We performed a study on standardizing the catch per unit effort (CPUE) for blue endeavour prawns (Metapenaeus endeavouri) caught in the Northern Prawn Fishery (NPF), one of Australia’s largest and most valuable prawn fisheries. Blue endeavour prawns constitute a significant proportion of the total NPF catches. However, there have been very limited studies on their population dynamics. This study assessed the effectiveness of Artificial Neural Networks (ANNs) for CPUE standardization, with a focus on blue endeavour prawns as a case study. Our approach involved developing new ANN models for CPUE standardization with two key ideas: using an architecture inspired by the catch equation to mitigate overfitting; and using the Tweedie distribution to manage uncertainties and zero counts in the catch data. Specifically, we grouped variables into three distinct modules based on the catch equation, with each representing catchability, fishing effort, and fish density, respectively. Parameter estimation for our ANNs was achieved by maximizing the likelihood using a coordinate descent approach, which alternates between optimizing the Tweedie distribution parameters (power and dispersion) and the standard neural net parameters. We conducted a comprehensive comparison among ANNs, generalized linear models, and generalized additive models. The findings suggest that customizing ANN structure improves model fitting and effectively mitigates the risk of overfitting. It also reveals a promising path for the application of neural networks in CPUE standardization.

Catch-per-unit-effort (CPUE) standardization is crucial for fishery stock assessment but often presents challenges due to spatial-temporal variations in species distribution and fishing effort. In this simulation study, we propose the use of customized artificial neural networks (ANNs) for modeling the spatial-temporal variations in CPUE standardization. This is achieved by encoding prior knowledge of the dependency structure between the variables into the architecture of the ANNs. We conducted numerical experiments on simulated data to compare our customized ANNs with Generalized Linear Models (GLMs), Generalized Additive Models (GAMs), and fully connected ANNs used in previous studies. Our simulated data cover three spatial-temporal dynamics scenarios with different degrees of species distribution shift over time: (1) steady fish distribution; (2) gradual directional shift over time; (3) sudden directional shift. In predicting the standardized CPUE in this simulation study, the customized ANNs demonstrated greater accuracy compared to the commonly used fully connected ANNs with an error reduction of over 70 %, more than 80 % compared to GLMs, and more than 40 % compared to GAMs, in terms of an error metric called the scaled mean absolute relative error. Our findings suggest that customized ANNs can serve as an alternative modeling tool alongside GLMs and GAMs in fisheries modeling.

The prior plays a central role in Bayesian inference but specifying a prior is often difficult and a prior considered appropriate by a modeler may be significantly biased. We propose multi-pass Bayesian estimation (MBE), a robust Bayesian method capable of adjusting the prior’s influence on the inference result based on the prior’s quality. MBE adjusts the relative importance of the prior and the data by iteratively performing approximate Bayesian updates on the given data, with the number of updates determined using a cross-validation method. The repeated use of the data resembles the data cloning method, but data cloning performs maximum likelihood estimation (MLE), while MBE interpolates between standard Bayesian inference and MLE; there are also algorithmic differences in how MBE and data cloning make repeated use of the data. Alternatively, MBE can be considered a method for constructing a new prior from the given initial prior and the data. We additionally provide a new non-asymptotic bound on the convergence of data cloning, and provide an MBE-like iterative heuristic approach which achieves faster convergence speed by boosting posterior variance. In numerical simulations on several simulated and real-world datasets, MBE provides robust inference results as compared to standard Bayesian inference and MLE.

Using length frequency distribution data (LFD) is cost-effective for estimating somatic growth in fish or invertebrates as length data are relatively easy to obtain. The recently developed R packages TropFishR and fishboot extend classic ELEFAN (Electronic LEngth Frequency ANalysis) programs and include more powerful optimization procedures and a bootstrap method for estimating uncertainties. Yet, the fundamental functions require users to provide search conditions (e.g. upper and lower limits for each parameter, length-class size, number of length-classes for the calculation of moving average), which can significantly affect the results. In this paper, we compare the ELEFAN approach with a Bayesian approach in analysing LFD, employing both standard and seasonal von Bertalanffy growth functions. We apply both approaches to a commercially valuable but poorly studied red endeavour prawn (Metapenaeus ensis) harvested in Australia's Northern Prawn Fishery. Sensitivity tests on ELEFAN confirm that any change in search settings would affect the results. Simulation studies on Bayesian growth models show that L-inf and K can be accurately obtained even with modal progression of only one year-class and using non-informative priors. However, age information, including the theoretical age at length zero (t(0)), is difficult to estimate and requires LFD from multiple age classes and informative priors. The Bayesian models yield mean parameters of: L-inf = 36.56 mm (carapace length), K = 2.74 yr(-1), and t(0) = -0.02 yr for the males, and L-inf = 51.81 mm, K = 1.94 yr(-1), and t(0) = -0.02 yr for the females. Seasonal oscillation models fit the LFD better, but the improvement is small and the estimated season-related parameters have large variances.

Sharks (class Chondrichthyes, subclass Elasmobranchii) typically have a long lifespan, slow growth rate and low fecundity, leading to low productivity and hence relatively high vulnerability to fishing. Information for managing fisheries catches of elasmobranchs is often lacking, because elasmobranchs are usually non-target species, often with low abundance. It is more feasible to develop management reference points for elasmobranchs based on their life-history information than through traditional stock assessments. The natural mortality rate (M) is the leading life-history parameter (LHP) required by many methods for developing reference points and is itself often derived indirectly from other LHPs. In this paper, we evaluate nine M estimators, using 15 shark stocks in the Western and Central Pacific Ocean as examples. We then compare four methods for developing fishing mortality reference points F-msy, including empirical modelling, the Euler-Lotka equation, an M-productivity-based method and the spawning potential rate converted method. Our analyses show that popular M estimators developed mainly from teleost data resulted in large deviations from the average and were not suitable for elasmobranchs. All four methods for estimating F-msy performed similarly. However, the empirical method is very simple and cost-effective and tended to produce smaller deviations from the average than the other three methods. Nonetheless, it is recommended that multiple methods should be used to minimize possible bias and reflect uncertainty, if the required LHPs are available.

Analysis of spawning biomass per-recruit has been widely adopted in fisheries management. Fishing mortality expressed as spawning potential ratio (SPR) often requires a reference point as an appropriate proxy for the fishing mortality that supports a maximum sustainable yield-F-MSY. To date, a single generic level betweenF(30%)andF(40%)is routinely used. Using records from stock assessments in the RAM Legacy Database (RAMLD), we confirm that SPR at MSY (SPRMSY) is a declining function of stock productivity quantified byF(MSY). We then use general linear models (GLM) and Bayesian errors-in-variables models (BEIVM) to show that SPR(MSY)can be predicted from life-history parameters (LHPs, including maximum lifespan, age- and length-at-maturation, growth parameters, natural mortality, and taxonomicClass) as well as gear selectivity. The calculated SPR(MSY)ranges from about 13% to 95% with a mean of 47%. About 64% of the stocks in the RAMLD require SPRMSY > 40%. Modelling SPR(MSY)reveals that LHPs plusClassexplain 61% of the deviance in SPRMSY. Faster-growing, low-survival, and short-lived species generally require a high SPR. With equal LHPs, elasmobranchs require about 20% higher SPR(MSY)than teleosts. WhenF(MSY)is estimated from fisheries that harvest older fish, increasing the vulnerable age by one year leads to about an 8% increase in SPRMSY. The BEIVM yields smaller variance and bias than the GLM. The models developed in this study could be used to predict SPR(MSY)reference points for new stocks using the same LHPs for calculatingF(x%), but without knowledge of the stock-recruitment parameters.