Imperfect detection and its consequences for monitoring for conservation

Abstract

Biodiversity monitoring is important to identify conservation needs and test the efficacy of management actions. Variants of “abundance” (\ud N\ud ) are among the most widely monitored quantities, e.g., (true) abundance, number of occupied sites (distribution, occupancy) or species richness. We propose a sampling-based view of monitoring that clearly acknowledges two sampling processes involved when monitoring \ud N\ud . First, measurements from the surveyed sample area are generalized to a larger area, hence the importance of a probability sample. Second, even within sampled areas only a sample of units (individuals, occupied sites, species) is counted owing to imperfect detectability \ud p\ud . If \ud p\ud < 1, counts are random variables and their expectation \ud E\ud (\ud n\ud ) is related to \ud N\ud via the relationship \ud E\ud (\ud n\ud ) = \ud N*p\ud . Whenever \ud p\ud < 1, counts vary even under identical conditions and underestimate \ud N\ud , and patterns in counts confound patterns in \ud N\ud with those in \ud p\ud . In addition, part of the population \ud N\ud may be unavailable for detection, e.g., temporarily outside the sampled quadrat, underground or for another reason not exposed to sampling; hence a more general way of describing a count is \ud E\ud (\ud n\ud ) = \ud N*a*p\ud , where \ud a\ud is availability probability and \ud p\ud detection, given availability. We give two examples of monitoring schemes that highlight the importance of explicitly accounting for availability and detectability. In the Swiss reptile Red List update, the widespread and abundant slow worm (\ud Anguis fragilis\ud ) was recorded in only 22.1% of all sampled quadrats. Only an analysis that accounted for both availability and detectability gave realistic estimates of the species’ distribution. Among 128 bird species monitored in the Swiss breeding bird survey, detection in occupied 1 km2\ud quadrats averaged only 64% and varied tremendously by species (3–99 %); hence observed distributions greatly underestimated range sizes and should not be compared among species. We believe that monitoring design and analyses should properly account for these two sampling processes to enable valid inferences about biodiversity. We argue for a more rigorous approach to both monitoring design and analysis to obtain the best possible information about the state of nature. An explicit recognition of, and proper accounting for, the two sampling processes involved in most monitoring programs will go a long way towards this goal