All ecologists know that wildlife surveys are almost never perfect. Individuals are often missed and even whole species may remain completely undetected even though they are present. Imperfect detection has important implications for wildlife monitoring and species distribution modelling (SDM) and yet it has largely been overlooked in the SDM literature. In an attempt to better understand how imperfect detection affects SDMs, we simulated a range of different detectability scenarios. Our results highlight how managers and conservation workers can improve SDM practice by accounting for detectability (Lahoz-Monfort et al, 2014).
As a nice example of the problems of imperfect detection, consider the Alaotran gentle lemur in northeast Madagascar. It is the only species of primate that inhabits a wetland, and surveys are carried out by canoe along channels cutting through the dense marsh vegetation. Individuals can be easily spotted when they stand at the edge of the vegetation, but their detection becomes extremely difficult as soon as they move slightly deeper into the reeds and papyrus. You don’t see them anymore, but they are still there.
Now you see it, now you don’t! As with so many species, the Alaotran gentle lemur (Hapalemur alaotrensis) is difficult to detect in the wild. Due to the dense structure of most of its marsh habitat, individuals are easily missed unless they stand right at the vegetation edge (as pictured left). In the picture on the right, two individuals stand slightly deeper among the papyrus. Can you spot them?
Factors leading to imperfect detection
Failing to spot animals has direct implications for wildlife monitoring. When not taken into account, it leads to the underestimation of population abundance and species occurrence rates.
Many things contribute to how detectable a species is in a survey. It might be that the habitat of the Alaotran gentle lemur makes it hard to spot, but their behavior being able to stay still and quiet when people are around is also factor. How much effort is put into finding them is also important as is the skill level of the people looking for them.
So detectability is the product of species and habitat characteristics, survey effort and observer skills, and it is therefore likely to vary in space and time. This variation amplifies the problem of imperfect detection as it can lead to the missing of relevant ecological relationships and trends, or the detection of spurious ones. Given that the modeling of species distributions is a fundamental tool in a wide range of applications in ecology, species management and conservation, it is important that the impact that imperfect detection can have is fully accounted for.
Impact of imperfect detection in the estimation of the distribution of a virtual species. In this example, occupancy increases with elevation while detectability decreases, (i.e., occupancy and detectability are negatively correlated) (color range from red to dark blue corresponds to values from 0 to 1). If disregarded, imperfect detection causes the underestimation of species occurrence probabilities, especially at higher elevations (left column), and can lead to the misidentification of critical habitat (right column, in orange).
Assessing the impact of imperfect detection
We used simulations to explore a range of plausible scenarios where detectability was either constant, a function of the same factor driving species occupancy, or a function of an independent covariate. Using these scenarios, we assessed the performance of three classes of modeling methods (based on presence-absence, presence-background and occupancy-detection data), looking at two key performance aspects: model calibration (ie, how well the predicted probabilities match the observed proportions of sites occupied) and discrimination ability (ie, how well the model distinguishes between occupied and empty sites).
Summary of the impacts that disregarding imperfect detection can have on the performance of SDMs as a function of the structure of the detection process with respect to the occupancy process. Symbol ✗ represents that imperfect detection affects performance, and ✓ that it does not. This applies to both presence-absence and presence-background methods. Note that presence-background methods such as Maxent produce a suitability index rather than a probability and are thus not necessarily calibrated even assuming perfect detection.
The fact that imperfect detection can bias the estimation of ecologically relevant state variables is not new. However, it’s largely been overlooked in the SDM literature. Furthermore, there has been confusion about its impacts and about the performance of models that explicitly account for detectability. Having this in mind, there are three key messages that arise from our results:
- Presence-background methods are also affected by imperfect detection. Even if false absence records are avoided, imperfect detection may imply that presence records do not represent a random sample of sites where the species occurs, which leads to the biased estimation of environmental relationships. This means that presence-background methods such as Maxent are not any more immune to imperfect detection than presence-absence methods.
- The impact of imperfect detection depends on its relationship with the environment. Disregarding imperfect detection is a greater problem when detectability is negatively correlated with species occupancy, or when it depends on independent covariates that are included as candidate occupancy predictors. Not only is model calibration impacted, but discrimination ability is also undermined (see summary table). In practice this means that what seems the best habitat for the species may simply represent sites where the species is more easily observable (see example maps).
- Modeling detectability improves the performance of species distribution models. Modeling occupancy and detection simultaneously does not necessarily require a greater sampling effort, but rather that data are collected so that they are informative about detectability. While all models may have similar ability to describe where the species is observed, only models that account for detectability provide a reliable estimation of where the species occurs (i.e., true SDM). Past comparative studies that assessed the performance of SDMs by evaluating their ability to predict detections rather than presences failed to reveal the actual benefits of accounting for detectability.
So what can we conclude from all of this? Our recommendation is that recording data in ways that allow the modeling of the detection process should become standard practice in future surveys. Possible methods include recording the detection/non-detection of the species in multiple visits to sites, or inter-detection times within a single survey visit.
When such data are not available, diligent consideration and reporting of the possible impacts of imperfect detection, including how it is likely to bias inference and prediction, should be considered a minimum standard of good species distribution modeling practice.