Optimization of an artificial neural network for identifying fishing set positions from VMS data: An example from the Peruvian anchovy purse seine fishery
Introduction
The ecosystem approach to fisheries management is increasingly calling for spatially explicit indicators (e.g. Pikitch et al., 2004, Babcock et al., 2005). Among those indicators, fishing effort is crucial for at least two reasons: (1) the need to control the compliance with spatially explicit management measures (such as inshore restrictions or Marine Protected Areas), and (2) the need to improve the interpretation of catch-per-unit of effort (CPUE) data in terms of fish stock abundance (see for instance problems related with spatial hyper aggregation, Rose and Kulka, 1999).
The effort deployed by fishing fleets can be spatially examined by VMS data. VMS provide high-resolution records of vessel positions on a regular time basis, ranging from few minutes to few hours depending on the fishery. However, while they are available for numerous fisheries and vessels (www.fao.org), this type of data has been barely used in fisheries science and management. One of the reasons is that VMS data do not provide explicit information on whether a vessel is fishing or not. Then, a first step for processing those data consists in estimating from position records where fishing events (set, haul, trawl, etc.) probably occurred. This is not a trivial task and several approaches have been undertaken, from very coarse ones (such as unique speed thresholds applied to different activities, Witt and Godley, 2007; and areas of high VMS poll density summed up with kernel home ranges, Harrington et al., 2007) to more refined ones (such as a combination of speed thresholds, directionality and other complementary rules; Deng et al., 2005, Mills et al., 2007). Although those methods usually detect quite properly true positives (a true fishing set detected as a fishing set), false positives and over (or under) estimation errors are rarely assessed. For the present case study, the Peruvian purse seine anchovy fishery, Bertrand et al. (2008) showed that the use of a simple speed threshold on raw VMS data leads to an overestimation of the number of fishing sets of 182%. Alternatively, a general linear regression modeling approach (GLM) identified 65% of true positives and 16% of false positives, leading to a global underestimation of the total number of fishing sets of 19%. Thus, none of these two approaches are satisfying since they strongly bias the estimation of the number of fishing sets.
To solve this problem for the Peruvian anchovy fishery, Bertrand et al. (2008) proposed an ANN approach based on a Multilayer Feed-Forward Network (MFN). This methodology was chosen because (1) ANNs do not require to know nor to assume any probability distribution function, (2) ANNs are adapted for working with large datasets linked by complex non linear relations, and (3) among ANNs, MFNs are commonly used because of their simplicity and the wide availability of software tools. This particular ANN is first trained on a subset of fishing trips for which fishing set positions are documented by at-sea observers (∼1% of the total fishing trips) and then used to estimate the location of fishing sets for the remaining trips monitored by VMS only. This ANN is designed to overcome the overestimation problem as it aims at: (1) accurately assessing the total number of fishing sets and (2) maximizing the true-to-false positives ratio. Bertrand et al. (2008) applied this tool on a rather limited spatial area (7°S–10°S along the Peruvian coast) and temporal window (2000–2002), correctly identifying 83% of the real fishing sets (true positives) with a total overestimation of 0.5%.
Based on those satisfying preliminary results, and before implementing such a tool in routine for the monitoring dashboard of the Peruvian anchovy fishery, there is a critical need to check the behavior and validate an optimal parameterization of the ANN when confronted to variable situations in time and space. We are particularly concerned in answering 3 main questions:
- (1)
At-sea observers data are usually not available on a day-to-day basis; then, when using the ANN in real-time during the fishing season, what is the effect of estimating fishing sets using an ANN trained on an earlier period?
- (2)
Two types of vessels, steel and wooden hulls, are participating to the industrial anchovy reduction fishery; although wooden vessels are only recently and gradually incorporated to the VMS monitoring. What may be the effect in the ANN behavior of this change in the fleet composition in VMS data?
- (3)
Management rules differ between two large regions in the coast of Peru (coastal restrictions, fishing bans, total allowable catch); does the ANN need different optimizations for the two regions or can we use a single ANN along the entire Peruvian coast?
Those questions are addressed performing a global optimization of the ANN. ANN optimization ranges from trial-and-error sensitivity analysis (e.g. Dedecker et al., 2004) to more complex and efficient approaches such as genetic algorithm or simulated annealing (e.g. Mühlenbein, 1990, Sexton et al., 1999, Bernardos and Vosniakos, 2007). These latter optimizations mainly concern network architecture (e.g. number of neurons and layers, and shape of activation functions) and convergence rules for the training algorithm (Bernardos and Vosniakos, 2007). Here, we refer to the optimization of (1) the internal structure and training algorithm of the ANN and (2) the “rules” used for choosing the relative size and composition of the DBs for the ANN training and inference. To address these two aspects with comparable methods, we use a trial-and-error sensitivity analysis.
In the next section, we present some characteristics of the Peruvian fishery, the data used in this study, the ANN architecture and training algorithm; and then we describe and perform a series of sensitivity tests. From the results obtained in terms of ANN performance, we draw practical recommendations for an optimal and robust use of this tool, in the specific case of the Peruvian anchovy fishery. Finally, we discuss on how this type of neural network approach can have wider potentials and could be implemented, with adjustments on the input variables, in any fishery relying on both VMS and at-sea observer data.
Section snippets
Some insights into the Peruvian anchovy reduction fishery, its monitoring and management and the data used
The Peruvian anchovy fishery is characterized by the remarkable size of its production (∼7 millions t.y−1 since 1999) and its sensitivity to the intense regional climatic variability on various spatio-temporal scales (Chavez et al., 2008). Indeed, climatic scenarios such as El Niño or la Niña events directly condition the extent of the anchovy habitat, modifying its catchability and driving its population dynamics (Bertrand et al., 2004a). To cope with this strong natural variability, fishing
Sensitivity analysis on the training databases
Training results are synthesized in Table 4. They are consistent with the preliminary ones in Bertrand et al. (2008). ANNs prove to be performing over the full available time series and the entire Peruvian coast. When considering training performances for DBs based on 1-year periods, the global error concerning the total number of fishing sets lies in [−2.8%; 1.7%]. It is significantly smaller than the ones of simple threshold on speed (+182%) and GLM (−19%) approaches evaluated on the
Summary and discussion
In this study, we tested the ability of an ANN trained on at-sea observer data to correctly infer fishing set positions in VMS data. We used an exhaustive fishery dataset from the purse seine anchovy fishery along the Peruvian coast from 2000 to 2007. Firstly, we have verified that an ANN approach provides better results than conventional criteria such as a simple threshold on speed or a GLM approach. For the studied fishery, a threshold on speed overestimates the number of fishing sets by
Acknowledgements
The authors would like to thank Matthieu Lengaigne for priceless discussions on artificial neural networks and sensitivity results. We are also grateful to Erich Díaz and Manuel Ochoa for valuable conversations about vessel trajectories. We warmly thank Jérémie Habasque for his help in the development of the user interface for fishing set estimation. We sincerely thank Charlotte Boyd as well as Johanna Holmgren for revising the English and for helping us make the article more readable.
Rocío Joo is a statistician and a doctoral student at the French Research Institute for Development (IRD), working with the Peruvian Institute for the Sea (IMARPE). She is the main author of the article and she has made the analysis for the calibration of the neural network described in it. She is also the main developer of the user interface for obtaining fishing set estimations.
References (39)
Developments on fisheries management in Peru: the new individual vessel quota system for the anchoveta fishery
Fish. Res.
(2009)- et al.
Patterns in the spatial distribution of Peruvian anchovy (Engraulis ringens) revealed by spatially explicit data
Prog. Oceanogr.
(2008) - et al.
Optimization of Artificial Neural Network (ANN) model design for prediction of macroinvertebrates in the Zwalm river basin (Flanders, Belgium)
Ecol. Modell.
(2004) - et al.
Interdecadal variability of anchoveta abundance and overcapacity of the fishery in Peru
Prog. Oceanogr.
(2008) - et al.
Multilayer feedforward networks are universal approximators
Neural Networks
(1989) - et al.
Application of neural networks to modeling nonlinear relationships in ecology
Ecol. Modell.
(1996) - et al.
Neural networks for the prediction and forecasting of water resources variables: a review of modeling issues and applications
Environ. Model. Software
(2000) Limitations of multi-layer perceptron networks—steps towards genetic neural networks
Parallel Comput.
(1990)- et al.
Optimization of neural networks: A comparative analysis of the genetic algorithm and simulated annealing
Eur. J. Op. Res.
(1999) - et al.
Neural networks in fisheries resources
Fish. Res.
(2008)
Forecasting with artificial neural networks: the state of the art
Int. J. Forecast.
A perspective on the use of spatialized indicators for ecosystem-based fishery management through spatial zoning
ICES J. Mar. Sci.
Optimizing feedforward artificial neural network architecture
Eng. App. Artif. Intell.
From small-scale habitat loopholes to decadal cycles: a habitat-based hypothesis explaining fluctuation in pelagic fish populations off Peru
Fish Fish.
Interactions between fish and fisher's spatial distribution and behavior: the anchovy (Engraulis ringens) fishery of Peru
ICES J. Mar. Sci.
Lévy trajectories of Peruvian purse seiners as an indicator of the spatial distribution of anchovy (Engraulis ringens)
ICES J. Mar. Sci.
Scale-invariant movements of fishermen: the same foraging strategy as natural predators
Ecol. Appl.
Neural Networks for Pattern Recognition
The northern Humboldt Current System: Brief history, present status and a view towards the future
Progr. Oceanogr.
Cited by (43)
Getting to good enough: Performance of a suite of methods for spatially allocating fishing effort to management areas
2018, Fisheries ResearchCitation Excerpt :The VMS-based methods used in this paper rely on simple speed windows. More sophisticated methods have been developed to analyze VMS data that, for example, rely on various spline interpolation techniques (e.g., Hintzen et al., 2010; Russo et al., 2011), artificial neural networks (Joo et al., 2011) and Bayesian models (Vermard et al., 2010; Bez et al., 2011). Many of these methods also make use of information on the instantaneous course of a vessel (e.g., Vermard et al., 2010), though this was not evaluated in our simple comparisons.
Rocío Joo is a statistician and a doctoral student at the French Research Institute for Development (IRD), working with the Peruvian Institute for the Sea (IMARPE). She is the main author of the article and she has made the analysis for the calibration of the neural network described in it. She is also the main developer of the user interface for obtaining fishing set estimations.
Sophie Bertrand is a fisheries ecologist at the French Research Institute for Development (IRD), working with IMARPE since 2001. Her research is based on the statistical analysis and modelling of spatially explicit data from exploited marine ecosystems (Vessel Monitoring System data for vessels, GPS tracking for birds and mammals, underwater acoustics for fish). She has been working with the Peruvian Vessel Monitoring System data since 2002, developing from them a series of metrics, a random walk modelling approach and the beta version of the neural network which behaviour is explored here.
Alexis Chaigneau is a physical oceanographer at the French Research Institute for Development (IRD). His research is focused on the circulation and mesoscale activity of the Humboldt Current System based on various satellite datasets and high-resolution model simulations. He has been working with IMARPE since 2006 and he contributed to the article mainly as an advanced programmer.
Miguel Ñiquen is a fisheries ecologist at the Peruvian Institute for the Sea (IMARPE). He is the head of the IMARPE's department in charge of monitoring and managing the Peruvian industrial fisheries, among which, the Peruvian anchovy purse seine fishery. He is the main daily user of the tool described in the paper and as such has provided its detailed book of specifications.