAM and GLM analyses Type

Parameter

Explanation

Units

Mean ± std

Range used (min, max)

Response

Anchovy catches Sardine catches Round herring catches Year Month Hour Latitude Depth

Catch per set Catch per set Catch per set Year of the catch Month of the catch Hour of the catch Latitude of the catch Depth at the centre of the grid cell Boat length % of disk illuminated Angle of elevation above horizon Mean SST Standard deviation of SST Temporal change in SST from previous image

ton ton ton

30.5 ± 26.2 9.5 ± 19.1 12.7 ± 18.9

hours °S m

32.9 ± 0.8 76.5 ± 52.2

0.1; 295 0.1; 254 0.1; 219 1987–1997 January–October 0; 24 30.4; 34.9 5; 290

m % degree

23.1 ± 3.8 44.3 ± 33.9 ⫺7.5 ± 38.2

11–37 0; 100 ⫺87; 86

°C °C °C

15.2 ± 1.7 0.41 ± 0.32 ⫺0.04 ± 1.1

10.3; 23.4 0; 2.99 ⫺5.1; 5.6

Predictors

Boat length Moon phase Moon elevation SST SSTstd SSTdif

282

J.J. Agenbag et al. / Progress in Oceanography 59 (2003) 275–300

have included two variables related to the lunar cycle: the Moon Elevation (angle above the horizon) and the Moon Phase (% of lunar disk illuminated, which is almost identical to the moon light intensity). 2.2.4. SST To characterise the thermal environment where catches were made, a number of variables were derived from weekly AVHRR SST composites produced at the Space Applications Institute (SAI) of the European Joint Research Centre (JRC) in Ispra, Italy (Cole & Villacastin, 2000). Data were provided through the European Union’s ENVIFISH programme. Images were transformed to a cylindrical equidistant (rectangular) projection with a spatial resolution of about 4.5 km and SST resolution of 0.1 °C. As catch locations are recorded on a 10⬘ latitude × 10⬘ longitude grid system, SST variables were calculated on the same grid scale. Although the image compositing procedure generated a relatively cloud-free product, residual cloud was still present in most images. Pixels obscured by clouds were excluded when calculating the averages and a variable was regarded as undefined if there were fewer than three good pixel values in the particular square. We initially calculated five SST variables, the mean SST for a grid square and four SST derivatives of which the first two were designed as scalar expressions of thermal gradients or more generally as indications of frontal intensity. The first was the thermal gradient: SSTgrad (°C) = √[(Ti + 1⫺Ti⫺1)2 + (Tj + 1⫺Tj⫺1)2] where T is the SST (°C) for the (i,j) coordinates of the pixel where the gradient is computed. It follows from the nominal 4.5 km image resolution that the distance considered when computing SSTgrad is about 9 km. The other estimate of thermal gradient was the standard deviation of SST (SSTstd) in a 3 × 3 pixel area centred on position (i,j). Because of the relatively low spatial resolution of SST images and the smoothing effect of the image compositing procedure, SSTgrad proved to possess a poor dynamic range compared with SSTstd and was therefore dropped from further analyses. The other two variables reflected changes in SST over time. The first is the temporal change of SST between two successive images: SSTdif (°C) = SSTthis week⫺SSTprevious week. Thus, SSTdif measure whether an area had been warming or cooling over the previous week. The second variable is the change in SST between the current image and that of the 1982–1997 climatology: SSTano (°C) = SSTthis week⫺SST1982–1997 climatology. Although these two variables are similar in concept, SSTdif represents a change in temperature over a short period of time and is therefore directly interpretable, whereas SSTano is a change relative to the long-term mean and is more difficult to interpret. In addition, initial exploration showed that the relationship between Catch-per-set and SSTano was very weak compared with that of SSTdif, so that SSTano was dropped from further analyses. 2.2.5. Fishing boat characteristics Boats of the South African purse seine fleet have a diverse range of characteristics, including size of vessel (11–37 m), crew size, engine power, net handling gear, freezer equipment and fishing strategies. It is therefore inevitable that Catch-per-set is influenced by the variable capabilities of the boats. Four parameters were originally considered to quantify these effects. Two of these, boat length and gross tonnage, are directly related to the physical characteristics of the boats. The other two, carrying capacity (maximum recorded offload) and fishing capacity (the maximum recorded catch) are related to the historic performance of the boat. These four measures of the capacity of the fishing boat were strongly correlated, and as a result we used only Boat length, which was the simplest and most direct measure of the boats’ capabilities. 2.2.6. Data validation The total number of catch records for the period 1987–1997 was 133 382 from 57 307 trips, an average of 2.33 catches per trip. Removal of those for which data fields were undefined (mainly SST indices caused by residual cloud in the images), reduced the data set to 120 063 records. A process of removing extreme values was then performed (see Table 1 for valid ranges). Because annual fishing quotas are normally filled by the last months of the year, catch frequency as well as tonnages become very low (Fig. 2). This

J.J. Agenbag et al. / Progress in Oceanography 59 (2003) 275–300

283

Fig. 2. Mean seasonal variability of anchovy, sardine and round herring catches from the South African fisheries for the years 1987– 1997 (shown as bar graphs and also as smoothed curves).

is most probably accompanied by changes in fishing strategy and therefore only data for the months January to October were retained. Finally, the zero catches were discarded which left 74 352, 63 670 and 40 653 records of anchovy, sardine and round herring catches, respectively. 2.3. Modelling approach To identify relationships between Catch-per-set and the suite of predictors describing the environment, we used the following modelling approach. As the expected relationships are likely to be non-linear, we first used a generalised additive modelling (GAM) approach. An additive model extends the linear model by allowing the linear functions of the predictors to be replaced by smooth functions of these predictors. The general form of the linear model,

冘

bXi ⫹ e

(1)

冘

fi(Xi) ⫹ e

(2)

n

Y⫽a⫹

i⫽1

becomes, n

Y⫽a⫹

i⫽1

where Y is the response, Xi are the predictors, a and b are constants, and e is the error. The fi are generally unknown and are estimated using scatterplot smoothers (MathSoft, 2001). Scatterplot smoothers enable the description of the relationship between a response and a predictor without imposing an a priori shape, by contrast to the linear modelling where a specific functional relationship has to be defined. In this study, we used least squared weighted smoothers (loess) to estimate these non-parametric functions. The loess smoother is equivalent to a suite of independent local linear regressions, each being applied over a small interval, defined as the span of the smoother. The degree of smoothing is dependent upon this span, the proportion of the data points that is used for the smoothing. We used the default initial span value of 2/3 in S-Plus (MathSoft, 2001), but this span is then iteratively adjusted by S-plus for every predictor until it reaches an optimal value for the model, so that the choice of this initial value is not critical. The response variable (Catch-per-set of anchovy, sardine or round herring) was transformed using a conventional log-transformation to reduce the skewness of the distributions. A visual analysis showed that this transformation also resulted in a normal distribution of the data, a necessary assumption for GLM. In addition, the GLM residuals were checked and found to conform to normal distributions.

284

J.J. Agenbag et al. / Progress in Oceanography 59 (2003) 275–300

Once the shape of the relationships between the response variable and each predictor was identified for each species using GAM (Figs. 3–5), simplified functions were used to parameterise these shapes in the general linear models (GLM). GLM was used to provide predictive equations. In defining simplified functions, the shapes obtained from the GAMs were reproduced as closely as possible using piecewise linear or polynomial regressions. Break points were inserted in some cases where clear breaks appeared to exist within regressions (Fig. 6). In building the GAMs and GLMs we tried to use continuous predictors where possible, rather than categorical predictors which would have entailed arbitrary definition of category limits. Categorical predictors also require more parameters to describe non-linear relationships and this would have made the predictive equations cumbersome and not very portable. Obviously, Year and Month effects had to be included, however, as categorical variables in all models. The range of variation of the smoothed terms obtained from the GAM models provide an indication of the relative importance of each of the predictors in explaining the observed variation of the response variable. Removing low contribution variables in GAM was validated by a step-wise GLM analysis in order to confirm redundancy. Because GAMs were used primarily to describe the shapes of the relationships between the response and each predictor, interactions between the predictors were not included in the GAMs, as it is computationally difficult to perform robust regressions with surfaces and 11 predictors on our large data set. Some interactions were, however, included in the

Fig. 3. Anchovy GAM model results. Functional relationships between anchovy Catch-per-set and predictors obtained from the model (r 2 = 12%).

J.J. Agenbag et al. / Progress in Oceanography 59 (2003) 275–300

285

Fig. 4. Sardine GAM model results. Functional relationships between sardine Catch-per-set and predictors obtained from the model (r 2 = 25%).

GLMs, but because of the very large number of possible interactions we only investigated those involving interannual and interseasonal changes in the relationships between Catch-per-set and the SST-derived predictors and spatial predictors. Table 2 shows that the interactions which explained a significant, though sometimes very small part, of the variance were the cross effects, Year:Month, Latitude:Year, Latitude:Month, Depth:Year, Depth:Month, SST:Year and SST:Month.

3. Results Outputs from the GAMs for anchovy, sardine and round herring are shown in Figs. 3–5, respectively. Each plot illustrates the non-linear relationship between the response variable (log of the Catch-per-set) and each predictor (adjusted for all other predictors in the model). The y-axis is a relative scale, so that a y-value of zero is the mean effect of the adjusted environmental variable on the response; a positive yvalue indicates a positive effect on the response, and a negative y-value indicates a negative effect on the response. As the range of the smoothed function indicates the relative importance of each predictor, all yaxes have been adjusted to approximately the same range of 2.0. Also shown in the graphs are the 95% confidence limits. These lines tend to diverge near the extremes of the range for continuous predictors as

286

J.J. Agenbag et al. / Progress in Oceanography 59 (2003) 275–300

Fig. 5. Round herring GAM model results. Functional relationships between round herring Catch-per-set and predictors obtained from the model (r 2 = 41%).

a consequence of fewer observations. Thus, the relative importance of each predictor should be judged predominantly over the range where confidence limits are narrowest. Results of GLMs were examined in the same graphical manner as the GAM results, but when displayed graphically the two sets of results are very similar and we, therefore, only present some examples of anchovy GAM versus GLM relationships (Fig. 6). In the GAM as well as the GLM models, all predictors were significant at the 99% confidence level, as is usually the case when very large sets of spatially and temporally auto-correlated data are used. It is more relevant to consider the relative contributions of the predictors, as provided by the GLM models and summarised in Table 2, to judge the meaning of a factor in the model. GLMs explain 20%, 34% and 54% of the total variance of Catch-per-set for anchovy, sardine and round herring, respectively. These were obtained after introduction of interactions (cross effects) between some of the predictors. Without interactions the GLM models for the three species explain 11%, 24% and 39% of the variance which is only slightly less than the 12%, 25% and 41% obtained with GAM models for the same species. This is a consequence of adjusting the predictors in the GLM models to be close approximations of the shapes of the smoothed relationships obtained with the GAM models. Among the interactions introduced in the GLM models (Table 2), the Year∗Month interaction had by far the largest influence. SST related parameters explained only small parts of the total variance compared to other parameters,

J.J. Agenbag et al. / Progress in Oceanography 59 (2003) 275–300

287

Fig. 6. Typical translations of GAM-derived relationships (left column) into simplified GLM functions (right column): (a) a second order parabolic function, (b) a fourth order polynom, (c) a linear function and (d) a linear function with a break point.

while the greatest contribution was made by the temporal variables (Year, Month, Year∗Month and Hour), which contributed between 60% and 65% of the explained variance. An example of the GLM formula obtained is given below for anchovy: log(Catch ⫺ per ⫺ set anchovy)(Year, Month, Hour, Latitude, Deth, Moon—phase, Moon—elevation, SST, SSTdif) ⫽ ⫺0.0087 ⫹ aYear ⫹ bMonth⫺4.43∗Hour⫺26.95∗Hour2⫺26.93∗Latitude⫺20.90∗Latitude2

288

J.J. Agenbag et al. / Progress in Oceanography 59 (2003) 275–300

Table 2 Percentage of variance explained by the GLM parameters (obtained from the GLM deviance tables)

Year Month Year:month Hour Latitude Latitude:year Latitude:month Depth Depth:year Depth:month Boat length Moon phase Moon elevation SST SST:year SST:month SSTstd SSTdif Total a b

Anchovy

Sardine

Round herring

3.7 0.7 6.0 1.8 1.9 0.8 0.4

12.1 1.6 6.7

6.1 14.0 10.4 5.1 8.7 1.6 0.6 3.7 1.2 0.4 0.3 1.0 0.1

a

4.4 1.6 0.6 1.4 0.5 0.3 2.8

a

0.4 0.3 1.3 0.4 0.1 0.9 0.2 0.2 0.1 0.3 19.6

a a

0.8 0.6 0.2

a

0.4

0.7 0.2 0.2 0.1

33.9

54.3

b

Effect not included (r 2 ⬍ 0.05). Effect removed by the stepwise procedure.

⫹ 0.00019∗Depth ⫹ [30.92∗Boat—length⫺38.67∗Boat—length2] ⫹ 0.0015∗Moon—phase ⫹ 0.0011∗Moon—elevation⫺91.34∗SST⫺4.69∗SST2 ⫹ 0.019∗SSTstd ⫹ 9.90∗SSTdif(SSTdif具2); (3) ⫺0.43∗SSTdif(SSTdif典2)] ⫹ c(Year∗Month) ⫹ d(Lat∗Year)∗LAT ⫹ e(Lat∗Month)∗LAT ⫹ f(Depth∗Year)∗DEPTH ⫹ g(Depth∗Month)∗DEPTH ⫹ h(SST∗Year)∗SST ⫹ i(SST∗Month)∗SSTe(Year, Month, Hour, Lat, Depth, Moon—phase, Moon—elevation, SST,SSTdif) where: aYear and bMonth are the parameter values for the Year and the Month categorical variables; c(Year∗Month) is the parameter value accounting for the cross effect of a given combination of Year and Month; d(SST∗Month) are the slopes value for the between effects considered; e(Year, Month, Hour, Lat, Depth, Moon—phase, Moon—elevation, SST, SSTdif) are the residual values. In the following section, we shall use the GLM results (Table 2) in conjunction with the relationships obtained from the GAM models (Figs. 3–5) to examine the influence of each of the predictors on Catchper-set variability and because most effects vary with species, we choose to discuss the results by species rather than by parameter or group of parameters.

4. Discussion We recognise that Catch-per-set is only an approximation of school size and that factors such as the characteristics of the fishing gear and boat may also affect catch size and we have compensated for this

J.J. Agenbag et al. / Progress in Oceanography 59 (2003) 275–300

289

in our models. There are, however, also other factors which affect this approximation but for which we cannot compensate in the model. Firstly, there is the problem of species composition in the case of a boat making several catches before offloading (discussed in Section 2.1). In recent years, an observer system was introduced which provides for sampling of the catches at sea but the system was not yet in operation during the 1987–1997 period. A second category of factors which could not be compensated for are those related to fishing strategies in response to the management of the mixed species fishery. To gain some insight into these it is necessary to consider briefly some aspects of the management and operation of the South African pelagic fishery. The fishery originally developed as a sardine fishery, but during the 1960s the sardine stock suffered a major collapse and the fishery has since been regarded and managed as an anchovy fishery. The sardine nevertheless remained the more valuable resource, being canned for human consumption and thus fetching higher prices by weight than any of the other species, which are almost exclusively used for oil and meal production. Annual catches of anchovies and sardines are controlled through a total allowable catch (TAC) or quota, fixed for each species separately, and annually adjusted according to the sizes of the stocks as determined by acoustic surveys (Hampton, 1992). The situation is complicated by juvenile sardine often schooling together with anchovy so that it is almost impossible to fish for anchovy without also catching significant quantities of sardine. As a result, the sardine TAC is split into two parts, a sardine ‘directed TAC’ aimed at catching relatively pure adult sardine for canning and a sardine ‘by-catch TAC’ which allows for juvenile sardine caught during anchovy directed fishing (Cochrane, Butterworth, De Oliveira, & Roel, 1998; De Oliveira, Butterworth, Roel, Cochrane, & Brown, 1998). Finally, because management stresses conservation and rebuilding of the valuable sardine resource, all fishing is terminated for the year once either of the two sardine TACs is reached. This stipulation has repercussions on our analysis based on Catch-per-set. Sardine catches are controlled by the industry management, in order to avoid overloading the canning facilities (which would mean turning valuable sardine into less profitable meal) and to avoid loosing part of its anchovy TAC by prematurely filling the sardine TAC. The consequence is that Catchper-set is probably a poorer approximation of school size for sardine than for the other two species. Catchper-set is probably a good proxy for round herring density, as it is a non-quota species, which means that there is effectively no restriction on exploitation and the fishing industry strives to augment its earnings by maximising catches of round herring whenever possible. The pelagic fishing fleet size decreased during the period of this study from 85 boats in 1987 to 60 in 1997 and range in size from 11 to 37 m. Smaller boats (⬍20 m) catch sardine as bait for line fishing. They carry ice and/or refrigeration equipment to keep the fish in good condition. As these boats deliberately make small catches only (⬍10 ton), Catch-per-set may in this case not be a good approximation of school size and hence fish density. Bigger boats fall in roughly two categories, based on whether or not they have cooling facilities. Boats without cooling facilities generally only catch anchovy, round herring and sometimes juvenile sardine for meal/oil production; Catch-per-set is expected to be a good indication of school size in this case. Boats with cooling facilities have a dual purpose: they are ordinarily the only vessels used to target sardine for canning, but they also fish for anchovy and round herring. Although these refrigerated boats cannot be distinguished from non-refrigerated boats on the basis of boat size alone, there is a tendency for the bigger and more modern boats to be refrigerated. Therefore, when fishing capacity (the maximum recorded Catch-per-set) is plotted as a function of Boat length (Fig. 7), three fairly distinct clusters are obtained: (1) Boats in the length range ca. 11–19 m are primarily bait boats and the maximum catch size increases slightly with boat size, (2) in the range 19–26 m, there is a tendency for the maximum catch to increase steeply with boat size, probably a consequence of boat size constraints on anchovy catches primarily, (3) beyond 26 m the trend flattens, perhaps because the limit of school size has been reached (saturation) or because these boats focus on sardine so that catch size is limited by managerial considerations.

290

Fig. 7.

J.J. Agenbag et al. / Progress in Oceanography 59 (2003) 275–300

Fishing capacity as a function of Boat length showing the three main categories of boats in the South African pelagic fleet.

4.1. Anchovy Temporal effects (Year, Month, Year∗Month and Hour) account for 12% of the total variance (63% of the variance explained by the GLM) and are, in combination, the most important predictor category. The largest contribution comes from the interannual variability of Catch-per-set and is reflected by the contribution of the Year effect (3.7% of the variance) which shows a decreasing trend from 1987 to1996 (Fig. 3). The annual anchovy biomass as estimated by acoustics (Fig. 8) is more variable than the mean annual Catch-per-set but the trends are similar and therefore leads us to surmise that the influence of the Year parameter is at least partly derived from a relationship between population size and Catch-per-set as was previously reported for other pelagic fisheries (Fre´ on, 1986, 1991; Wada & Matsumiya, 1990). The Month effect is comparatively moderate (i.e. there is only a weak seasonal effect in the Catch-per-set variation) with slightly bigger Catch-per-set values at the beginning of the fishing season. The trend bears some similarity to the mean seasonal catches, which tend to decrease after June (Fig. 2). GLM results (Table 2) indicate that the Year∗Month interaction is the most important predictor in the anchovy model (6.0% of the variance), which means that the seasonal variance of the Catch-per-set is not a fixed pattern but changes interannually. While there are many factors which could affect the seasonal pattern, the two most obvious ones are: variations in the spawning/recruitment cycle and stock-size fluctuations. On average, 70% of the annual tonnage of anchovy caught consists of fish less than one year old (Butterworth & Berg, 1993) which implies that the seasonal Catch-per-set pattern would be altered by early- or late arrival of the recruits, as well as by their relative abundance. Examination of the Catch-per-set seasonal trends of the individual years reveals a change in the seasonal pattern during the years when the anchovy stock was at lowest levels, particularly 1989, 1990 and 1996 (Fig. 8). In each of these years Catch-per-set was noticeably reduced during part of the year. Although this observation might suggest that the data set be partitioned into two parts (high and low stock size) the effect is not regarded as sufficiently consistent to allow sensible partitioning. A clear and relatively strong predictor for anchovy is the hour of the catch: although anchovy are actually caught in slightly greater quantities during the night than during the day (a tendency which is more pronounced in the other two species), the GAM results show that the Catch-per-set is larger during the day with a maximum around midday. This relationship is partly in agreement with Thomas and Schu¨ lein’s (1988) results for the Namibian anchovy. They found the catch frequency (the number of pure catches per hour) to be highest during the day (10h00) and that adult anchovy formed fewer but larger schools during the day. They could, however, find no meaningful day/night differences for juveniles. Our data set does not allow us to discriminate between catches of adults and juveniles. It is also interesting to note that

J.J. Agenbag et al. / Progress in Oceanography 59 (2003) 275–300

Fig. 8.

291

A comparison of mean annual Catch-per-set with biomass estimates from acoustic surveys for anchovy (a) and sardine (b).

Catch-per-set is at a maximum at a time when feeding activity is at a minimum (Armstrong, James, & Valde´ s Szeinfeld, 1991), which perhaps implies that the fish tend to disperse in smaller schools when feeding. Among the ‘spatial’ predictors, the latitudinal (or regional) predictor essentially indicates that anchovy Catch-per-set is highest, and fairly uniform, within the West Coast fishing grounds but decreases to the north of about 32° and also on the Agulhas Bank, i.e. south of about 34°S (Fig. 3)—a pattern also seen in the mean spatial distribution of Catch-per-set (Fig. 9). The relationship has a weak relative minimum between 32°S and 33°S, which corresponds with the location of St. Helena Bay where, as shown in Fig. 1, on average the largest quantities of anchovy are caught. These apparently contradictive results most probably arise from the spatial bias imposed on the distribution of the mean annual catch by the three fishing ports located in this St. Helena Bay. In other words, the fishers probably find a trade-off between distance to sail and catch size. This conclusion is supported by the observation that, although less pronounced, local maxima in the mean annual catch also occurs in the vicinity of the other fishing ports, i.e. Saldanha Bay, Cape Town/Hout Bay and Hermanus/Gans Bay (Fig. 1), while Fig. 9 shows that these areas are not associated with a noticeably increased mean Catch-per-set. In the GLM model, Latitude accounted for 1.9% of the Catch-per-set variance while the Latitude:Year and Latitude:Month interactions explained 0.8% and 0.4% of the variance, respectively (Table 2). The relative importance of the interactions indicates that the spatial character of Catch-per-set varies interannually and seasonally as one would expect.

292

Fig. 9.

J.J. Agenbag et al. / Progress in Oceanography 59 (2003) 275–300

Mean Catch-per-set per 10⬘ × 10⬘ grid square for anchovy (a), sardine (b) and round herring (c) averaged for the period 1987–1997.

The other ‘spatial’ predictor, Depth, has no significant effect on the anchovy GLM model (Table 2), while the GAM relationship (Fig. 3) shows that Catch-per-set is almost completely invariant with depth up to about 100 m whereafter there is a decreasing trend, but relatively few anchovy catches (⬍3%) are made beyond 150 m. The relationship thus reflects the known tendency for anchovy to shoal in nearshore waters (Hampton, 1987) and can also be seen in the distribution map of mean annual catch (Fig. 1). Acoustic spawner biomass surveys have found indications of an eastward and offshore movement of adult anchovy on the Agulhas Bank (Roel, Hewitson, Kerstan, & Hampton, 1994). This tendency is not shown by our results, probably because commercial catches are mainly made on the West Coast with only a relatively small proportion on the Agulhas Bank and even fewer at the end of the year when the spawners are concentrated on the Agulhas Bank. It was anticipated that Catch-per-set would be strongly influenced by the predictor boat length but it turned out that its contribution to the variance explained by the GLM model is smaller than expected (1.3%, Table 2). The reason for this is, however, evident from the GAM relationship (Fig. 3). In this curve, the three classes of fishing vessel previously discussed are clearly seen, i.e. the length range 11–19 m

J.J. Agenbag et al. / Progress in Oceanography 59 (2003) 275–300

293

which mainly consists of bait boats fishing for sardine, the 20–26 m range consisting of the ‘anchovy’ boats and finally the large refrigerated boats (⬎ 26 m) which tend to specialise in catching sardine for canning purposes. As indicated by the confidence lines, most of the anchovy catches (about 80%), were made by the boats in the relatively small length range 20–26 m and this has the effect of reducing the influence of boat length on the model. The predictor Moon phase, which is directly proportional to moon light intensity, contributes 0.4% to the variance explained by the GLM model (Table 2) and shows a slight positive trend (Fig. 3). The moon elevation effect shows a similar weak positive trend for positive elevation angles (i.e. when the moon is above the horizon) and explains 0.1% of the variance. Although the lunar effect is obviously quite small, results show that anchovy Catch-per-set increases slightly with increasing levels of lunar illumination. This result is in agreement with Thomas and Schu¨ lein (1988) and also with our observation of increased Catchper-set during day time, which suggests that there is perhaps a general positive effect of light level on anchovy schooling behaviour. SST and related parameters have relatively small effects on the GLM model (1.7%). GAM results (Fig. 3) nevertheless show a strong negative relationship with SST greater than about 15 °C; in colder water the slope is effectively zero. At 0.9% of the explained variance, SST is the most important of the SST-related variables in the GLM model (Table 2) and its relationship with Catch-per-set is in agreement with the known tendency for anchovy to shoal in cool nearshore waters (Hampton, 1987). We also included in the model the interactions between SST and Year and Month to account for interannual and seasonal variability; the effects are only moderate (in each case 0.2% of the variance), indicating that the SST relationship is interannually and seasonally fairly stable. A weak positive trend is obtained with SSTstd in the 0.0–1.0 °C range (Fig. 3). Since SSTstd is seen as a measure of horizontal temperature gradients, the result suggests there is not a strong tendency for anchovy to concentrate in frontal regions, which is perhaps not surprising considering that strong gradients are not expected to occur in the cool coastal waters inhabited by this species. SSTdif, the difference in SST between the current week and the previous week, explains 0.3% of the variance in the GLM model because of the moderate positive slope over the main data range, ⫺2.0 to +2.0 °C (Fig. 3). Beyond +2.0 °C, the slope is strongly positive. Within the study area, SST temporal changes, on the time scale of a week, are predominantly associated with the pulsing of Ekman upwelling so that the SSTdif relationship seen in conjunction with the SST relationship implies that while anchovies generally prefer colder SSTs, they also tend to avoid areas of newly upwelled water. 4.2. Sardine As was the case for anchovy, temporal variables make the greatest contribution (20.5%) to the total variance explained by the GLM model (33.9%) and of these the Year effect is the most important (12.1%). The mean annual sardine Catch-per-set demonstrates a positive trend similar to the annual biomass estimated by acoustics (Fig. 8) which leads us to the same conclusion reached for anchovy, i.e. that effect of the Year parameter is at least partly derived from a relationship between population size and Catch-perset. Sardine Catch-per-set is relatively low over the first three months and then increases to a plateau in July and finally to very high values in September and October (Fig. 4). The contribution of the Year∗Month interaction is the second highest of the GLM model predictors (6.7%, Table 2) indicating that there is substantial interannual variation in the seasonal trend of Catch-per-set. The sardine fishery, unlike anchovy, is not heavily dependent on interannual variations in recruitment and the seasonal variations associated with the recruitment process. Although there are indications that the seasonal trend might be affected by the sardine population size, these are not sufficiently consistent to be advanced as a plausible explanation for strong Year∗Month interaction effect. It is interesting to note that sardine catch size is apparently not affected by Hour (Fig. 4). Examination of the data, however, showed that 72% of the sardine catches were made during the night, which concurs with Thomas and Schu¨ lein’s (1988) findings for the

294

J.J. Agenbag et al. / Progress in Oceanography 59 (2003) 275–300

Namibian fishery. But they also found that, contrary to our result, individual catches significantly increased during the day. As far as the spatial aspects are concerned, sardine catch size increases quite regularly from north to south (Fig. 4) and also slightly with Depth (Fig. 4). The latitudinal change is only slightly discernible in the distribution of mean Catch-per-set (Fig. 9), but the Depth effect is clearly seen in the more offshore location of the higher values. Both the GAM results and Fig. 9 thus confirm the previously recorded tendency for sardine to aggregate in deeper water near the shelf edge (Anders, 1975, Armstrong et al., 1987). The contributions of both the Latitude and Depth effects are greater in the GLM for sardine (4.4% and 1.4%, Table 2) than in the anchovy model, but the temporal interactions with these variables are generally very similar for the two species. The relationship between Catch-per-set and Boat length (Fig. 4) derived from GAM seems rather complex, but within the most common boat size range, 20–26 m, the relationship is very similar to the one observed for the anchovy, i.e. a slight positive trend in Catch-per-set with increasing boat size. In the case of sardine, there is no flattening (and decline) of the curve beyond 26 m, as was obtained for anchovy, because, as discussed, these bigger boats specialise in fishing for sardine. This may be the reason why ‘boat size’ has a greater effect on the variance explained by the sardine GLM (2.8% compared with 1.3% for anchovy). The negative trend obtained for the boat range 11–19 m is probably an artefact. Of the 32 boats in this category, all but two, are bait boats which fish exclusively for sardine but on such a small scale, in terms of their TACs and tonnage of individual catches, that boat size cannot be expected to exert much influence on Catch-per-set. The two lunar predictors, phase and elevation, have no appreciable effect in the sardine GLM (⬍0.05% of variance) and the GAM derived relationships (Fig. 4) are completely flat. Thomas and Schu¨ lein (1988) reported that catches of Namibian sardine were somewhat larger during full moon than during new moon, but that the effect was less pronounced than in the case of the other two species investigated by them. If one compares the sardine results with those of anchovy, one gains the impression that sardines are less light-sensitive than anchovies. SST parameters seem to have more or less the inverse effect on sardine catches as on anchovy. Sardine Catch-per-set increases almost linearly with SST; SSTstd has little effect apart from a weak negative trend towards large SSTstd and there is a clear, almost linear, negative trend with respect to SSTdif (Fig. 4). Their contributions to the variance explained by the GLM model are however very similar to the anchovy model (Table 2), that is equally small. GAM results nevertheless provide some insight into the conditions preferred by sardine and suggest that sardine tend to concentrate in warmer water than anchovy; there is little or no tendency to aggregate near fronts and they seem to prefer areas where recent cooling had taken place. Because, on the time scale of the interval between successive images, cooling is most probably a consequence of upwelling and sardines, which is the most herbivorous of the three species, may be attracted to newly upwelled water as a consequence of the associated increase in phytoplankton abundance. The time interval on which SSTdif is based is, however, too short for a significant zooplankton growth and thus would not offer better feeding conditions to the more zooplanktivorous anchovies and round herring. 4.3. Round herring As with the other two species, temporal parameters are the dominant predictors in the round herring GLM model (Table 2) due to quite clear and strong relationships between Catch-per-set and the Year, Month and Hour parameters demonstrated by the GAM results (Fig. 5). The round herring GLM results are, however, different in the sense that the Month effect is notably greater than the Year effect (14.0% and 6.1%, respectively). It is also seen that the GAM derived relationship of Catch-per-set versus Month closely reflects the mean seasonal trend in catches (Fig. 2), indicating that both Catch-per-set and the mean quantity of round herring caught drops off very sharply in the second half of the year. This explains the

J.J. Agenbag et al. / Progress in Oceanography 59 (2003) 275–300

295

pronounced effect of the Month predictor on the GLM. The relationship between Catch-per-set and Hour (Fig. 5) is the exact opposite of anchovy and indicates that Catch-per-set for round herring is a maximum during the night. Unlike anchovy, which remains in the surface layers throughout the day, round herring performs a pronounced diel vertical migration (Armstrong et al., 1991). The fish are mainly near the sea bed during the day, migrate in dense schools to and from the surface at dusk and dawn and tends to disperse on the surface during the night. The amplitude of variation of the round herring loess curve is larger than that of anchovy, suggesting that Catch-per-set in more dependent on Hour and this is confirmed by the larger part of the variance explained by the GLM model (5.1% compared with 1.8% in the case of anchovy). The round herring Year:Month interaction is the second most important predictor (10.4%) and probably a consequence of the strong Month effect because, as in the case of sardine, round herring catches are not subject to short term variations arising from recruitment fluctuations. In fact, inspection of the interannual variation of the Catch-per-set seasonal trends indicates that the round herring trend is the most stable of the three species. Spatial predictors Latitude and Depth indicate that the largest round herring catches are made in the northern part of the fishing grounds (~31°S), with a secondary maximum around 34°S and that catch size rapidly increases with depth to a maximum at about 150 m (Fig. 5). The Depth relationship is in agreement with the distribution obtained from research cruises, which have demonstrated that most round herring are located between the 100 and 200 m depth contours (Roel & Armstrong, 1991). The Latitude and Depth effects, respectively accounts for 8.7% and 3.7% of the variance explained by the GLM and their relationships with Catch-per-set are fairly well reflected by the distribution of mean Catch-per-set shown in Fig. 9. The interactions of these predictors with Year and Month are comparatively small and generally similar to those of the other two species. The Depth:Year interaction is notably larger in the round herring model (1.2%) compared with the 0.4% and 0.5% in the anchovy and sardine models, respectively, probably as a result of the round herring Catch-per-set being much more depth dependant than the other two species. The relationship between round herring Catch-per-set and Boat length derived from GAM is similar to that of anchovy (compare Figs. 3 and 5). This is to be expected because both species are used for the production of meal and are consequently targeted mainly by the 20–26 m non-refrigerated boats. Adult round herring, however, school with adult sardine, which is probably the reason why the GAM derived relationship for round herring does not have an appreciable negative slope within the range of 26–37 m boats which specialise in catching sardine, whereas the anchovy curve demonstrates a fairly prominent downward trend in catch size over this range. Within the middle 20–26 m range of boat lengths, which made most of the round herring catches, the slope of the GAM curve is much smaller than that of anchovy, indicating that Boat length has less influence on round herring catch size than on anchovy. This is also seen in the smaller contribution to the variance explained by the round herring GLM, 0.3% compared with 1.3% for anchovy and 2.8% for sardine (Table 2). As was the case with both anchovy and sardine, lunar predictors, Moon phase and Moon elevation have only small effects on the round herring GLM, explaining a mere 1.0% and 0.1% of the variance, respectively (Table 2). The GAM derived relationships (Fig. 5) show that Catch-per-set decreases slightly with increasing phase (⬎60% of the disk) and increasing positive elevations. In other words, catch size decreases with increasing moon illumination. Referring to the discussion of the effects of these variables on anchovy catches, we note again the apparent parallelism between the effects of moon light and variable Hour: round herring catches are smaller during the day (high solar light levels) and also when lunar light levels are high. We have exactly observed the inverse in the case of anchovy, while in the case of sardine light level, whether solar or lunar, seems to have no effect on shoaling behaviour. The GAM relationship for SST in Fig. 5, indicates a negative trend for round herring Catch-per-set with increasing SST but as can be seen from the 95% confidence limits, most catches were made in a narrow temperature range, about 15–18 °C, and in within this range the curve is flat, resulting in SST having no effect in the GLM model (Table 2).While one would not expect surface temperature variations to have a

296

J.J. Agenbag et al. / Progress in Oceanography 59 (2003) 275–300

major effect on the distribution of a species such as round herring which lives near the bottom and only migrates to the surface at night, the result is still surprising considering the relatively strong positive relationship with Depth. Within the Benguela upwelling system, SST normally warms with distance from the coast, i.e. with bathymetry, so that from this point of view one would have expected some positive relationship between Catch-per-set and SST. The absence of the expected relationship suggest, on the other hand, that the GAM/GLM methodology is able to separate the effects of the parameters in spite of covariance. The two other SST parameters, SSTstd and SSTdif make small contributions to the variance explained by the model (0.2% and 0.1%, respectively). The GAM curves (Fig. 5) show that the majority of round herring catches were made within a narrow range of SSTstd values (0.0–0.8 °C) but within this range there is quite a strong tendency for catch size to increase with increasing SSTstd. Since SSTstd is seen as a measure of horizontal temperature gradients, the result is in agreement with the common perception that round herring tends to concentrate in the vicinity of thermal fronts. Such a tendency has not been scientifically demonstrated before but may be justified in terms of the fact that round herring feeds exclusively on zooplankton and the generally accepted hypothesis that frontal zones support increased zooplankton populations (Bakun, 1996). SSTdif has almost no effect in the GLM model because, in the data range, ⫺2 to 2 °C, where the majority of catches were made, the curve of round herring Catch-per-set versus SSTdif is effectively flat (Fig. 5). The relationship however suggests that this species tend to prefer a temporally more stable SST environment. 4.4. Summary of predictor effects Temporal variables had the greatest effect on the GLM for all species. In the case of anchovy and sardine, for which we have annual biomass estimates, trends in the GAM Catch-per-set versus Year relationships are similar to the trends in biomass which confirms the applicability of mean Catch-per-set as index of annual (global) biomass and, by extrapolation, supports the concept of using it also as an indication of density. The seasonal effect varies considerably from species to species. It is greatest in the case of round herring, but it is likely that a major part of the Month variance for this species arises from small catches of adult round herring during the second half of the year (Roel & Armstrong, 1991) causing the sharp drop in Catch-per-set in May/June. Although, during our data period, all three species were caught most frequently during the night, anchovy Catch-per-set was actually greatest at midday and it is noted that anchovy catch size reaches a maximum at the time of the day when feeding activity is normally at a minimum, which suggests that fish tend to form bigger and/or more compact schools at that stage. Round herring catch size was largest at midnight, which is in agreement with the diel vertical migration behaviour of mature round herring. Sardine catch size is apparently not influenced by time of the day. Spatial parameters, Latitude and Depth, have the second greatest influence on the Catch-per-set (on average 25% of the explained variance) and have to be considered in combination. Spatial effects as perceived from the GAMs and GLMs are well reflected on the averaged Catch-per-set patterns represented in Fig. 9. A comparison of this figure with Fig. 1 shows that there is a low spatial match between the areas of maximum mean annual catches and the level of local abundance (density) suggested by Catchper-set. The mean annual catch distribution in Fig. 1 is closely related to the index catch-per-unit-area which we rejected as an adequate index of density because, amongst other reasons, it was regarded as too susceptible to biases introduced by non-uniform sampling by the fishing fleet. The fact that the higher mean catches (catch-per-unit-area) are generally closer to the coast (and thus to the fishing harbours marked on Fig. 1) than the peaks in the Catch-per-set distribution is indicative of such a bias and implies that fishing is often not conducted in the locations of greatest fish density. In the case of anchovy there also seems to be a temporal mismatch in the sense that more anchovy is caught at night, while Catch-per-set is actually greater during the day. These observations suggest that the fishers have difficulty predicting the

J.J. Agenbag et al. / Progress in Oceanography 59 (2003) 275–300

297

best spatio-temporal strata for their operations but may very well be consequences of a strategy aimed at minimisation of sailing time and better visual detection of fish schools at night. Contrary to what we expected, the effect of boat size on Catch-per-set turned out to be relatively small. The reason for this being that a very large proportion of the catches had been made by boats in the narrow size range, 20–26 m. Within this range, Catch-per-set generally increases slightly with boat size, as one would expect. The bait boats (⬍20 m) deliberately make small catches so that the relationship between Boat size and Catch-per-set breaks down. The large refrigerated boats (⬎26 m) tend to specialise in fishing for sardine, with the result that sardine is the only one of the three species that shows an almost continuous positive trend in Catch-per-set for Boat length ranging from 20 to 37 m. Relationships between Catch-per-set and the lunar parameters, Moon phase and Moon elevation are weak but suggest that anchovy and round herring schooling behaviour are affected by the level of lunar illumination. Anchovy Catch-per-set increases with Moon phase as well as Moon elevation, in other words, they form bigger and/or denser schools at increasing levels of illumination. In the case of round herring the effect is reversed, in the sense that catch size (school size) decreases with both lunar and solar illumination. Sardine catches are not affected by either. While available evidence thus suggest that light level (solar and lunar) directly affects fish behaviour, it must be borne in mind that it may also influence fishing efficiency in the sense that it is easier to detect fish schools on dark nights through the bioluminescence of plankton induced by the fish (Fre´ on & Misund, 1999). A anthropogenic response such as this could easily be misinterpreted as a fish behaviour response and is a good example of the potential pitfalls encountered when employing commercial catch records to derive relationships between fish and the environment. In the GLM models, SST parameters (SST, SSTstd and SSTdif) generally explain only a small part of the variance in Catch-per-set of all three species, but the GAM relationships nevertheless provide insight into conditions preferred by the different species and are generally consistent with existing knowledge. Judging by the effect on anchovy catch size, this species prefers cooler temperatures, but tends to avoid newly upwelled water. By contrast, sardines show a preference for warmer temperatures than anchovy and tend to concentrate in regions which have experienced cooling over the preceding week. A preference for this combination of conditions may be related to the sardine being the most herbivorous of the three species and, on the time scale of a week, phytoplankton abundance is expected to increase significantly in waters cooled by upwelling events but this interval is not sufficient to obtain a significant increase in zooplankton. Round herring catch size decreases with temperature but the species was most commonly caught in a narrow temperature range, 15–18 °C, typical of the frontal region between coastal upwelled water and warmer oceanic water. This is consistent with the round herring SSTstd relationship which suggests that this fish tends to concentrate at thermal fronts and also with the feeding preferences of this zooplanktivorous species.

5. Conclusion and perspectives The combination of GAM and GLM modelling seems to have been an effective approach to assess some environmental effects on fish density. The GAM procedure, which was principally employed as a means of formulating the functional shapes for the GLM model, turned out to be a very useful tool for visualising the relationships between the various predictors and Catch-per-set. The redundancy amongst the predictors seems to be negligible, as indicated by the GAM functional shapes and by the negligible effect of the GLM stepwise procedure. It was also very pleasing to note that the GAM methodology seems able to separate the effects of some variables that clearly covary, for example, round herring Catch-per-set was found to be strongly affected by Depth, but not at all by SST which strongly covaries with Depth over large scales in the Benguela region. The GLM procedure not only provided predictive equations, but also more robust indications of the relative importance of predictors. Although the environmental variables,

298

J.J. Agenbag et al. / Progress in Oceanography 59 (2003) 275–300

contributed only a small proportion of the variance explained by the GLMs, they generally provide meaningful and interesting relationships. The lunar predictors, for example, are relatively unimportant in terms of explaining variations in Catch-per-set, but provide information on the effect of light on schooling behaviour, a subject poorly studied with respect to marine pelagic fish. Some of the relationships are in completeor partial agreement with results previously described in the literature and thus instils confidence in those results which are not verifiable or are contrary to expectations. While our modelling approach seems to have performed satisfactorily and that useful results were obtained, it is worth considering means of improvement. Within the constraints of available data, quantification of the fish density is a major problem. There are few options and we have adopted Catch-per-set as the most promising one, but it is, as discussed, influenced by factors such as fishing gear, managerial objectives and fishing strategy which are difficult to quantify. We suspect that much of the variance in Catch-per-set not explained by our models might be due to these anthropogenic factors and to the practice of determining species composition from a single sample taken from the vessel’s hold at offloading and assigning this composition to all catches made prior to offloading. While it seems very unlikely that one would be able to formulate an alternative density index which exclude these factors without introducing some form of spatial bias, it might be possible to reduce their influence by careful partitioning of the data set. The known divergence of schooling behaviour between juveniles and adults of the pelagic species in South African waters could be another important source of variation not taken into account by our current models. The data set of commercial catch records used for our analysis does not contain age-related information, but such data exist and it might be worth considering means of merging it with the catch data. Such a merge will, however, severely reduce the number of usable catch records, perhaps to the point where a modelling approach, such as we have used, will no longer be viable. We have observed good relationships between the average annual Catch-per-set and population estimates (Fig. 8), which suggests that model results might be improved by introduction of the annual population biomass. In particular, this is expected to explain a significant part of the Year effect. The set of environmental variables used in the analysis was constrained by the need for high spatiotemporal coverage to match the catch data. In practice, this requirement virtually restricted the choice to remotely sensed parameters, which at the time was limited to SST. There is some scope for improvement, for example, the introduction of an upwelling intensity index, and more importantly, estimates of phytoplankton abundance as now obtainable from ocean colour images, could significantly improve our models’ performance.

Acknowledgements This research was financially supported by the EU ENVIFISH programme (contract number IC18-CT980329) as well as the French/South African IDYLE programme. We further gratefully acknowledge our colleagues Jan van der Westhuizen for assistance and advice regarding the use and interpretation of the commercial catch records as well as practical aspects pertaining to the South African pelagic fishery, Philipe Cury for his valuable assistance with GAM modelling and Laurent Drapeau without whose advice on all manner of statistical problems we probably could not have managed at all.

References Anders, A. S. (1975). Pilchard and anchovy spawning along the Cape east coast. South African Shipping News and Fishing Industry Review, 30(9), 53–57.

J.J. Agenbag et al. / Progress in Oceanography 59 (2003) 275–300

299

Armstrong, M. J. (1986). Assessments of appropriate yield for the pilchard resource in ICSEAF Division 1.6. Collection of Scientific Papers. International Commission for the Southeast Atlantic Fisheries, 13, 25–44. Armstrong, M. J., Berruti, A., & Colclough, J. (1987). Pilchard distribution in South African waters, 1983–85. South African Journal of Marine Science, 5, 871–886. Armstrong, M. J., Chapman, P., Dudley, S. F. J., Hampton, I., & Malan, P. E. (1991). Occurrence and population structure of pilchard Sardinops ocellatus, round herring Etrumeus whiteheadi and anchovy Engraulis capensis off the east coast of southern Africa. South African Journal of Marine Science, 11, 227–250. Armstrong, M. J., James, A. J., & Valde´ s Szeinfeld, E. S. (1991). Estimates of annual consumption of food by anchovy and other pelagic fish species off South Africa during the period 1984–1988. South African Journal of Marine Science, 11, 227–250. Bakun, A. (1996). Patterns in the ocean. Ocean processes and marine population dynamics. La Paz, BCS Mexico: California Sea Grant College System, NOAA in cooperation with Centro de Investigaciones Biologicas del Noroeste, 323 pp. Borchers, D. L., Buckland, S. T., Priede, I. G., & Ahmadi, S. (1997). Improving the precision of the daily egg production method using generalized additive models. Canadian Journal of Fisheries and Aquatic Sciences, 54, 2727–2742. Butterworth, D. S., & Berg, M. O. (1993). The development of a management procedure for the South African anchovy resource. In S. J. Smith, J. J. Hunt, & D. Rivard (Eds.), Risk evaluation and biological reference points for fisheries management (pp. 83– 99). Canadian special publication of fisheries and aquatic sciences, 120. Castillo, J., Barbieri, M. A., & Gonzalez, A. (1996). Relationships between sea surface temperature, salinity, and pelagic fish distribution off northern Chile. ICES Journal of Marine Science, 53, 139–146. Cochrane, K. L., Butterworth, D. S., De Oliveira, J. A. A., & Roel, B. A. (1998). Management procedures in a fishery based on highly variable stocks and with conflicting objectives: Experiences in the South African pelagic fishery. Reviews in Fish Biology and Fisheries, 8, 177–214. Cole, J., & Villacastin, C. (2000). Sea surface temperature variability in the northern Benguela upwelling system, and implications for fisheries research. International Journal of Remote Sensing, 21, 1597–1617. Crawford, R. J. M. (1981). Distribution, availability and movements of round herring Etrumeus teres off South Africa, 1964–1976. Fisheries Bulletin of South Africa, 14, 141–149. Cruickshank, R. A., & Boyd, A. J. (1985). An environmental basin model for west coast pelagic fish distribution. (pp541–549). International Symposium on the most important Upwelling Areas off Western Africa (Cape Blanco and Benguela), Instituto De Investigaciones Pesqueras, Barcelona 1985, V1. De Oliveira, J. A. A., Butterworth, D. S., Roel, B. A., Cochrane, K. L., & Brown, J. P. (1998). The application of a management procedure to regulate the directed and bycatch fishery of South African sardine Sardinops sagax. In S. C. Pillar, C. L. Moloney, A. I. L. Payne, & F. A. Shillington (Eds.), Benguela dynamics: Impacts of variability on shelf-sea environments and their living resources. South African Journal of Marine Science, 19, 449–469. Di Natale, A., & Mangano, A. (1986). Moon phases influence on CPUE: A first analysis of swordfish driftnet catch data from the Italian fleet between 1990 and 1991. International Commission for the Conservation of Atlantic Tunas. Collective Volume of Scientific Papers, 44, 264–267. Fiedler, P. C. (1983). Satellite remote sensing of the habitat of spawning anchovy in the southern California Bight. CalCOFI Report, XXIV, 202–209. Fiedler, P. C., & Bernard, H. J. (1987). Tuna aggregation and feeding near fronts observed in satellite imagery. Continental Shelf Research, 7, 871–881. Fox, C. J., O’Brien, C. M., Dickey-Collas, M., & Nash, R. D. M. (2000). Patterns in the spawning of cod (Gadus morhua L.), sole (Solea solea L.) and plaice (Pleuronectes platessa L.) in the Irish Sea as determined by generalized additive modelling. Fisheries Oceanography, 9, 33–49. Fre´ on, P. (1986). Re´ ponses et adaptations des stocks de clupe´ ides d’Afrique de l’ouest a` la variabilite´ du milieu et de l’exploitation. The`se de Doctorat, Universite´ d’Aix-Marseille II (France). Fre´ on, P. (1991). Seasonal and interannual variations of mean catch per set in Senegalese sardine fisheries: Fish behaviour or fishing strategy? In T. Kawasaki, S. Tanaka, Y. Toba, & A. Tanigushi (Eds.), Long-term variability of pelagic fish Populations and their environment (pp. 135–145). Oxford: Pergamon Press. Fre´ on, P., & Misund, O. A. (1999). Dynamics of pelagic fish distribution and behaviour: Effects on fisheries and stock assessment. Cambridge: University Press. Gaudreau, N., & Boisclair, D. (2000). Influence of acoustic estimates of the abundance of fish performing daily horizontal migration in a small oligotrophic lake. Canadian Journal of Fisheries and Aquatic Sciences, 57, 581–590. Hampton, I. (1987). Acoustic study on the abundance and distribution of anchovy spawners and recruits in South African waters. In A. I. L. Payne, J. A. Gulland, & K. H. Brink (Eds.), The Benguela and comparable ecosystems. South African Journal of Marine Science, 5, 901–917. Hampton, I. (1992). The role acoustic surveys in the assessment of pelagic fish resources on the South African continental shelf. In A. I. L. Payne, K. H. Brink, & R. Hilborn (Eds.), The Benguela trophic functioning. South African Journal of Marine Science, 12, 1031–1050.

300

J.J. Agenbag et al. / Progress in Oceanography 59 (2003) 275–300

James, A. G. (1987). Feeding ecology diet and field-based studies on feeding selectivity of the Cape anchovy Engraulis capensis Gilchrist. In A. I. L. Payne, J. A. Gulland, & K. H. Brink (Eds.), The Benguela and comparable ecosystems. South African Journal of Marine Science, 5, 673–692. James, A. G. (1988). Are clupeoid microphagists herbivorous or omnivorous? A review of the diets of some commercially important clupeoids. South African Journal of Marine Science, 7, 161–177. Kerstan, S. (1993). In Marine science for a sustainable future. Eighth South African marine science symposium, 17–22 October 1993, Langebaan, South Africa. Relationships between pilchard distribution and environmental factors. Lasker, R., Pela´ ez, J., & Laurs, R. M. (1981). The use of satellite infrared imagery for describing ocean processes in relation to spawning of the northern Anchovy (Engraulis mordax). Remote Sensing of the Environment, 11, 439–453. Laurs, R. M., Fiedler, P. C., & Montgomery, D. R. (1984). Albacore tuna catch distributions relative to environmental features observed from satellites. Deep-Sea Research, 31, 1085–1099. Louw, G. G., Van der Lingen, C. D., & Gibbons, M. J. (1998). Differential feeding by sardine Sardinops sagax and anchovy Engraulis capensis recruits in mixed shoals. In S. C. Pillar, C. L. Moloney, A. I. L. Payne, & F. A. Shillington (Eds.), Benguela dynamics: Impacts of variability on shelf-sea environments and their living resources. South African Journal of Marine Science, 19, 227–232. Luecke, C., & Wurtsbaug, W. A. (1993). Effects of moonlight and daylight on hydroacoustic estimates of pelagic fish abundance. Transactions of the American Fisheries Society, 122, 112–120. Maravelias, C. D., & Reid, D. G. (1995). Relationship between herring (Clupea harengus, L.) distribution and sea surface temperature in the northern North Sea. In C. Bas, J. J. Castro, & J. M. Lorenzo (Eds.), International symposium on middle-sized pelagic fish. Scientia Marina, 59, 427–438. Mathsoft (2001). S-PLUS 4 guide to Statistics. Seattle: Data Analysis Products Division. Pati, S. (1981). Observations on the lunar and tidal influence on gill netting in the Bay of Bengal. Fishery Technology, 18, 25–27. Podesta´ , G. P., Browder, J. A., & Hoey, J. J. (1993). Exploring the association between swordfish catch rates and thermal fronts on US longline grounds in the western North Atlantic. Continental Shelf Research, 13, 253–277. Richardson, A. J., Mitchell-Innes, B. A., Fowler, J. L., Bloomer, S. F., Field, J. G., Verheye, H. M., Hutchings, L., & Painting, S. J. (1998). The effect of sea temperature and food availability on the spawning success of the Cape anchovy Engraulis capensis in the southern Benguela. In S. C. Pillar, C. L. Moloney, A. I. L. Payne, & F. A. Shillington (Eds.), Benguela dynamics. Impacts of variability on shelf-sea environments and their living resources. South African Journal of Marine Science, 19, 275–290. Roel, B. A., & Armstrong, M. J. (1991). The round herring Etrumeus whiteheadi, an abundant, underexploited clupeoid species off the coast of southern Africa. South African Journal of Marine Science, 11, 267–287. Roel, B. A., Hewitson, J., Kerstan, S., & Hampton, I. (1994). The role of the Agulhas Bank in the life cycle of pelagic fish. South African Journal of Science, 90, 185–196. Shannon, L. V., Crawford, R. J. M., Brundrit, G. B., & Underhill, L. G. (1988). Responses of fish populations in the Benguela ecosystem to environmental change. Journal du Conseil pour l’Exploration de la Mer, 45, 5–12. Stretta, J. M. (1977). Tempe´ rature de surface et peˆ che thonie`re dans la zone frontale du cap Lopez (Atlantique tropical oriental) en juin et juillet 1972, 1974 et 1975. Cahiers ORSTOM, se´ rie oce´ anographie, 15, 163–180. Swartzman, G., Huang, C., & Kaluzny, S. (1992). Spatial analysis of Bering Sea groundfish survey data using generalized additive models. Canadian Journal of Fisheries and Aquatic Sciences, 49, 1366–1378. Thomas, R. M. (1986). Growth of larval pelagic fish in the South-East Atlantic from daily otolith rings in 1982/83 and 1983/84. South African Journal of Marine Science, 4, 61–78. Thomas, R. M., & Schu¨ lein, F. H. (1988). The shoaling behaviour of pelagic fish and the distribution of seals and gannets off Namibia as deduced from routine fishing reports. South African Journal of Marine Science, 7, 179–192. Van der Lingen, C. D. (1994). Effect of particle size and concentration on the feeding behaviour of adult pilchard Sardinops sagax. Marine Ecology Progress Series, 109, 1–13. Van der Lingen, C. D., Hutchings, L., Merkle, D., Van der Westhuizen, J. J., & Nelson, J. (2001). Comparative spawning habitats of anchovy (Engraulis capensis) and sardine (Sardinops sagax) in the southern Benguela upwelling ecosystem. In G. H. Kruse, N. Bez, T. Booth, M. Dorn, S. Hills, R. N. Lipcius, D. Pelletier, C. Roy, S. J. Smith, & D. Witherell (Eds.), Spatial processes and management of marine populations (pp. 185–209). University of Alaska Sea Grant, AK-SG-01-02, Fairbanks. Wada, T., & Matsumiya, T. (1990). Abundance index in purse seine fishery with searching time. Nippon Suisan Gakkaishi, 56, 725–728. Wallace-Fincham, B. P. (1987). The food and feeding of Etrumeus whiteheadi Wongratana 1983, off the Cape Province of South Africa. M.Sc. thesis, 117 pp. Williams, K. F. (1977). Sea surface temperature maps to assist tuna fisheries off New South Wales, Australia. G. H. Tomczak (Ed.), FAO Fisheries Technical Paper No. 170, pp. 38–55.