A comparison of clustering approaches to create sites for occupancy models from opportunistic biodiversity surveys

Abstract Opportunistic surveys, such as those archived by citizen science biodiversity inventory programs, can provide species observations over large spans of space and time. Like most biodiversity data, opportunistic reports suffer from imperfect detection of organisms, which can distort species distribution models used for science and natural resource management. Occupancy models and their variants account for imperfect detection, but opportunistic surveys are usually not collected using the sampling protocols required for occupancy modelling. The Site Clustering Problem addresses this issue by formulating the construction of sites from unstructured observations as a clustering task. We developed and evaluated a workflow for building occupancy models from opportunistic data, comprised of steps to cluster observations, form site geometries and fit models. We compared eight clustering approaches as potential solutions to the Site Clustering Problem on 31 avian species in Oregon, USA. We evaluated the clustering approaches based on predictive performance of the downstream occupancy models and analysed the impact of species traits such as home range size and habitat type on the results. In simulation studies, we measured the difference between estimated parameters and true data-generating parameters. Our evaluation revealed several insights. Clustering methods from machine learning that accounted for similarity in both environmental and geographic space allowed observations from nearby geospatial coordinates in similar environments to be grouped together, and some of these methods were top performers. That said, the simple approach of grouping two to ten co-located observations as sites was also competitive. This study explored a set of options for bridging the gap between large databases of opportunistic biodiversity surveys and the sampling frameworks underlying models that correct for imperfect detection of organisms. Our results provide preliminary guidance for modellers seeking to create sites for occupancy models as well as directions for future work. In particular, we expect that the development of clustering approaches that can consider species-specific spatial constraints will improve solutions to the Site Clustering Problem.