A Bayesian hierarchical occupancy model for track surveys conducted in a series of linear, spatially correlated, sites

1. What they did

Many wildlife agencies monitor animals by recording whether their signs, such as tracks, scats or scent marks, are present or absent. To get reliable results from these surveys, it helps to know how often signs are missed or misidentified. Wildlife managers in Minnesota, United States, wanted a way to follow the status and trends of river otters, because trappers had asked for a harvest season in the south-east of the state, where managers feared otter populations might be small and isolated.

The authors used otter snow-track data that the Minnesota Department of Natural Resources collected in 2003:

• The survey. Observers flew in a helicopter along 56.3 km of the Mississippi River, 1, 2 or 3 days after snowfalls of more than 2.5 cm. There were five snowfall events, 20 flights and three observers. Seven flights covered only the lower half of the route.
• Recording tracks. Observers logged a Global Positioning System (GPS) waypoint when they met an otter track and every 5 seconds while the track continued. When a track was hard to identify, the pilot circled until the observer was comfortable deciding.
• Making sites. Afterwards, the river was divided into 140 segments of 402 m, and each segment was scored as track present or absent for each flight.

They then built a 3-level occupancy model that separates three things: whether a site lies within at least one otter’s home range (occupancy), whether an otter has walked through and left tracks since the last snowfall (track laying, which depends on movement), and whether the observer recorded a track. The model allowed both missed tracks and false records, and let each observer have their own detection rates. A second, “spatial” version allowed neighbouring sites to be linked, because otters travel long distances and one otter’s trail can cross several sites.

They tested both versions with 81 simulated scenarios, fitted them to the Mississippi River data, and used further simulations to work out how many flights, snow events, days and observers a future survey would need.

2. What they found

Simulations

• Both models gave precise estimates with little bias.
• When neighbouring sites were linked in the simulated data, the spatial model was better at estimating how likely tracks were to be laid. The baseline model tended to overestimate it.
• The spatial model’s 95% credibility intervals for the occupancy rate held the true value in 95% of scenarios, against 91% for the baseline model. For the track-laying probability the figures were 93% and 78%.
• Estimates of the occupancy rate became more precise as tracks became more common.

The otter data

• A run test showed that sites were not independent of each other.
• The spatial model estimated that 82% of sites were occupied, compared with 59% for the baseline model.
• The spatial model estimated lower track-laying rates (0.12 against 0.15 per day for occupied sites), but strong links between neighbouring sites: when an otter did leave tracks, the trail was likely to cross site boundaries.
• Observers recorded a track that was present about 61% to 69% of the time, and recorded a track where none was present 1% to 5% of the time, depending on observer and model.
• Observer-to-observer differences were notable. The authors suggest two likely reasons: tracks near site boundaries assigned to the wrong site after the flights, and the difficulty of correctly identifying otter tracks.

Survey design

• To reach the precision goal (a standard error below 0.05 for the occupancy rate), a survey needed at least 15 to 20 helicopter flights spread over at least three snowfall events, with three surveys on consecutive days after each snowfall.
• The number of days surveyed after each snowfall mattered most, the number of snowfall events came second, and the number of observers mattered least.

3. What they concluded

• The model brings together false detections, links between neighbouring sites, and the way detection depends on time since snowfall. The authors knew of no earlier study that combined all three.
• Their pilot data suggest that detection was driven mainly by otter movement after snowfalls, and much less by the difficulty of seeing tracks from the air once they were there.
• They recommend at least 15 to 20 flights over at least three snowfall events, with at least three successive daily surveys after each event.
• In future, sites should be defined on a grid before flying, with GPS used to place each record in its site, to reduce misclassification.
• False positives deserve more attention in occupancy models because they can strongly affect occupancy estimates. The authors strongly recommend testing such models with simulated data.
• The approach should suit many other surveys of animal sign, such as scent station surveys for carnivores, and surveys that sample clusters of neighbouring sites.

4. How tracking was or was not used

The whole survey rests on tracks and signs identification from the air: observers in a helicopter looked for otter tracks in fresh snow along the river and recorded where they saw them. When a track was hard to identify, the pilot circled the area until the observer was ready to decide whether an otter track was present. No trailing on the ground is reported. The model treats tracks as building up day by day after each snowfall until the next snowfall, wind or melting erases them.

5. Tracker or observer certification

Not reported. Three observers flew the surveys, but nothing is said about their experience or skill. This matters to the findings: the model estimated that observers recorded tracks that were present only about 61% to 69% of the time, and the authors name the difficulty of correctly identifying otter tracks as one likely cause of observer differences, citing Evans et al. (2009) on observer reliability in counts of river otter tracks.

6. Other technologies, skills or innovations

• Bell OH-58A+ helicopter for aerial snow-track surveys.
• GPS waypoints logged every 5 seconds while a track was in view.
• Open-source software: Program R and WinBUGS, linked through the R package BRugs. The R code to simulate data and fit both models is provided with the paper.
• A model component for how tracks accumulate after a snowfall, which the authors compare to temporary emigration in mark–recapture models.

7. Implications for conservation (from the paper)

• Agencies need efficient, cost-effective monitoring tools that can separate real population trends from changes in detection.
• The model gives managers a tool for designing surveys and for judging whether they can detect the effects of management, such as higher harvest pressure or efforts to reduce wetland drainage and pollution.
• Following a river in a single flight path is efficient, but it makes neighbouring sites similar; being able to account for this avoids flying less efficiently or throwing away data.