Targeting in a longline tuna fishery: a step-by-step guide

This guide uses a single iteration of the tuna operating model to compare the three practical approaches to fisher targeting in CPUE standardisation (Hoyle et al. 2014):

  • Cluster: discrete fishing tactics from clustering the catch composition;
  • DPC: continuous targeting from the principal components of the catch composition (Winker et al. 2013, 2014);
  • HBF: the number of hooks between floats as a recorded gear covariate.

We use two scenarios:

  1. Fixed: species depth varies in space with the thermocline, but the pattern is the same every year. This is the main case.
  2. Moving: the thermocline also moves between years (ENSO-like), so the depth of the fish, and with it the catchability of a given HBF, changes from year to year.

Everything else, including the seed, is identical in the two scenarios.

library(cpueTargeting)
library(ggplot2)

# Optional extra steps (spatial resolution, year interactions) take longer;
# set to TRUE to run them.
run_extras <- FALSE
# Replicates over many seeds (e.g. overnight); set to TRUE to run them.
run_sims <- FALSE

The operating model in brief

The operating model extends the spatial multispecies simulator of Thorson et al. (2017, Appendix A), itself a generalisation of the simulator of Ono et al. (2015), which is implemented for the bathymetric Run 22 scenario in simulate_bathymetric(). The tuna extension adds a pelagic horizontal habitat, species vertical distributions, longline hook depth by HBF and vertical targeting. A full description of both models is in the background document.

In brief:

  • Four species in a conceptual Southern Hemisphere ocean (north = warm): yellowfin (YFT) and bigeye (BET) prefer warm water, southern bluefin (BFT) and albacore (ALB) cooler water.
  • Each species has a normal depth profile. Within each pair one species is shallow (YFT, ALB) and one deep (BET, BFT). A thermocline anomaly moves all profiles up (shallow thermocline) or down (deep thermocline).
  • Each HBF level has a typical hook depth (roughly as in Bigelow et al. 2002, Fig. 2). Each set deviates from it by a lognormal error (line angle, currents, …), so HBF is an imperfect proxy for where the hooks are.
  • The catchability of a set for a species is the vessel’s catchability times the overlap between hook and fish depth distributions.
  • Prices favour ALB in years 1-10 (shallow, cool water), BET in years 11-20 (deep, warm water) and both BET and BFT in years 21-30 (deep water, warm north and cool south), loosely following the historical shift from albacore to bigeye targeting. The first switch changes both location and depth; in the last decade the fleet splits between two deep targets in different areas. Skippers choose where to fish (spatial targeting) and which HBF to use (vertical targeting) from the expected catch value.
  • Effort increases towards 800 sets per year and is cut by 25% in years 21-30 (~17,000 sets in total).
  • True trends differ between species. B_start sets where each stock starts (YFT 0.6, BET 0.7, BFT 0.2, ALB 0.9 of carrying capacity) and q how hard it is fished. ALB is fished down to about half while targeted and then levels off; BET declines steadily to about 30%; BFT recovers from 20% to almost 50% and declines again once it is targeted; YFT dips and then recovers once effort is cut. Use effort, q, B_start and r to change these trends.

Where the hooks fish and where the fish are. Each panel is one HBF level: the grey bars show the expected depth distribution of the hooks, the coloured areas the depth distribution of each species. Where a species’ area overlaps the grey bars, a set with that HBF can catch it; the numbers give the resulting vertical availability, which multiplies catchability in the operating model. Shallow gear (HBF 5) catches mostly YFT and ALB; deep gear (HBF 20) catches relatively more BET and BFT.

plot_tuna_vertical_distributions()

How to control the scenario

All settings live in one list, tuna_settings(). Print it to see the defaults, change elements with modifyList() (or $), and pass the list to simulate_tuna(). Settings you leave out keep their default values.

s <- tuna_settings()
str(s)                                        # all settings and defaults
s2 <- modifyList(s, list(effort = 500, set_cv = 0.4))
sim <- simulate_tuna(s2, seed = 123)

The settings come in five groups. Species-specific settings are vectors in the order of species: YFT, BET, BFT, ALB.

Targeting through prices

Targeting changes are driven by species prices: each year skippers send effort to the cells, and choose the HBF, with the highest expected catch value (price x expected catch). price is a matrix with one row per year and one column per species. The default has three 10-year blocks:

Years Most valuable Where it lives Fleet goes to
1-10 ALB shallow, cool south shallow gear, south
11-20 BET deep, warm north deep gear, north
21-30 BET and BFT deep, warm north and cool south deep gear, north and south

This loosely follows the historical shift from albacore (canning) to bigeye (sashimi) targeting in the Japanese longline fleet (Hoyle and Okamoto 2013), followed by increasing interest in southern bluefin as it recovers. The first switch changes both where and how deep the fleet fishes. In the last decade both targets are deep, so the fleet splits by area rather than by depth, which helps to separate spatial from vertical targeting.

tuna_settings()$price[c(1, 11, 21), ]   # one row from each block
##      [,1] [,2] [,3] [,4]
## [1,]  1.0  0.5  0.5  2.0
## [2,]  1.0  2.0  1.0  0.5
## [3,]  0.5  2.0  2.0  0.5

To define your own sequence, build a new matrix, e.g. a single switch from ALB to BET:

s <- tuna_settings()
s$price <- matrix(c(rep(c(1.0, 0.5, 0.5, 2.0), 15),   # years 1-15: ALB
                    rep(c(1.0, 2.0, 1.0, 0.5), 15)),   # years 16-30: BET
                  ncol = 4, byrow = TRUE)              # columns: YFT, BET, BFT, ALB

Constant prices (the same row every year) switch off targeting changes, which is a useful control scenario.

Presets. The default has strong targeting: the target species is worth four times as much as the others and skippers respond almost deterministically, so nominal CPUE is biased by about 15-20%, enough to separate the standardisation methods in a single run. Two alternatives:

  • tuna_settings("mild"): the same stocks and fishery but moderate targeting (prices differ by a factor of two, skippers respond less strictly). Nominal CPUE is only slightly biased and the methods are hard to tell apart without replicates: when targeting is weak, the choice of method matters less than noise.
  • tuna_settings("stress_test"): strong targeting of lightly fished, unfished stocks (q = 0.2, B_start = 1), the settings of the first runs. True trends are nearly flat, so nominal CPUE changes mainly through targeting.
base <- tuna_settings("mild")   # instead of tuna_settings() in Step 1

How strongly skippers respond:

Setting Effect Default
preference how strongly effort concentrates on high-value cells (spatial targeting); 0 = effort spread evenly 1.5 (mild: 1.25)
HBF_choice_temp how strictly skippers pick the best HBF (vertical targeting); small = strict, large = nearly random 0.15 (mild: 0.2)

Species habitat

Setting Meaning Default (YFT, BET, BFT, ALB)
sst_opt, sst_tol preferred SST and tolerance (deg C): horizontal distribution 25, 24, 20, 21; 3, 3.5, 3, 3.5
depth_mu, depth_sd mean and SD of the depth profile (m): vertical distribution 90, 260, 240, 110; 60, 50, 55, 60
B0, r carrying capacity (t) and intrinsic growth rate 400, 300, 150, 350 thousand t; 0.30, 0.18, 0.25, 0.25
B_start initial biomass as a fraction of carrying capacity 0.6, 0.7, 0.2, 0.9
dispersal mean movement distance per year (cells) 1.5

Closer sst_opt values or larger sst_tol mix the species more in space; closer depth_mu or larger depth_sd mix them more in depth, which weakens the contrast between shallow and deep gear. Check the result with plot_tuna_vertical_distributions(s) and, after simulating, plot_availability(sim).

Longline gear

Setting Meaning Default
HBF_levels HBF values the fleet can use 5, 10, 15, 20
HBF_depth median hook depth for each HBF level (m) 110, 160, 205, 250
set_cv set-level error in hook depth: how well HBF predicts where the hooks are (0 = perfectly) 0.2
hook_cv spread of hooks within a set 0.35

set_cv is the main “error in variable” knob for the HBF covariate.

Environment

Setting Meaning Default
sst_mean, sst_gradient, sst_noise SST field: mean, south-north increase, random variation (deg C) 23, 10, 1
thermocline_sd fixed spatial pattern of the thermocline anomaly (m) 30
thermocline_annual_sd basin-wide annual shift (m); ~10-15 is realistic 0
thermocline_annual_field_sd new spatial pattern each year (m); ~5-10 0

Fleet and data

Setting Meaning Default
n_years number of years (the price matrix needs at least as many rows) 30
effort sets per year: one value per year, or a single value for a logistic increase towards it logistic increase towards 800, cut by 25% in years 21-30
n_vessels, q, vessel_cv number of vessels, fishing catchability (how hard stocks are fished), between-vessel CV 30, 3, 0.2
q_index catchability at which CPUE is recorded: sets the zeros and noise of the data independently of q (set it equal to q to couple them) 0.2
tweedie_power, tweedie_phi set-level noise in catchability 1.2, 0.1
obs_res positions in Data are reported as centres of obs_res x obs_res blocks 5

Step 1: define the scenarios

The thermocline anomaly in cell i and year t is fixed[i] + shift[t] + annual[i, t]:

Setting Meaning Fixed Moving
thermocline_sd SD of the fixed spatial pattern (m) 30 30
thermocline_annual_sd SD of the basin-wide annual shift (m, AR(1)) 0 15
thermocline_annual_field_sd SD of a new spatial pattern each year (m) 0 10

The annual values give typical year-to-year anomalies of about 20-30 m and occasionally 50 m, a realistic range for interannual changes in thermocline depth over most of the tropical Pacific.

base <- tuna_settings()      # effort: up to ~700 sets per year, cut by 25% after year 20

scenarios <- list(
  Fixed  = modifyList(base, list(thermocline_sd = 30)),
  Moving = modifyList(base, list(thermocline_sd = 30,
                                 thermocline_annual_sd = 15,
                                 thermocline_annual_field_sd = 10))
)

Step 2: simulate

With the same seed, both runs share the SST field, the fixed thermocline pattern and the vessels.

seed <- 123
sims <- lapply(scenarios, simulate_tuna, seed = seed)

Step 3: the environment

SST (warm north, cool south) drives the horizontal distribution; the fixed thermocline pattern shifts the species’ depth:

sims$Fixed$Plots$Env

sims$Fixed$Plots$Thermocline

The resulting species densities:

sims$Fixed$Plots$SpDens

The same distributions in three dimensions. The latitude x depth section averages over x: YFT and BET sit in the warm north, BFT and ALB in the cool south, and within each pair one species is shallow and one deep. The layer maps show the density in each depth layer, i.e. what a set fishing at that depth would encounter.

sec <- plot_sections(sims$Fixed)
sec$section

sec$layers

In the Moving scenario the basin-wide thermocline goes up and down between years, and the spatial pattern changes on top of it:

plot_thermocline(sims$Moving)$series

plot_thermocline(sims$Moving)$maps

Step 4: why the thermocline matters for HBF

When the thermocline is shallow (orange), the deep species move up towards the hooks of shallow sets; when it is deep (blue), they move below the reach of all but the deepest sets. The same HBF therefore catches a different mix of species depending on the thermocline.

plot_thermocline_effect(shifts = c(-30, 0, 30))

Step 5: how the fishery responds

Gear use

The fleet sets deeper gear once BET is valuable (from year 11):

sims$Fixed$Plots$HBF + ggtitle("Fixed: proportion of sets by HBF")

Vertical availability by HBF

The average overlap between hooks and fish for sets with each HBF. In the Fixed scenario the lines are fairly flat, so an HBF factor is a reasonable proxy for vertical availability. In the Moving scenario the catchability of a given HBF changes from year to year with the thermocline.

for (sc in names(sims)) print(plot_availability(sims[[sc]]) + ggtitle(sc))

Nominal CPUE versus true biomass

for (sc in names(sims)) {
  print(sims[[sc]]$Plots$CPUE + ggtitle(paste(sc, "- nominal (points) vs true (blue)")))
}

Step 6: standardise

Each model is year plus a targeting term; +S adds a spatial smoother s(Xs, Ys) on the reported 5 x 5 blocks.

Model Targeting term Treats targeting as Uses
Base none ignored catch per set
Base+S none, + space ignored catch per set + position
Cluster+S k-means clusters of catch composition (4 = PCs + 1) discrete all species’ catches
DPC+S s(PC1) + s(PC2) + s(PC3) of catch composition continuous all species’ catches
HBF+S HBF factor gear covariate HBF from logbooks

One addition: DPC+S (2 PCs) uses only the first two principal components, as in the 2023 CPUE guidelines. The default (n_pc = 3) uses three, which with four species is all the information in the catch composition.

Positions are reported only as 5 x 5 blocks. This is deliberate: logbooks rarely report positions at the resolution at which fish and fleets interact, and a longline set may span a front. The spatial term therefore cannot resolve all spatial targeting, which is the realistic, data-limited case.

engine = "bam" uses mgcv::bam(discrete = TRUE) for speed; use engine = "gam" to reproduce the guideline approach exactly.

models <- c("Base", "Base+S", "Cluster+S", "DPC+S", "HBF+S")

fits_main <- lapply(sims, standardize_cpue, models = models,
                    spatial_res = 5, n_pc = 3, engine = "bam")

fits_dpc2 <- lapply(sims, standardize_cpue, models = "DPC+S",
                    spatial_res = 5, n_pc = 2, engine = "bam", label = " (2 PCs)")

fits <- Map(bind_comparisons, fits_main, fits_dpc2)
fits$Fixed$formulas
##                                                                      Base 
##                                                                    "year" 
##                                                                    Base+S 
##                                                 "year + s(Xs, Ys, k = 8)" 
##                                                                 Cluster+S 
##                                            "year + s(Xs, Ys, k = 8) + FT" 
##                                                                     DPC+S 
## "year + s(Xs, Ys, k = 8) + s(PC1, k = 3) + s(PC2, k = 3) + s(PC3, k = 3)" 
##                                                                     HBF+S 
##                                           "year + s(Xs, Ys, k = 8) + HBF" 
##                                                             DPC+S (2 PCs) 
##                 "year + s(Xs, Ys, k = 8) + s(PC1, k = 3) + s(PC2, k = 3)"

The combinations Cluster_HBF+S and DPC_HBF+S can be added to models in the same way.

Step 7: compare the indices with the truth

Indices

plot_indices(fits, models = c("Base", "Base+S", "Cluster+S",
                              "DPC+S (2 PCs)", "DPC+S", "HBF+S"))

Annual indices are noisy. To compare the trends, show only loess smoothers (type = "smooth"), or the annual values as points with the smoother (type = "points+smooth"); span sets the smoothness:

plot_indices(fits, models = c("Base", "Cluster+S", "DPC+S", "HBF+S"),
             type = "smooth", span = 0.5)

Accuracy

Root mean squared error of the log index relative to the true biomass trend:

plot_diagnostics(fits)

rmse <- do.call(rbind, lapply(names(fits), function(sc) {
  data.frame(scenario = sc, fits[[sc]]$diagnostics)
}))
rmse_wide <- reshape(rmse, idvar = c("scenario", "species"), timevar = "model",
                     direction = "wide")
names(rmse_wide) <- sub("rmse_log.", "", names(rmse_wide), fixed = TRUE)
knitr::kable(rmse_wide, digits = 3, row.names = FALSE)
scenario species Base Base+S Cluster+S DPC+S HBF+S DPC+S (2 PCs)
Fixed YFT 0.209 0.209 0.091 0.106 0.040 0.057
Fixed BET 0.256 0.254 0.132 0.130 0.081 0.082
Fixed BFT 0.168 0.147 0.056 0.309 0.056 0.051
Fixed ALB 0.254 0.214 0.116 0.078 0.074 0.096
Moving YFT 0.199 0.199 0.097 0.110 0.146 0.065
Moving BET 0.249 0.248 0.129 0.127 0.177 0.087
Moving BFT 0.192 0.152 0.076 0.307 0.150 0.063
Moving ALB 0.259 0.220 0.124 0.086 0.127 0.107

What to look for

  • Spatial targeting. Base (no space) is biased for species whose targeting is mainly spatial, e.g. BFT when the fleet moves into cool water. Base+S removes most of that bias.
  • Vertical targeting. The shallow species (YFT, ALB) drop in nominal CPUE at year 11, and BET rises, when the fleet switches to deeper gear. Space alone cannot remove this; DPC and HBF can.
  • Cluster versus DPC. Targeting in this simulator is continuous (a graded choice among locations and HBF levels), so a few discrete clusters (by default the number of PCs plus one, i.e. 4) capture only part of it. Clusters are recommended only when discrete strategies exist (Hoyle et al. 2014).
  • Moving thermocline. An HBF factor assumes the same catchability for an HBF level in every year, so it cannot follow the annual changes, and may do worse than Base+S because skippers choose HBF partly in response to the thermocline. DPC uses the realised catch mix of each set, which reflects where the hooks actually fished relative to the fish that year, so it can partly correct for it.
  • Number of PCs. With two PCs, DPC misses part of the targeting signal for some species (e.g. BFT); the third PC adds it. With four species, three PCs use all the information in the catch composition.
  • Limits of DPC. The catch composition only shows relative catchability. If a shift changes the availability of several species in the same proportion, the mix does not change and DPC cannot see it. And the more PCs are used, the more risk that real changes in relative abundance are absorbed by the “targeting” terms (Winker et al. 2014 select the number of PCs with a scree test for this reason).

A single iteration is only an illustration; repeat over seeds before drawing conclusions.

Step 8: what do the principal components capture?

The DPC models use the PC scores of each set as predictors. To see what they represent, plot_pcs() shows the species loadings, the scores in space, against the realised hook depth of the set (latent truth from the simulator, not used in any model) and over time.

pcs_fixed <- plot_pcs(fits$Fixed, pcs = c("PC1", "PC2", "PC3"))

What the PCs mean in terms of species:

pcs_fixed$loadings

Where each tactic is used. Compare with the SST field: warm-water (YFT, BET) versus cool-water (BFT, ALB) catches show up as a north-south pattern.

pcs_fixed$space

How the scores change with the depth at which the hooks actually fished, for each HBF. A PC that changes with hook depth within an HBF level carries information about the realised fishing depth that the HBF factor does not.

pcs_fixed$depth

How the average tactic changes over time, following the price blocks:

pcs_fixed$time

The same plots for the Moving scenario; in particular, the scores over time now also follow the thermocline:

pcs_moving <- plot_pcs(fits$Moving, pcs = c("PC1", "PC2", "PC3"))
pcs_moving$depth

pcs_moving$time

Replicates: is the ranking robust?

One iteration is only an illustration. run_replicates() repeats the simulation and standardisation for many seeds. Results are saved after every fit and finished seeds are skipped, so the run can be interrupted and resumed (e.g. overnight). Set run_sims <- TRUE to run it.

reps <- run_replicates(
  scenarios, seeds = 1:20,
  models = c("Base", "Base+S", "Cluster+S", "DPC+S", "HBF+S"),
  extra = list(list(models = "DPC+S", n_pc = 2, label = " (2 PCs)", engine = "bam")),
  spatial_res = 5, n_pc = 3, engine = "bam",
  file = "replicates_fixed_moving.rds")
plot_replicates(reps)

Afterwards, the results can be loaded and plotted without re-running:

reps <- readRDS("replicates_fixed_moving.rds")
plot_replicates(reps)

Optional: with or without space, and at what resolution

Set run_extras <- TRUE to run this step. Every model is fitted with and without the smoother at three reporting resolutions: 1 x 1 (exact cell), 5 x 5 and 10 x 10 cells.

targeting <- c("Base", "Cluster", "DPC", "HBF")
resolutions <- c("1x1" = 1, "5x5" = 5, "10x10" = 10)
fits_res <- lapply(resolutions, function(res) {
  standardize_cpue(sims$Fixed, models = c(targeting, paste0(targeting, "+S")),
                           spatial_res = res, spatial_k = if (res == 1) 30 else NULL,
                           engine = "bam")
})
plot_diagnostics(fits_res[["5x5"]])
plot_diagnostics(fits_res, models = paste0(targeting, "+S"))
  • Space matters most for species whose targeting is mainly spatial. It cannot replace the targeting term for gear depth within a pair (YFT vs BET, ALB vs BFT).
  • At 1 x 1 the smoother needs a larger basis (spatial_k) and may start to absorb part of the targeting signal.

Optional: interactions with year

When the effect of gear changes over time, one idea is to let the HBF effect vary with time. Two versions can be added through add_models:

  • HBF_period: a separate HBF effect in each 10-year price block. It cannot follow year-to-year thermocline changes.
  • HBF_year_re: a random HBF x year interaction, which gives each year its own (shrunk) HBF effects.
extra <- c(
  HBF_period  = "year + s(Xs, Ys, k = 8) + HBF_period",
  HBF_year_re = "year + s(Xs, Ys, k = 8) + HBF + s(year, HBF, bs = 're')"
)
fits_int <- lapply(sims, standardize_cpue,
                   models = c("HBF+S", "DPC+S"), add_models = extra, engine = "bam")
plot_diagnostics(fits_int)
plot_indices(fits_int, models = c("HBF+S", "DPC+S", "HBF_period", "HBF_year_re"))

Points to keep in mind:

  • With a year x HBF interaction the “year effect” is no longer unique: the index refers to the reference HBF (the most common level), and a different reference gives a different trend.
  • A more defensible alternative is to average the predictions over the HBF levels (or over the fished cells and gear configurations) in each year, giving a marginal index.
  • The interaction can absorb part of the abundance signal of species that are caught mainly with one gear configuration. Compare the year effects with the truth, not only the fit statistics.

A note on habitat-based standardisation

Habitat-based standardisation (HBS; Bigelow et al. 2002) estimates nothing about the gear. It weights each set’s effort by the predicted overlap between the hook depths of its HBF and the species’ depth distribution from a habitat model. It needs a validated habitat model and ocean data, so it is a specialised approach and not part of the comparison here. It can be fitted with models = "HBS+S"; in this simulator its habitat model knows the species’ depth preferences and the average thermocline, but not the annual anomalies.

References

Bigelow, K.A., Hampton, J. and Miyabe, N. (2002) Application of a habitat-based model to estimate effective longline fishing effort and relative abundance of Pacific bigeye tuna (Thunnus obesus). Fisheries Oceanography 11, 143-155.

Hoyle, S.D., Langley, A.D. and Campbell, R.A. (2014) Recommended approaches for standardizing CPUE data from pelagic fisheries. WCPFC-SC10-2014/SA-IP-10.

Ono, K., Punt, A.E. and Hilborn, R. (2015) Think outside the grids: an objective approach to define spatial strata for catch and effort analysis. Fisheries Research 170, 89-101.

Thorson, J.T., Fonner, R., Haltuch, M.A., Ono, K. and Winker, H. (2017) Accounting for spatiotemporal variation and fisher targeting when estimating abundance from multispecies fishery data. Canadian Journal of Fisheries and Aquatic Sciences 74, 1794-1807.

Hoyle, S.D. and Okamoto, H. (2013) Target changes in the tropical WCPO Japanese longline fishery, and their effects on species composition. WCPFC-SC9-2013/SA-IP-04.

Winker, H., Kerwath, S.E. and Attwood, C.G. (2013) Comparison of two approaches to standardize catch-per-unit-effort for targeting behaviour in a multispecies hand-line fishery. Fisheries Research 139, 118-131.

Winker, H., Kerwath, S.E. and Attwood, C.G. (2014) Proof of concept for a novel procedure to standardize multispecies catch and effort data. Fisheries Research 155, 149-159.