Citizen science decisions: A Bayesian approach optimises effort

Abstract

Volunteer citizen scientists are an invaluable resource for classifying large numbers of images that are used for species monitoring. Citizen science projects often rely on the “wisdom of the crowd” through majority vote methods to produce accurate classifications and assume all volunteer citizen scientists have equal ability. We use a Bayesian framework to estimate iNaturalist NZ user accuracies and simultaneously collectively classify the observations. We calculate the probability that the inferred observation classification from the Bayesian framework is correct for each observation given the assumed true user accuracies. We refer to this probability as the classification certainty. Our results show that 50% of images were classified by more volunteer citizen scientists than required to reach a minimal desired collective classification certainty level and more than one third of identifications were above the number required to meet the minimal desired classification certainty. Over 60% of observations that are yet to be considered research grade have a high classification certainty that has already surpassed the desired minimal level and could therefore be upgraded to research grade with no additional identifications. With more sophisticated collective classification methods than a simple majority vote procedure citizen science data and volunteer citizen scientists effort could be utilised more optimally.