Dataset DOI: 10.5061/dryad.0p2ngf2hj

This repository contains the data and R code used to estimate the abundance and habitat-specific density of Aldabra giant tortoises (Aldabrachelys gigantea) on Aldabra Atoll, Seychelles, using hierarchical Bayesian distance sampling.

The data and code support the manuscript:

Inferring the abundance of an island megaherbivore from semi-structured distance sampling

Repository contents

• AnalysisCode.R
  Main R script containing data preparation, model definition, MCMC fitting, posterior extrapolation, and figure generation.
• torts.rds
  Individual tortoise observation data. Each row represents one tortoise detected during a transect survey, including perpendicular distance from the transect line and associated survey metadata.
• M.rds
  Transect-section level survey data. Each row represents one sampled transect section during a survey occasion, including observed count, habitat type, survey year/season, and transect metadata.

File formats

The data are stored as .rds files, which are native R serialized data objects. They can be read into R using:

readRDS("torts.rds")
readRDS("M.rds")

Study design and key terminology

Transects were divided into 50 m analytical units. In the manuscript these are referred to as transect sections. In parts of the model code, the same 50 m units are referred to as sites. These terms therefore refer to the same analytical sampling unit in this repository.

Perpendicular distances from the transect line were recorded for detected tortoises and used to fit a distance-sampling detection function. The main analysis focuses on wet-season surveys, when tortoise availability for detection is expected to be higher and more consistent than in the dry season.

The abundance estimates should be interpreted as applying primarily to visible adult and sub-adult tortoises. Small juveniles are likely to be under-represented because they are cryptic and less consistently detectable during transect surveys.

Habitat codes

The following habitat codes occur in the data and/or analysis:

The analysis code harmonises some historical habitat labels before model fitting:

CRB -> EXP
CSB -> SND
MSC -> SMX

Data dictionary: torts.rds

torts.rds contains individual-level tortoise observations. Each row is a single detected tortoise.

Data dictionary: M.rds

M.rds contains transect-section level count data. Each row represents a sampled 50 m transect section during a survey occasion.

Note that variable M within the dataframe M.rds is a legacy/helper index column retained from earlier data preparation. This column is not used in the final analysis and is removed in AnalysisCode.R before model fitting. It should not be confused with M_super, the fixed superpopulation size used in the Bayesian data-augmentation model.

Required R packages

The analysis uses the following R packages:

library(tidyverse)
library(nimble)
library(coda)
library(ggforce)
library(patchwork)
library(scales)

Install any missing packages before running the analysis, for example:

install.packages(c("tidyverse", "coda", "ggforce", "patchwork", "scales"))
install.packages("nimble")

How to run the analysis

1. Download this repository.
2. Open R or RStudio and set the working directory to the folder containing:

AnalysisCode.R
torts.rds
M.rds

1. Run the full script:

source("AnalysisCode.R")

Alternatively, open AnalysisCode.R in RStudio and run the code section by section.

Analysis workflow

The script performs the following steps:

1. Load and prepare data
   The script reads torts.rds and M.rds, filters surveys to the wet season, harmonises habitat labels, creates survey occasion identifiers, and links individual tortoise observations to transect-section metadata.
2. Define the hierarchical distance-sampling model
   The model is implemented in nimble. The observation component uses a half-normal detection function and a custom distance likelihood implemented via the zeros trick. The state process estimates latent abundance for each transect section and survey year using a superpopulation formulation.
3. Run MCMC
   The model is fitted using nimbleMCMC. The default settings in the script are:

nburnin = 75000
niter   = 250000
nchains = 2
thin    = 20

This step is computationally intensive and may take a long time depending on the computer used.

1. Extrapolate to atoll-wide abundance
   Posterior samples of section-level abundance are aggregated by habitat and year, converted to density using the surveyed strip area, and scaled by mapped habitat area to estimate total population size. The extrapolation uses habitat-specific mapped proportions and excludes areas not included in the abundance calculation as described in the manuscript.
2. Generate figures
   The script produces visualisations of:
   • annual atoll-wide population estimates,
   • habitat-specific detection coefficients,
   • habitat-specific section-level abundance,
   • fitted detection functions and observed distance data,
   • the posterior distribution of the global abundance-variation parameter.

Model outputs

The fitted MCMC object can be saved by uncommenting the following line in AnalysisCode.R:

saveRDS(out1, "Hybrid_Model_Final_Run.rds")

Once saved, the model can be reloaded without refitting:

out1 <- readRDS("Hybrid_Model_Final_Run.rds")

This is recommended because the full MCMC run is computationally expensive.

Reproducibility

The analysis is fully reproducible from the included .rds files and AnalysisCode.R, provided the required R packages are installed. Results may vary slightly between runs because the MCMC initial values are randomly generated unless a random seed is set before model fitting.

To improve reproducibility, users may set a seed before running the model:

set.seed(123)

Citation

When using these data or code, please cite the associated manuscript:

Hillyard et al. Inferring the abundance of an island megaherbivore from semi-structured distance sampling. Ecology and Evolution.