Imperfect detection of individuals/species when counting is a well-known problem in almost all monitoring data. Imperfect detection may also occur in habitat monitoring, especially for small and linear structures. The resolution of remotely sensed data may be insufficient for them and they may be missed in field surveys. If detection probability is constant per species or habitat, as well as for a given observer and a given site, then it should not affect estimates of temporal trends. However, if detection probability varies through time, it dangerously confounds the trends we want to document with our monitoring data. We may erroneously conclude that a species or habitat is growing or declining when in fact the opposite is true. On this page, we outline general principles; on a complementary page, we explain methods that allow estimating detection probability.
Imperfect detection is of particular concern in species monitoring but it may also be a problem in habitat monitoring. In habitat monitoring it is primarily a problem for small and linear structures and for habitats that are difficult to discriminate from similar, more common ones and therefore maybe overlooked. Imperfect detection can occur both in field surveys and in remote sensing surveys.
In species monitoring it is rarely possible to perfectly detect and count all individuals. It may come to a surprise but experience has shown that even in large species living in semi-open habitats, such as elephants or giraffes, individuals may be overlooked.
Consequences of imperfect detection. If detection probability is constant per species or habitat, as well as for a given observer and a given site, then it should not affect estimates of temporal trends. However, if detection probability varies through time, it dangerously confounds the trends we want to document with our monitoring data. Temporal increase in experience of observers, increasing shyness of observed animals, temperature influence on activity, especially in cold-blooded animals, seasonal or diurnal changes in activity, method of observing a species, changes of habitat, and additional species-dependent factors may all cause temporal changes in detection probability. Also, spatial variations in the identity of the observer bias inferences of spatial patterns of abundance because of spatial variation in detection probability. The worst case is, if detectability varies with population size.
Imperfect detection of individuals may not be a dramatic problem. It just makes that count data must not be considered (and analysed) as censuses or density estimates. Under explicitly stated and validated assumptions, they remain the best available indices of abundance at large spatial and temporal scales.
Accounting for imperfect detection. There are several possibilities to account for imperfect detection. A frequently used approach is to carefully standardize the surveys. Many components of detection probability can be controlled for in the sampling design (same sites, same observers, same dates of visit) (Designing surveys to account for detectability) and we can hope that the number of counted individuals is linearly related to actual abundance. However, all factors that may change detection probability should be carefully considered before analysing data from standardized surveys. The following graph illustrates this for the common wall lizard (Podarcis muralis).
Factors influencing detectability of common wall lizards (Podarcis muralis).
A disadvantage of this approach is that we do not really know whether we were successful in our attempts to standardize our method sufficiently to obtain constant detection probabilities.
Other methods account for detection probability in the analysis of the data. The simplest way is to split the data into separate data sets for which detection probability is constant, e.g. a set for each observer or for each habitat type. Then each data set could be analysed separately. In large-scale monitoring schemes that analyse data with log-linear Poisson regression models, factors that influence detection probability can be included as a co-variate. General limits to these approaches are loss of statistical efficiency (overfitting of parameters and low accuracy of estimates). Other approaches are based on sampling designs where information is collected to estimate detection probability (mark-recapture, distance methods, double sampling, ground-truthing in habitat monitoring). Basically, these studies require that a subset of the whole population is known (e.g. by natural or artificial marks or complete coverage in field-based habitat monitoring) and detection is estimated as the fraction of the known subset sampled. For suitable study designs see Designing surveys to account for detectability and for estimation methods see Statistical methods to estimate detection probability.
- Bart, J., Droege, S., Geissler, P., Peterjohn, B. & Ralph, C.J. 2004. Density estimation in wildlife surveys. Wildl. Soc. Bull. 32: 1242-1247.
- Elzinga CL, Salzer DW, Willoughby JW, Gibbs JP (2001) Monitoring plant and animal populations: a handbook for field biologists. Blackwell Science, Malden, MA
- Thomas, L. 1996. Monitoring long-term population change: why are there so many analysis methods? Ecology 77: 49-58.
EuMon core team; May 2013