This tutorial illustrates the application of the TropFishR package to perform a single-species fish stock assessment with length frequency (LFQ) data. According to Sparre and Venema (1998), this includes following steps: (1) estimation of biological stock characteristics (growth and natural mortality), (2) exploration of fisheries aspects (fishing mortality rate and selectivity), (3) assessment of stock status. The order of the methods is important as they build upon each other in a sequential way. Data from literature might be used to skip a step in this workflow or to compare them to the outcomes of this routine.
The current version of TropFishR (v1.6.3) requires R \(>= 3.0.0\) and can be downloaded from CRAN as follows.
Alternatively, the package can be installed from GitHub:
The package is loaded into the R environment with:
The tutorial will make use of a synthetic LFQ data set included in
the package (“synLFQ7”). We load the data set into the R environment
with data(synLFQ7)
and rename it with:
lfq <- synLFQ7
.
Growth, natural mortality, recruitment patterns and the stock-recruitment relationship are important biological stock characteristics and input parameters for population dynamics and yield per recruit models.
Commonly used growth parameters are the asymptotic length (\(L_{\infty}\)), the growth coefficient
(\(K\)) and the theoretical length at
age zero (\(t_{0}\)) of the von
Bertalanffy growth function (VBGF). The ELEFAN (ELectronic LEngth
Frequency ANalysis) methods allow to estimate \(L_{\infty}\) and \(K\) from LFQ data by restructuring the data
and fitting growth curves through the restructured LFQ data (Pauly 1980). I recommend to start by
visualising the raw and restructured LFQ data, which aids in determining
an appropriate bin size and the moving average of the restructuring
procedure. The function lfqModify
allows to change the bin
size by setting the argument bin_size
to a numeric. The
function lfqRestructure
is used for the restructuring
process, where the argument MA
allows to control the number
of bins used for the moving average and the argument
addl.sqrt
allows to apply an additional squareroot
transformation in the restructuring process, which reduces the weighting
of large individuals.
## set seed value for reproducible results
set.seed(1)
## adjust bin size
lfq_bin2 <- lfqModify(lfq, bin_size = 2)
## plot raw and restructured LFQ data
ma <- 7
lfq_bin2_res <- lfqRestructure(lfq_bin2, MA = 7, addl.sqrt = FALSE)
opar <- par(mfrow = c(2,1), mar = c(2,5,2,3), oma = c(2,0,0,0))
plot(lfq_bin2_res, Fname = "catch", date.axis = "modern")
plot(lfq_bin2_res, Fname = "rcounts", date.axis = "modern")
par(opar)
Based on visual inspection of the length-frequency distributions, a bin size of 2 cm and a moving average of 7 seems appropriate for this data set and will be used in the following.
In TropFishR, there are 4 different methods based on the ELEFAN
functionality: (i) K-Scan for the estimation of K for a fixed value of
\(L_{inf}\), (ii) Response Surface
Analysis (RSA), (iii) ELEFAN with simulated annealing
(ELEFAN_SA
), and (iv) ELEFAN with a genetic algorithm
(ELEFAN_GA
), where the last three methods all allow to
estimate K and \(L_{\infty}\)
simultaneously. The K-Scan method does not allow to test if different
combinations of \(L_{\infty}\) and K
might result in a better fit. RSA with a range around \(L_{\infty}\) can be used to check different
combinations. However, RSA can be very slow and does not allow to
optimise over the parameters C and \(t_s\) of the seasonalised VBGF (soVBGF). It
only allows to compare the score of ELEFAN runs with manually fixed C
and \(t_s\) values. In contrast, the
newly implemented ELEFAN method ELEFAN_SA
using a simulated
annealing algorithm (Xiang et al. 2013)
and ELEFAN_GA
using genetic algorithms allow for the
optimisation of the soVBGF (M. H. Taylor and
Mildenberger 2017). The optimisation procedure in the simulated
annealing algorithm gradually reduces the stochasticity of the search
process as a function of the decreasing “temperature” value, which
describes the probability of accepting worse conditions.
Here, we will apply the two new ELEFAN functions
ELEFAN_SA
and ELEFAN_GA
. Both functions
require to define the lower (low_par
) and upper
(up_par
) bounds of the search space for all growth
parameters. For this data set, we assume a wide search space for \(K\) and use the maximum length in the data
to define the search space for \(L_{\infty}\) (C. C.
Taylor 1958; Beverton 1963). We assume the full parameter space
for the other parameters (\(t_{anchor}\), \(C\), \(t_s\)).
## coarse estimate of Linf
linf_guess <- max(lfq_bin2$midLengths) / 0.95
## lower search space bounds
low_par <- list(Linf = 0.8 * linf_guess,
K = 0.01,
t_anchor = 0,
C = 0,
ts = 0)
## upper search space bounds
up_par <- list(Linf = 1.2 * linf_guess,
K = 1,
t_anchor = 1,
C = 1,
ts = 1)
Then ELEFAN_SA
can be applied as follows.
## run ELEFAN with simulated annealing
res_SA <- ELEFAN_SA(lfq_bin2, SA_time = 60*0.5, SA_temp = 6e5,
MA = ma, seasonalised = TRUE, addl.sqrt = FALSE,
init_par = list(Linf = linf_guess,
K = 0.5,
t_anchor = 0.5,
C=0.5,
ts = 0.5),
low_par = low_par,
up_par = up_par)
## show results
res_SA$par
res_SA$Rn_max
Note that the computing time can be controlled with the argument
SA_time
and the results might change when increasing the
time, in case the stable optimum of the objective function was not yet
reached (stable optimum is indicated by overlapping blue and green dots
in the score graph). Due to the limitations of the vignette format the
computation time was set to 0.5 minutes, which results already in
acceptable results of \(L_{\infty}\) =
123.05, K = 0.2, \(t_{anchor}\) = 0.27,
C = 0.38, and \(t_s\) = 0.42 with a
score value (\(Rn_{max}\)) of 0.29. I
recommend to increase SA_time
to 3 - 5 minutes to increase
chances of finding the stable optimum. Another new optimisation routine
is based on generic algorithms and is applied by the function
ELEFAN_GA
.
## run ELEFAN with genetic algorithm
res_GA <- ELEFAN_GA(lfq_bin2, MA = ma, seasonalised = TRUE,
maxiter = 50, addl.sqrt = FALSE,
low_par = low_par,
up_par = up_par,
monitor = FALSE)
## show results
res_GA$par
res_GA$Rn_max
The generation number of the ELEFAN_GA
was set to only
50 generations (argument maxiter
), which returns following
results: \(L_{\infty}\) = 123.79, K =
0.22, \(t_{anchor}\) = 0.33, C = 0.49,
and \(t_s\) = 0.31 with a score value
(\(Rn_{max}\)) of 0.27. As with
ELEFAN_SA
the generation number was hold down due to the
vignette format and should be increased in order to find more stable
results. According to (Pauly 1980) it is
not possible to estimate \(t_{0}\)
(theoretical age at length zero) from LFQ data alone. However, this
parameter does not influence results of the methods of the traditional
stock assessment workflow (catch curve and yield per recruit model) and
can be set to zero. The ELEFAN methods in this package do not return
starting points as FiSAT II users might be used to. Instead, they return
the parameter t_anchor
, which describes the fraction of the
year where yearly repeating growth curves cross length equal to zero;
for example a value of 0.25 refers to April 1st of any year. The maximum
age is estimated within the ELEFAN function: it is the age when length
is 0.95 \(L_{\infty}\). However, this
value can also be fixed with the argument agemax
, when
alternative information about the maximum age of the fish species is
available.
The jack knife technique allows to estimate a confidence interval around the parameters of the soVBGF (Quenouille 1956; J. Tukey 1958; J. W. Tukey 1962). This can be automated in R with following code:
## list for results
JK <- vector("list", length(lfq_bin2$dates))
## loop
for(i in 1:length(lfq_bin2$dates)){
loop_data <- list(dates = lfq_bin2$dates[-i],
midLengths = lfq_bin2$midLengths,
catch = lfq_bin2$catch[,-i])
tmp <- ELEFAN_GA(loop_data, MA = ma, seasonalised = TRUE,
maxiter = 50, addl.sqrt = FALSE,
low_par = low_par,
up_par = up_par,
monitor = FALSE, plot = FALSE)
JK[[i]] <- unlist(c(tmp$par, list(Rn_max=tmp$Rn_max)))
}
## bind list into dataframe
JKres <- do.call(cbind, JK)
## mean
JKmeans <- apply(as.matrix(JKres), MARGIN = 1, FUN = mean)
## confidence intervals
JKconf <- apply(as.matrix(JKres), MARGIN = 1, FUN = function(x) quantile(x, probs=c(0.025,0.975)))
JKconf <- t(JKconf)
colnames(JKconf) <- c("lower","upper")
## show results
JKconf
Depending on the number of sampling times (columns in the catch
matrix) and the maxiter
, this loop can take some time as
ELEFAN runs several times, each time removing the catch vector of one of
the sampling times.
The fit of estimated growth parameters can also be explored visually and indicates high similarity with true growth curves and a good fit through the peaks of this data set.
## plot LFQ and growth curves
plot(lfq_bin2_res, Fname = "rcounts",date.axis = "modern", ylim=c(0,130))
lt <- lfqFitCurves(lfq_bin2, par = list(Linf=123, K=0.2, t_anchor=0.25, C=0.3, ts=0),
draw = TRUE, col = "grey", lty = 1, lwd=1.5)
lt <- lfqFitCurves(lfq_bin2, par = res_SA$par,
draw = TRUE, col = "darkblue", lty = 1, lwd=1.5)
lt <- lfqFitCurves(lfq_bin2, par = res_GA$par,
draw = TRUE, col = "darkgreen", lty = 1, lwd=1.5)
For further analysis, we use the outcomes of ELFAN with genetic algorithm by adding them to the Thumbprint Emperor data list.
The instantaneous natural mortality rate (M) is an influential parameter of stock assessment models and its estimation is challenging (Kenchington 2014; Powers 2014). When no controlled experiments or tagging data is available the main approach for its estimation is to use empirical formulas. Overall, there are at least 30 different empirical formulas for the estimation of this parameter (Kenchington 2014) relying on correlations with life history parameters and/or environmental information. We apply the most recent formula, which is based upon a meta-analysis of 201 fish species (Then et al. 2015) and assign the results to our data set. This method requires estimates of the VBGF growth parameters [\(L_{\infty}\) and K; Then et al. (2015)].
## estimation of M
Ms <- M_empirical(Linf = lfq_bin2$par$Linf, K_l = lfq_bin2$par$K, method = "Then_growth")
## assign M to data set
lfq_bin2$par$M <- as.numeric(Ms)
The result is a natural mortality of 0.28 year\(^{-1}\).
Another prerequisite for the stock status estimation is knowledge on
fishing mortality rate (F) (usually derived by subtracting natural
mortality from total mortality) and gear selectivity is necessary. The
length-converted catch curve allows the estimation of the instantaneous
total mortality rate (Z) of LFQ data and the derivation of a selection
ogive. Here we skip an in-depth selectivity exploration, because more
data would be required for this assessment. For a comprehensive
description of selectivity estimation refer to Millar and Holst (1997). The following approach
assumes a logistic selection ogive, typical for trawl-net selectivity,
which may provide an appropriate first estimate in the case of LFQ data
derived from a mixture of gears. The total mortality rate is estimated
with a sample of the catch representative for the whole year. Besides,
changing the bin size, the function lfqModify
allows
rearranging the catch matrix in the required format (catch vector per
year) and optinal pooling of the largest length classes with only a few
individuals into a plus group.
## define plus group as largest length class smaller than Linf
plus_group <- lfq_bin2$midLengths[max(which(lfq_bin2$midLengths < lfq_bin2$par$Linf))]
## summarise catch matrix into vector and add plus group
lfq_catch_vec <- lfqModify(lfq_bin2, vectorise_catch = TRUE, plus_group = plus_group)
The catch curve can then be applied to the vectorised data set with
catchCurve(lfq_catch_vec)
. This triggers an interactive
plotting function where the user selects points to be used in the
regression analysis by clickling on the first and last point to be
included. Please find for more information with
help(catchCurve)
. Alternatively, the argument
reg_int
defines the first and last point to be used in the
regression analysis and, thus, allows to avoid the interactive plotting
function. For this data set, the 18th and 55th point seem to be
reasonable first and last points of the regression analysis. The
argument calc_ogive
allows the estimation of the selection
ogive. Note, that you might require the argument
catch_columns
if your data set spans multiple years as it
allows to choose the columns (years after vectorisation) of the catch
matrix which will be summarised for the analysis. Here, we do not need
this argument as the catch matrix only includes catches from 2016. If
data of several years are available, the catch curve can either be
applied to the years separately (e.g. catch_columns=1
for
the first year), or to the data of several years combined
(e.g. catch_columns=c(1,2)
).
## run catch curve
res_cc <- catchCurve(lfq_catch_vec, reg_int = c(18,55), calc_ogive = TRUE)
## assign estimates to the data list
lfq_catch_vec$par$Z <- res_cc$Z
lfq_catch_vec$par$FM <- as.numeric(lfq_catch_vec$par$Z - lfq_catch_vec$par$M)
The catch curve analysis returns a Z value of 0.5 \(year^{-1}\). By subtracting M from Z, the fishing mortality rate is derived: 0.22 \(year^{-1}\). The selectivity function of the catch curve estimated a length at 50% probability of capture (\(L_{50}\)) of 35.81 cm.
The exploitation rate is defined as \(E = F/Z\) and relative to the reference level of 0.5, provides a simple indication of the stock status Gulland (1983).
For this data set, the exploitation rate is equal to 0.44 and, thus, does not indicate overfishing.
Prediction models (or yield per recruit models, e.g. Thompson and Bell model) allow to evaluate the status of a fish stock in relation to reference levels and to infer input control measures, such as restricting fishing effort or regulating gear types and mesh sizes. The model requires information about the parameters of the length-weight relationship (\(a\) and \(b\)) and the optional maturity parameters allow to estimate the Spawning Potential Ratio (SPR). By default the Thompson and Bell model assumes knife edge selection (\(L_{25}\) = \(L_{50}\) = \(L_{75}\))1. However, we can use \(L50\) and \(L75\) values, e.g. from the catch curve to define a selection ogive.
## assign length-weight parameters to the data list
lfq_catch_vec$par$a <- 0.015
lfq_catch_vec$par$b <- 3
## assign maturity parameters
lfq_catch_vec$par$Lmat <- 35
lfq_catch_vec$par$wmat <- 5
## list with selectivity parameters
selectivity_list <- list(selecType = "trawl_ogive",
L50 = res_cc$L50, L75 = res_cc$L75)
The parameter FM_change
determines the range of the
fishing mortality for which to estimate the yield and biomass
trajectories. In the second application of this model, the impact of
mesh size restrictions on yield is explored by changing \(L_{c}\) (Lc_change
) and F
(FM_change
, or exploitation rate, E_change
)
simultaneously. The resulting estimates are presented as an isopleth
graph showing yield per recruit. By setting the argument
stock_size_1
to 1, all results are per recruit. If the
number of recruits (recruitment to the fishery) are known, the exact
yield and biomass can be estimated. The arguments curr.E
and curr.Lc
allow to derive and visualise yield and biomass
(per recruit) values for current fishing patterns.
## Thompson and Bell model with changes in F
TB1 <- predict_mod(lfq_catch_vec, type = "ThompBell",
FM_change = seq(0,1.5,0.05),
stock_size_1 = 1,
curr.E = lfq_catch_vec$par$E,
s_list = selectivity_list,
plot = FALSE, hide.progressbar = TRUE)
#> [1] Fishing mortality per length class not povided, using selectivity information to derive fishing mortality per length class.
## Thompson and Bell model with changes in F and Lc
TB2 <- predict_mod(lfq_catch_vec, type = "ThompBell",
FM_change = seq(0,1.5,0.1),
Lc_change = seq(25,50,0.1),
stock_size_1 = 1,
curr.E = lfq_catch_vec$par$E,
curr.Lc = res_cc$L50,
s_list = selectivity_list,
plot = FALSE, hide.progressbar = TRUE)
## plot results
par(mfrow = c(2,1), mar = c(4,5,2,4.5), oma = c(1,0,0,0))
plot(TB1, mark = TRUE)
mtext("(a)", side = 3, at = -0.1, line = 0.6)
plot(TB2, type = "Isopleth", xaxis1 = "FM", mark = TRUE, contour = 6)
mtext("(b)", side = 3, at = -0.1, line = 0.6)
## Biological reference levels
TB1$df_Es
#> F01 Fmax F05 F04 E01 Emax E05
#> 1 0.2104208 0.2705411 0.1302605 0.1703407 0.4242468 0.5454602 0.262629
#> E04 YPR_F01 YPR_Fmax YPR_F05 YPR_F04 YPR_F01.1 BPR_Fmax BPR_F05
#> 1 0.3434379 1588.519 1620.227 1365.019 1514.626 7668.384 6159.381 10971.29
#> BPR_F04 SPR_F01 SPR_Fmax SPR_F05 SPR_F04
#> 1 9072.983 0.3386681 0.2699639 0.4890713 0.4026261
## Current yield and biomass levels
TB1$currents
#> curr.Lc curr.tc curr.E curr.F curr.C curr.Y curr.V curr.B
#> 1 NA NA 0.4415174 0.2189868 0.3265313 1583.163 0 7483.369
#> curr.SSB curr.SPR
#> 1 7250.44 0.330245
Please note that the resolution of the \(L_c\) and F changes is quite low and the range quite narrow due to the limitations in computation time of the vignette format. The results indicate that the fishing mortality of this example (F = 0.22) is smaller than the maximum fishing mortality (\(F_{max} =\) 0.27), which confirms the indication of the exploitation rate (E = 0.44). The prediction plot shows that the yield could be increased when fishing mortality and mesh size is increased. The units are grams per recruit.
For management purposes, fish stock assessments are mainly conducted
for single species or stocks, which describe the management units of a
population. There is much to be gained from multi-species and ecosystem
models, but data requirements and complexity make them often unsuitable
for deriving management advice. For data-poor fisheries, a traditional
fish stock assessment solely based on length-frequency (LFQ) data of one
year (as presented here) is particularly useful. LFQ data comes with
many advantages over long time series of catch and effort or
catch-at-age data (Mildenberger, Taylor, and
Wolff 2017). In this exercise, the exploitation rate and results
of the yield per recruit models indicate sustainable exploitation. The
exploration of stock status and fisheries characteristics can of course
be extended, but go beyond the scope of this tutorial, which is thought
to help getting started with the length-based assessment of the
TropFishR package. Further details about functions and their arguments
can be found in the help files of the functions (help(...)
or ?..
, where the dots refer to any function of the
package). Also the two publications by Mildenberger, Taylor, and Wolff (2017) and by
M. H. Taylor and Mildenberger (2017)
provide more details about the functionality and context of the
package.
Note that the length at capture has 2 abbreviations \(L_{50}\) and \(L_c\).↩︎