Non-detection by citizen scientists modeled as a function of visit characteristics

Citizen science is an increasingly valuable source of information about biodiversity. It is challenging to use this information for analysis of distribution and trends. The lack of a protocol leads to observation bias due to unequal detection and reporting probabilities, caused by different preferences and habits of citizen scientists. With an immediate focus on occupancy, it is hard to extract value from citizen science data because of such observation bias. We propose to incorporate multiple characteristics of excursions in analyses of data collected by citizen scientists to improve estimates of the probability that a species is not detected and reported, even though it does occur. By limiting these models to areas that are known to be occupied, detection can be modeled separately without considering variation in occupancy. We apply this idea to 150 common species in the Southwest Delta of The Netherlands, and illustrate the data selection, modeling process and results using four species. The strongest features to predict detection are the number of species reported during a visit (list length), earlier observations of the target species by the same observer, and the day of year. We compare three approaches to predict the probability that a species is not detected during any of the visits to an area. Predictions based on only the number of visits were outperformed by predictions that also take the list length into account. Predictions based on all features combined consistently beat both other approaches, across all 10 species groups that were compared. We thus show that explicitly modelling the characteristics of all visits to an occupied area results in estimation of non-detection probabilities, while providing insight into the causes of detection and reporting bias. Like in occupancy models, predictions of our model can be combined per area to update occupancy estimates and thereby correct bias in citizen science data when aiming to map a species’ distribution.