Effect of sampling strategies on the response curves estimated by plant species distribution models

Abstract

Species distribution models (SDMs) rely on species presence/absence or abundance data and environmental variables to estimate species response curves. Therefore, the quality (and quantity, i.e., sample size) of the data to describe the species distribution determines the quality of the estimate of the species-environment relationship. However, SDMs are seldom fitted on high-quality data collected strictly for that purpose. Usually, SDMs rely on a collection of opportunistic datasets sampled from previous projects or public repositories with different objectives. Here, we aim at assessing how the sampling strategy capturing the geographic distribution of a species affects the accuracy and precision of its response curves along environmental gradients, as estimated by parametric SDMs. We simulated the occurrence of two virtual plant species across the Abruzzo region (Italy). We assumed that the two virtual plants were similarly affected by precipitation, but one had a wider realised niche for temperature (i.e., higher thermal tolerance), and, as a result, a wider distribution extent. Then, we sampled occurrence data for the two species following five different sampling strategies: random, stratified, systematic, topographic, and uniform (the latter performed within the environmental space). In addition, we simulated a spatially biased sampling design by collecting presence/absence data close to roads. To account for sample size, we also repeated our simulations along a gradient of increasing sampling effort, i.e., number of sampled locations. In total, we ran 500 replicates for each combination of sampling design and effort. For each replicate, we fitted SDMs using binomial generalised linear models and extracted the model coefficients for precipitation and temperature to be compared with the true coefficients from the virtual species’ model. We evaluated the quality of the estimated response curves by computing the following measures: bias (accuracy), variance (precision), and mean squared error (accuracy and precision). Our results suggest that a proper estimate of the species response curve can be obtained when the choice of the sampling strategy is guided by the species’ ecology. In particular, species with wide tolerances to environmental drivers may be better modelled using data uniformly collected within the environmental space, while none of the tested sampling designs seemed to substantially outperform the others for modelling species with a narrow realised niche.

Testing the effect of sample prevalence and sampling methods on probability- and favourability-based SDMs

“…However, despite this achievement, the species response curves estimated by SDMs are still conditioned by the collected data (i.e., presence samples, presence/absence samples, pseudo-absences, background points) used for the model calibration. Indeed, different survey strategies may influence the accuracy and the quality of predictions (Hirzel and Guisan, 2002;Thibaud et al, 2014;Bazzichetto et al, 2022). Therefore, an efficient sampling method is crucial for avoiding spatial heterogeneity in the sampling intensity (e.g., incomplete sampling and over-sampling) of species occurrences and pseudo-absences/background points (Inman et al, 2021).…”

Causal inference and large‐scale expert validation shed light on the drivers of SDM accuracy and variance

“…The predictive performance of a SDM may be decomposed broadly into its accuracy and precision (Bazzichetto et al, 2022).…”

“…The predictive performance of a SDM may be decomposed broadly into its accuracy and precision (Bazzichetto et al, 2022). Accuracy is a measure of how close the model's predictions are to the “truth” on average.…”

Causal inference and large-scale expert validation shed light on the drivers of SDM accuracy and variance

“…The performance of a SDM may be decomposed broadly into its accuracy and precision (Bazzichetto et al, 2022). Accuracy is a measure of how close the model's predictions are to the "truth" on average.…”

“…Methodological decisions include the choice of SDM algorithm or ensemble of algorithms (Fukuda & De Baets, 2016;Hao et al, 2020), environmental covariates (Arenas-Castro et al, 2022;Bucklin et al, 2015;De Marco & Nóbrega, 2018), and strategies to mitigate undesirable properties of the occurrence data (Barbet-Massin et al, 2012;Beck et al, 2014;Chapman et al, 2019;Dudík et al, 2005;Fourcade et al, 2014;Phillips et al, 2009). Data characteristics include the extent of spatial clustering and geographic bias (Bazzichetto et al, 2022;Beck et al, 2014;Steen et al, 2020), the expertise of data collectors (Steen et al, 2019), the ratio of presences to absences (Fukuda & De Baets, 2016), coverage of species' geographic ranges (Konowalik & Nosol, 2021), and sample size (Feeley & Silman, 2011;Hernandez et al, 2006;Stockwell & Peterson, 2002;Wisz et al, 2008). And finally, species traits include range size relative to the study extent (Santika, 2011) and niche breadth (Hernandez et al, 2006;Tessarolo et al, 2021), amongst others.…”