Enhancing Absence Inference in Occupancy Modelling of Citizen Science Data

Rachel Drake

Citizen Science data have become increasingly widespread in applied ecological research. The structure of this data can range from rigorous protocols to the submission of single observations to public databases. Under the framework of occupancy models, this range of data structures all utilise the same methodology for absent observations, where a presence record infers an absence of all other species at that location and time. This study aims to both develop methods for improving absence inference in both checklist and opportunistic data types, as well as providing novel insight into possible bias present in model validation. To accomplish this we use eBird and iNaturalist data to compare both checklist and opportunistic data types across two species designed to test different ecological bird behaviours as well as observer protocols. We accomplish this using the unmarked package in R, alongside open source environmental data. Our results reveal that occupancy models applied to both complete and opportunistic datasets yield divergent ecological inferences. In certain instances, these differences are as extreme as inverted trends in species occupancy across specific regions. The estimated effects of ecological parameters on species occupancy also exhibit inversion within the model covariates. Additionally, all models underwent comparison against two benchmark datasets, each derived from either checklist or opportunistic data. Predictions from these models exhibit significant variation across datasets and model types. The findings emphasise the importance of preprocessing opportunistic data in occupancy modelling, along with careful planning of model validation to mitigate potential sources of bias. This holds significant value for researchers employing occupancy modelling with citizen science data, aiming to provide ecologically robust inference. Keywords: occupancy modelling, citizen science, eBird, preprocessing, model validation.