Explorers vs. followers: A behavioural approach to spatial bias correction in species distribution modelling

In recent years, the increase in data availability through citizen science data collection has raised questions about the quality of this data. Species distribution models can be severely impacted by non-random spatial distributions of records. Multiple methods exist to correct for spatial bias and most of them imply that the sampling is uneven in space and determined by the observers’ choices of where to search for observations. Most methods for addressing sampling biases in opportunistic datasets assume that each observer behaves uniformly, which in practice may not be the case. We focus our study on a widely-used correction method, chosen for its adaptable framework, and assess its effectiveness in mitigating biases from a group of observers with varying behaviours. This method includes a covariate in the model as a bias proxy and corrects for this bias by setting this covariate equal to a constant upon prediction. We differentiate two observer behaviours: exploring and following. Under this paradigm, explorers select destinations far away from the current set of observed points, while followers choose destinations at or near one of the observed points. As such, it is worth investigating whether the current approaches to correcting for observer bias hold under varying observer behaviours, or whether a data-driven approach based on modelled observer behaviour may lead to better predictions. To do so, we developed a new software platform, obsimulator, to simulate patterns of points driven by observer behaviour. We established a correction method based on a bias incorporation approach using k-nearest neighbours. We found that including a bias covariate and setting it to a constant for prediction yields the best results and the strength of the correction differs between cohorts of observers. Additionally, the optimal number of neighbouring points and smoothing parameters depends on the ratio of explorers versus followers in the observers’ cohort. • Bias correction often fails to address non-uniform observer behaviour. • Proxy covariate bias correction offers flexibility for diverse observer types. • Optimal correction strategies vary by observer type and specific cohort. • Data-driven approach based on modelled observer behaviour improves model predictions. • Accounting for diverse behaviours improves citizen science data model accuracy.