Spatial thinning and class balancing: Key choices lead to variation in the performance of species distribution models with citizen science data

Abstract

Spatial biases are a common feature of presence–absence data from citizen scientists. Spatial thinning can mitigate errors in species distribution models (SDMs) that use these data. When detections or non‐detections are rare, however, SDMs may suffer from class imbalance or low sample size of the minority (i.e. rarer) class. Poor predictions can result, the severity of which may vary by modelling technique. To explore the consequences of spatial bias and class imbalance in presence–absence data, we used eBird citizen science data for 102 bird species from the northeastern USA to compare spatial thinning, class balancing and majority‐only thinning (i.e. retaining all samples of the minority class). We created SDMs using two parametric or semi‐parametric techniques (generalized linear models and generalized additive models) and two machine learning techniques (random forest and boosted regression trees). We tested the predictive abilities of these SDMs using an independent and systematically collected reference dataset with a combination of discrimination (area under the receiver operator characteristic curve; true skill statistic; area under the precision‐recall curve) and calibration (Brier score; Cohen's kappa) metrics. We found large variation in SDM performance depending on thinning and balancing decisions. Across all species, there was no single best approach, with the optimal choice of thinning and/or balancing depending on modelling technique, performance metric and the baseline sample prevalence of species in the data. Spatially thinning all the data was often a poor approach, especially for species with baseline sample prevalence <0.1. For most of these rare species, balancing classes improved model discrimination between presence and absence classes using machine learning techniques, but typically hindered model calibration. Baseline sample prevalence, sample size, modelling approach and the intended application of SDM output—whether discrimination or calibration—should guide decisions about how to thin or balance data, given the considerable influence of these methodological choices on SDM performance. For prognostic applications requiring good model calibration (vis‐à‐vis discrimination), the match between sample prevalence and true species prevalence may be the overriding feature and warrants further investigation.

Data integration methods to account for spatial niche truncation effects in regional projections of species distribution

“…While many SDMs are based on PO data similar to ours owing to spatial thinning where occurrences are removed to reduce the effect of spatial sampling bias on parameter estimates (see Steen et al. [2020] for a review), multiple occurrences per coarse‐grain pixels are frequently considered in the context of data integration using a Poisson point process (PPP; Dorazio 2014, Fithian et al. 2015 a , Fletcher et al.…”

“…They are thus different from classical PO data obtained from databases of opportunistic records (e.g., GBIF) where multiple occurrences can occur in a given cell. While many SDMs are based on PO data similar to ours owing to spatial thinning where occurrences are removed to reduce the effect of spatial sampling bias on parameter estimates (see Steen et al [2020] for a review), multiple occurrences per coarse-grain pixels are frequently considered in the context of data integration using a Poisson point process (PPP; Dorazio 2014, Fithian et al 2015a, Fletcher et al 2016, Koshkina et al 2017. Although we here focus on PO data obtained from IUCN range maps or atlases where only one occurrence per coarse-grain cell is considered, the PO-based models presented below are based on a PPP and are therefore also valid with multiple occurrences.…”

Towards ecological science for all by all

“…birds), and a non-random skew of observations towards anthropogenic land-uses ('road-side bias'), whether the focus be on any, rare or non-native species. Thus, the study by Petersen et al (2021) provides useful context for the remaining three papers under the theme of quality control which consider how to account for different forms of bias generated through citizen science data related to site selection and observer retention over time (Dambly et al, 2021), artificial light conditions (Ditmer et al, 2021), and choice of methods in generating less biased SDMs (Steen et al, 2021).…”

Novel methods to correct for observer and sampling bias in presence‐only species distribution models

“…However, various methods of bias correction have been suggested and combined in the literature. For example, spatial thinning of species observations according to the study area’s resolution (Aiello‐Lammens et al., 2015; Kiedrzyński et al., 2017; Steen et al., 2020), model‐based corrections (Komori et al., 2020; Stolar & Nielsen, 2015), or combining species observations with large survey data (Fithian et al., 2015; Fletcher et al., 2016) have been recommended. Yet overall, sampling background data with the target‐group strategy—that is, using similar sampling design/bias to sample background points as observed species presence points—remains currently the most popular approach (Botella et al., 2020; Hertzog et al., 2014; Kramer‐Schadt et al., 2013; Phillips et al., 2009; Righetti et al., 2019).…”