Seabird Risk Assessment

This is an example of carrying out a Seabird Risk Assessment, following the Spatially Explicit Fisheries Risk Assessment framework (SEFRA; see Chapter 3 of this PDF). Within the SEFRA framework, seabird fatalities in fisheries are assumed to be proportional to both the fishing effort, and the seabird density (birds per square kilometer). The constant of proportionality is the vulnerability, and may be estimated from observer data.

The fisheries and observer data used here are simulated. This project acts as a template for carrying out a risk assessment analysis, showing the format of input files and the process needed for carrying out the assessment. The assessment is carried out on simulated on seabird captures, in commercial longline fisheries in the southern hemisphere, assuming a known vulnerability. Only the great albatrosses group (Diomedea) are included.

The risk assessment prepares input data, runs a Bayesian model, and carries out simple analysis and display of the results.

Process

To run this test example, clone the project from Github (git clone https://github.com/seabird-risk-assessment/seabird-risk-assessment.git). If you have docker installed on an Ubuntu system, you can then run make to run the test project. A docker image containing all the necessary software is available from DockerHub.

Otherwise, install the following software:

Software requirements

• R, with the packages data.table, lubridate, sf, rjags, coda, stringr, ggplot2 to process the data, orchestrate the model, and to get the results; and rmarkdown, knitr, kableExtra, and shinystan for the viewing the results and generating the report.
  • ggpplot2: Currently, a development version of ggplot2 is required to plot the maps. First, install the devtools package, then run devtools::install_github("tidyverse/ggplot2")
  • sf: The current vesrion of sf has a bug that prevents the maps from being generated. Install the development version to avoid this bug: devtools::install_github("r-spatial/sf")
• Stan, to run the Bayesian model
• GNU make, is not necessary but highly recommended to run the whole project at once. Windows users can get GNU make along with other common programs part the UNIX-based environment by installing the R package Rtools. The installation guide can be found here.
• Pandoc is required to compile the report. This should be pre-installed with R-studio.

This project has been tested in Ubuntu 16.04.

Running the risk assessment

If you have GNU make installed, you can run the entire project by running make in the base directory of the project. This can be run from within the docker image. Before running make ensure that the environment variable RUN is not set. In a Linux system (such as within the docker container), this can be achieved by running RUN= make.

The relationship between the individual scripts is shown in build diagram. If you don't have make available you can follow the dependencies shown in that digram.

When the assessment is complete, there will be a report model-report.html in each of the model directories, e.g, in the model/stan directory.

Data

Spatial grid

A spatial grid is provided in the input-data/southern-hemisphere-5-degree directory. This grid is 5-degree by 5-degree, covering the southern hemisphere. It is provided as a shapefile (in WGS84 latitude-longitude projection) that can be loaded into GIS tools, or queried from software such as R.

The grid has been clipped to coastlines used by the American Bird Conservancy in the development of their seabird distributions. For each cell, the shape file has the following fields:

• grid_id : integer index, unique for each grid cell
• lon : longitude of the center of the cell
• lat : latitude of the center of the cell
• area : area of each cell in kilometers squared

Note that a grid is not necessarily required by the SEFRA method, however it is convenient to have a common resolution for preparing the data.

Fisheries effort data

Observed and total fishing effort.

For both observed and total fishing effort, a csv file (input-data/fishing-effort-observed.csv or input-data/fishing-effort-total.csv respectively) is provided with the following fields:

• fishery_group : Fishery group; single fishery or group of fisheries. Fisheries of the same group are assumed to share the same vulnerability to capture.
• year : 4-digit year
• month : 1 to 12 (month of the year)
• grid_id : index to the spatial grid
• effort : annual effort (the effort measure may be different between fisheries, for surface longline fisheries, number of hooks is a standard measure)

In this example, the fishery group was taken as the flag state of the countries with the highest fishing effort, with smaller fisheries grouped as other. The data may include other columns (such as RFMO, flag state, data source).

The fishing effort was based on publicly available data from RFMOs, but has been randomised, and so does not represent actual fishing effort. Five percent of the fishing effort in each cell was assumed to be observed.

Seabird data

Studied species and groupings

To identify seabird species, use the three-letter FAO codes from the ASFIS database. An extract from this data for seabird species is in data/birds/fao_birds.csv. The codes are intended to allow captures to be linked to the corresponding taxonomic data. The FAO do not include all seabirds, and so custom codes may be needed.

The list of species considered in the study is in input-data/species-list.csv, with the code, scientific name, and common name of each taxon. This is also where the species groups are defined, with taxa of the same group sharing the same vul_group code. Taxa of the same group are assumed to have the same vulnerability to capture.

Some seabird captures are not recorded at the species level but at a higher taxonomic level, because of identification issues.
These groups of species feature in the species list, with the hierarchy between taxa defined by specifying the parent code of the higher taxonomic level. Currently, only one group of unidentified species is used in the model. The column base_species indicates the taxa for which the estimation of captures and the reporting are carried out (TRUE for reported taxa, FALSE otherwise)

At-sea distributions

A polygon shapefile for each taxon (e.g. input-data/species-distributions/Diomedea_amsterdamensis.shp), in the ESRI format, with the same geometries as in the generic grid, and with the following attributes:

• grid_id : index to the spatial grid
• species_code : code of the taxon
• taxa : scientific name of the taxon
• density : bird density within each grid polygon, normalised so that the product of density and the area (in km^2) of each polygon sums to one across all polygons. When multiplied by the total population size, the density is in birds per km^2

These shapefiles are from the American Bird Conservancy.

Demographic parameters

A csv file (input-data/demographic-parameters-processed.csv), containing the mean and uncertainty of key demographic parameters used in the study for each taxa.

The ields are:

• species : species code
• parameter : one of:
  • survival_current : Current adult annual survival rate, including human-caused mortality
  • survival_optimal : Optimal adult annual survival rate, which would occur in absence of human-caused mortality. Used for the Population Sustainability Threshold
  • age_first_breeding : Mean age at first reproduction
  • population_total : Total number of individuals (typically derived from the number of annual breeding pairs)
• mean, sd : mean and standard deviation of the parameter
• mean_log, sd_log : mean and standard deviation of the parameter in the log-scale (for log-normal distributions)
• min, max : minimum and maximum value used for truncating the generated distributions, to keep the parameters within plausible range.

Observed captures

A csv file (input-data/observed-captures.csv) with the following fields:

• species : code for each seabird species
• species_group : code for each seabird species group
• fishery_group : fishery group
• taxa : scientific name of the species
• year : 4-digit year of capture
• month : 1 to 12 (month of the year)
• grid_id : index to the spatial grid
• captures : total number of captures (live and dead)
• alive : number of observed live captures
• dead : number of observed dead captures

To generate the seabird bycatch data, the total number of incidents was drawn randomly from a Poisson distribution, with a mean depending on the overlap between the fisheries and the species, and the vulnerability coefficients.

The vulnerability intercept was set to 0.05, and for the fishery group, species group, and the interaction, the vulnerability values were drawn randomly from a log-normal distribution of mean 0 and standard deviation of 0.5 (on the log scale).

From the number of incidents, the number of observable captures, then the number of live and dead captures, were drawn randomly from a binomial distribution, assuming a probability of of a capture being observable was taken to be 0.5, and the probability that a capture was released alive of 0.3. A proportion (30%) of observable captures were assumed to be unidentified, and were as unidentified wandering albatross.

The vulnerability and parameter values used to simulate the data can be found in the folder test-data.

Model

The Bayesian model to estimate the vulnerability to capture and the number of annual potential fatalities is currently coded in the open source software Stan.

The estimation of vulnerability and the predictions of annual potential fatalities are carried out simultaneously. The estimation is done on the observed fishing effort aggregated by species, species group, and fishery group, summing the fishing effort and the number of captures over the period 2004 to 2014. The predictions are carried out on observed fishing effort (to compare observed and predicted captures), as well as on the total fishing effort, aggregated by species, species group, fishery group, grid cell, and month. The aggregation for the predictions are on data between 2012 and 2014, to reflect recent fishing effort and distribution, taking the annual average of fishing effort, so that the predicted captures are also annual. The periods of the data used in the aggregation for the estimation or the prediction stages are specified in the file parameters.r.

From the MCMC output of the model, a posterior sample of the annual potential fatalities is returned for each stratum, which can be summed over strata to get for example the total number of annual potential fatalities by species, or by fishery group.

The aggregation levels of the data to be used in the model are flexible and may be changed in the script that prepares the model data model/jags/prepare-model-data.r.

Report

Results are reported and summarised in a R Markdown document in report/report.Rmd, generating an HTML report that can be viewed in any web browser after being compiled by the R package knitr.

In particular, the document reports on:

• the diagnostics of the model, including the comparison of observed and predicted captures
• the estimated annual potential fatalities (APF) by species
• the spatial distribution of APF
• the comparison of the model results with the fixed parameters that were used to simulate the data

An open project

This project is licensed under the open-source MIT licence, which permits re-use or adaptation of this code. The seabird distributions provided by the American Bird Conservancy are not covered by this licence.

You are welcome to make contributions to this project. Note however, that this is a test risk assessment. Please do not commit any fisheries or bycatch data to this project.