By Georgia Garrard (This article was first published in the February 2013 issue of Decision Point, The Monthly Magazine of the Environmental Decisions Group)
Imperfect detectability of plants and animals during ecological surveys is a major issue. Get it wrong and you run the risk of endangered species being lost when development is permitted but the species was still present. Or, on the other hand, you risk an invasive species breaking out when control efforts are stopped too soon because the invading organism was thought to be not there. Imperfect detectability can lead to biased estimates of abundance or occupancy, impaired ability to detect change or response to management action and ultimately leads to poorly informed management decisions.
These days the problem of imperfect detectability is widely recognised, and considerable research has been done in this area by the Environmental Decisions Group.
A range of methods exist for estimating detectability, including distance sampling, mark-recapture, N-mixture models, zero-inflated occupancy models and time-to-detection models. Each of these models has its own set of assumptions, data requirements and applications. Often, the data requirements of these models are heavy; data, including abundance counts, presence-absence observations and times-to-detection, are variously required from multiple sites and multiple observers. What that adds up to is that, at this time, estimates of detection probability are available for relatively few species.
If you don’t have a species-specific model to guide how much effort you need to expend to detect a species that is present (with confidence), what can you do? Maybe you could focus on some character trait of that species as a generic guide? For example, if you were aware that it took a certain amount of effort to detect a weed species with red flowers, maybe that knowledge could help you decide how much effort would be required to find an unrelated species that also had red flowers. We asked whether we can learn about the influence of species traits on detectability, and used trait-based models to predict the detectability of species for which no species-specific model exists.
“Trait-based models should provide sensible bounded estimates of detectability on which to base survey design and effort requirements.”
Using a time-to-detection model (see Garrard et al. 2013 and Box 1), we investigated the influence of a range of species traits on the detectability of grassland plant species. Examples of the traits investigated were local abundance, height, likelihood of flowering at the time of survey, flower colour, leaf area, number of similar grassland species and whether the species grows in clumps.
Box 1 Have we looked hard enough?
How long do we need to spend surveying in a single visit to a site to achieve a reasonable chance that we will detect a threatened plant species if it’s present (Garrard et al. 2008)?
I investigated detectability issues in the Western (Basalt) Plains grassland community on the northern and western fringes of Melbourne. This vegetation community is listed as critically endangered, however, its close proximity to Melbourne’s urban growth boundary means there is continual pressure for development in areas where remnants occur.
The Western Plains grassland is home to a number of nationally- endangered plant species, including the spiny rice-flower, Pimelea spinescens subsp. spinescens. I determined how ‘detectable’ this species was during a flora survey. During a multi- site, multi-observer field study I collected data on the time at which this species was first detected in 1 ha plots. I also looked at variables that were likely to influence the rate of detection, such as the experience of the observer, the date and time of day, weather conditions and the cover of the dominant grass species, kangaroo grass (Themeda triandra).
What I found suggests that the chance of failing to detect this species is potentially very high. Even at sites where the species was known to exist, only around half of the observers detected it. The average time to detection was strongly influenced by observer experience and the cover of kangaroo grass at the site.
So, how hard do we need to look for this species? Under the most favourable survey conditions, the average time to detection is 26 minutes per hectare. But if we want to be 80% certain that we will detect the species if it is present, we need to allocate around 42 minutes per hectare. And to increase this certainty to 95% requires a survey effort of 78 minutes per hectare.
However, the survey effort required to achieve a reasonable probability of detecting the species if it is present may increase significantly under sub-optimal survey conditions. For example, to achieve a 95% probability of detection using a less experienced observer would increase the survey effort required to over 2 hours per hectare. Under the average conditions experienced during my field study, the average time to detection of this species was around 3.5 hours per hectare. Importantly, this is significantly higher than the level of effort commonly expended during ecological impact assessment surveys in the region.
We found that local abundance has a clear influence on detectability, with species that occur in higher numbers having lower detection times (higher detection rates) than those occurring in small numbers (Figure 1). Species are also more likely to be detected if they are unique or in their peak flowering month at the time of survey, although these results are less definitive.
“It’s impossible to consider constructing a species-specific detectability model for every threatened species.”
Our results also show that flower colour may have a large effect on detectability, with pink and red flowered species potentially more easily detected than those with inconspicuous or yellow flowers. This makes sense in native grasslands, where there are many yellow flowers and few pink or red flowers. The influence of flower colour is still very uncertain. We used only coarse categories and it will be interesting to see whether more objective measures of flower colour can help resolve this uncertainty.
Using trait-based detectability models, we were able to predict average times-to-detection reasonably well for new species (Figure 2).
I’m probably biased, but I think trait-based detectability models are an exciting development in the field of detectability research and have enormous potential to improve our environmental decision making. With more than 1300 nationally-listed threatened plant species and another 400 animal species in Australia alone, it’s impossible to consider constructing a species-specific detectability model for every threatened species. Consider the effort expended on developing guidelines for the detection of the spiny rice flower (see box 1).
While they may not perfectly predict individual species’ detection probabilities, trait-based models should provide sensible bounded estimates of detectability on which to base survey design and effort requirements.
- A new, trait-based model of detectability (ggarrardresearch.wordpress.com)