Species Occupancy Modeling for Detection Data Collected Along a Transect

1. What they did

Occupancy is the share of survey sites where a species is present. Because animals can be present but missed, occupancy surveys need a way to estimate how often a species goes undetected. For wide-ranging animals such as tigers, surveyors often walk transects and record each sign they find. Until now these records were analysed by cutting each transect into segments (for example 1 km pieces) and noting whether the species was found in each piece. The choice of segment length is arbitrary.

The authors, statisticians at the University of Kent and a Fauna & Flora International (FFI) programme manager, built new occupancy models that use the exact position of each detection along the transect, without cutting it into segments:

• A simple model that treats detections as scattered at random along the transect (a Poisson process).
• A clustered model that allows detections to come in bunches (a two-state Markov modulated Poisson process, or 2-MMPP). In this model the surveyor moves between stretches where signs are found often and stretches where they are found rarely, for example where a species only uses part of a site, or where animals cross or follow a trail now and then.

They worked out formulas for study design, tested both models with computer simulations, and then applied them to footprint surveys of the Sumatran tiger. Survey teams walked transects in 89 sites (17 × 17 km squares) in and around Kerinci Seblat National Park, Sumatra, Indonesia, in 2007 and 2008, noting the location of every tiger footprint found. They covered 2826.5 km in total, from 1.8 to 108.3 km per site.

2. What they found

Study design

• For a fixed total effort, the best transect length per site depends on how common the species is. For rare species it is better to survey more sites; for commoner species, fewer sites surveyed more intensively.

Simulations

• With random (unclustered) detections, cutting transects into segments gave occupancy estimates with more bias and more scatter than the continuous model, especially with long segments.
• For rare and elusive species (few occupied sites and few signs per km), all estimates were biased and less precise than the theory predicted, so the authors recommend using simulations when planning a survey.
• When detections were clustered, the simple model underestimated occupancy, sometimes badly (for example 0.38 instead of 0.75). The clustered model gave estimates close to the true values.

Sumatran tiger footprints

• Tiger footprints were found in 66 of the 89 sites, a naïve occupancy of 0.74.
• The clustered model fitted the data far better than the simple model (its AIC was almost 60 units lower). It estimated occupancy at 0.96 (95% confidence interval 0.831 to 1.000), compared with 0.82 from the simple model.
• In occupied sites, footprints were met at about 0.23 per km in some stretches and about 0.03 per km in others. The authors describe this as about ten times higher in some areas than in others. The estimated chance of being in the high-rate state was about a third (0.35).
• The data said little about how long each stretch lasted, because the transects (mostly 3 to 21 km) were short compared with these stretches. A simpler model, in which each transect has either a high or a low rate throughout, fitted almost as well and gave practically the same estimates.
• Checks of the distance to the first footprint and the distances between footprints confirmed the better fit of the clustered model.

3. What they concluded

• Modelling detections as a continuous process is a natural way to analyse data collected along transects, and it removes the need to choose a segment length. Cutting transects into larger segments can give poorer occupancy estimates, because information is lost.
• Ignoring clustering in detections makes occupancy look lower than it is.
• The clustered model suited the Sumatran tiger footprint data better than the simple model. In their introduction the authors note that tigers are believed to use ridges often to move across their territory, and the transects often followed ridges, so some clustering was expected.
• The same models apply to surveys that record detections over continuous time, such as camera traps and bird point counts.

4. How tracking was or was not used

Survey teams walked transects in the forest and recorded the location of every tiger footprint they found. These footprint positions along each transect are the data the models analyse. The paper does not describe how footprints were identified or how the transects were chosen beyond often following ridges.

5. Tracker or observer certification

Not reported. The paper does not describe the survey teams or their skills. The acknowledgements thank Yoan Dinata, Agung Nugroho, Iding Achmad Haidir and Maryati for help with data collection and entry. The models assume every recorded footprint is a true tiger detection, so the occupancy estimates depend on correct footprint identification.

6. Other technologies, skills or innovations

• New continuous-detection occupancy models (Poisson process and Markov modulated Poisson process), fitted by maximum likelihood in MATLAB.
• A formula (using the Lambert W function) and a table for choosing the best transect length per site.
• Survival-analysis tools (Kaplan–Meier curves) used to check how well the models fit the distances between footprints.

7. Implications for conservation (from the paper)

• Occupancy is widely used in wildlife monitoring and conservation, and these models can give less biased estimates from transect sign surveys of wide-ranging species such as tigers.
• Survey effort can be planned to give the most precise occupancy estimate for a given budget of kilometres walked.